Tag Archives: Apache

Amazon QuickSight Update – Geospatial Visualization, Private VPC Access, and More

Post Syndicated from Jeff Barr original https://aws.amazon.com/blogs/aws/amazon-quicksight-update-geospatial-visualization-private-vpc-access-and-more/

We don’t often recognize or celebrate anniversaries at AWS. With nearly 100 services on our list, we’d be eating cake and drinking champagne several times a week. While that might sound like fun, we’d rather spend our working hours listening to customers and innovating. With that said, Amazon QuickSight has now been generally available for a little over a year and I would like to give you a quick update!

QuickSight in Action
Today, tens of thousands of customers (from startups to enterprises, in industries as varied as transportation, legal, mining, and healthcare) are using QuickSight to analyze and report on their business data.

Here are a couple of examples:

Gemini provides legal evidence procurement for California attorneys who represent injured workers. They have gone from creating custom reports and running one-off queries to creating and sharing dynamic QuickSight dashboards with drill-downs and filtering. QuickSight is used to track sales pipeline, measure order throughput, and to locate bottlenecks in the order processing pipeline.

Jivochat provides a real-time messaging platform to connect visitors to website owners. QuickSight lets them create and share interactive dashboards while also providing access to the underlying datasets. This has allowed them to move beyond the sharing of static spreadsheets, ensuring that everyone is looking at the same and is empowered to make timely decisions based on current data.

Transfix is a tech-powered freight marketplace that matches loads and increases visibility into logistics for Fortune 500 shippers in retail, food and beverage, manufacturing, and other industries. QuickSight has made analytics accessible to both BI engineers and non-technical business users. They scrutinize key business and operational metrics including shipping routes, carrier efficient, and process automation.

Looking Back / Looking Ahead
The feedback on QuickSight has been incredibly helpful. Customers tell us that their employees are using QuickSight to connect to their data, perform analytics, and make high-velocity, data-driven decisions, all without setting up or running their own BI infrastructure. We love all of the feedback that we get, and use it to drive our roadmap, leading to the introduction of over 40 new features in just a year. Here’s a summary:

Looking forward, we are watching an interesting trend develop within our customer base. As these customers take a close look at how they analyze and report on data, they are realizing that a serverless approach offers some tangible benefits. They use Amazon Simple Storage Service (S3) as a data lake and query it using a combination of QuickSight and Amazon Athena, giving them agility and flexibility without static infrastructure. They also make great use of QuickSight’s dashboards feature, monitoring business results and operational metrics, then sharing their insights with hundreds of users. You can read Building a Serverless Analytics Solution for Cleaner Cities and review Serverless Big Data Analytics using Amazon Athena and Amazon QuickSight if you are interested in this approach.

New Features and Enhancements
We’re still doing our best to listen and to learn, and to make sure that QuickSight continues to meet your needs. I’m happy to announce that we are making seven big additions today:

Geospatial Visualization – You can now create geospatial visuals on geographical data sets.

Private VPC Access – You can now sign up to access a preview of a new feature that allows you to securely connect to data within VPCs or on-premises, without the need for public endpoints.

Flat Table Support – In addition to pivot tables, you can now use flat tables for tabular reporting. To learn more, read about Using Tabular Reports.

Calculated SPICE Fields – You can now perform run-time calculations on SPICE data as part of your analysis. Read Adding a Calculated Field to an Analysis for more information.

Wide Table Support – You can now use tables with up to 1000 columns.

Other Buckets – You can summarize the long tail of high-cardinality data into buckets, as described in Working with Visual Types in Amazon QuickSight.

HIPAA Compliance – You can now run HIPAA-compliant workloads on QuickSight.

Geospatial Visualization
Everyone seems to want this feature! You can now take data that contains a geographic identifier (country, city, state, or zip code) and create beautiful visualizations with just a few clicks. QuickSight will geocode the identifier that you supply, and can also accept lat/long map coordinates. You can use this feature to visualize sales by state, map stores to shipping destinations, and so forth. Here’s a sample visualization:

To learn more about this feature, read Using Geospatial Charts (Maps), and Adding Geospatial Data.

Private VPC Access Preview
If you have data in AWS (perhaps in Amazon Redshift, Amazon Relational Database Service (RDS), or on EC2) or on-premises in Teradata or SQL Server on servers without public connectivity, this feature is for you. Private VPC Access for QuickSight uses an Elastic Network Interface (ENI) for secure, private communication with data sources in a VPC. It also allows you to use AWS Direct Connect to create a secure, private link with your on-premises resources. Here’s what it looks like:

If you are ready to join the preview, you can sign up today.

Jeff;

 

Introducing container-diff, a tool for quickly comparing container images (Google Open Source Blog)

Post Syndicated from jake original https://lwn.net/Articles/739384/rss

Google has announced that it has released its container-diff tool under the Apache v2 license.
container-diff helps users investigate image changes by computing semantic diffs between images. What this means is that container-diff figures out on a low-level what data changed, and then combines this with an understanding of package manager information to output this information in a format that’s actually readable to users. The tool can find differences in system packages, language-level packages, and files in a container image.

Users can specify images in several formats – from local Docker daemon (using the prefix `daemon://` on the image path), a remote registry (using the prefix `remote://`), or a file in the .tar in the format exported by “docker save” command. You can also combine these formats to compute the diff between a local version of an image and a remote version.”

Visualize AWS Cloudtrail Logs using AWS Glue and Amazon Quicksight

Post Syndicated from Luis Caro Perez original https://aws.amazon.com/blogs/big-data/streamline-aws-cloudtrail-log-visualization-using-aws-glue-and-amazon-quicksight/

Being able to easily visualize AWS CloudTrail logs gives you a better understanding of how your AWS infrastructure is being used. It can also help you audit and review AWS API calls and detect security anomalies inside your AWS account. To do this, you must be able to perform analytics based on your CloudTrail logs.

In this post, I walk through using AWS Glue and AWS Lambda to convert AWS CloudTrail logs from JSON to a query-optimized format dataset in Amazon S3. I then use Amazon Athena and Amazon QuickSight to query and visualize the data.

Solution overview

To process CloudTrail logs, you must implement the following architecture:

CloudTrail delivers log files in an Amazon S3 bucket folder. To correctly crawl these logs, you modify the file contents and folder structure using an Amazon S3-triggered Lambda function that stores the transformed files in an S3 bucket single folder. When the files are in a single folder, AWS Glue scans the data, converts it into Apache Parquet format, and catalogs it to allow for querying and visualization using Amazon Athena and Amazon QuickSight.

Walkthrough

Let’s look at the steps that are required to build the solution.

Set up CloudTrail logs

First, you need to set up a trail that delivers log files to an S3 bucket. To create a trail in CloudTrail, follow the instructions in Creating a Trail.

When you finish, the trail settings page should look like the following screenshot:

In this example, I set up log files to be delivered to the cloudtraillfcaro bucket.

Consolidate CloudTrail reports into a single folder using Lambda

AWS CloudTrail delivers log files using the following folder structure inside the configured Amazon S3 bucket:

AWSLogs/ACCOUNTID/CloudTrail/REGION/YEAR/MONTH/HOUR/filename.json.gz

Additionally, log files have the following structure:

{
    "Records": [{
        "eventVersion": "1.01",
        "userIdentity": {
            "type": "IAMUser",
            "principalId": "AIDAJDPLRKLG7UEXAMPLE",
            "arn": "arn:aws:iam::123456789012:user/Alice",
            "accountId": "123456789012",
            "accessKeyId": "AKIAIOSFODNN7EXAMPLE",
            "userName": "Alice",
            "sessionContext": {
                "attributes": {
                    "mfaAuthenticated": "false",
                    "creationDate": "2014-03-18T14:29:23Z"
                }
            }
        },
        "eventTime": "2014-03-18T14:30:07Z",
        "eventSource": "cloudtrail.amazonaws.com",
        "eventName": "StartLogging",
        "awsRegion": "us-west-2",
        "sourceIPAddress": "72.21.198.64",
        "userAgent": "signin.amazonaws.com",
        "requestParameters": {
            "name": "Default"
        },
        "responseElements": null,
        "requestID": "cdc73f9d-aea9-11e3-9d5a-835b769c0d9c",
        "eventID": "3074414d-c626-42aa-984b-68ff152d6ab7"
    },
    ... additional entries ...
    ]

If AWS Glue crawlers are used to catalog these files as they are written, the following obstacles arise:

  1. AWS Glue identifies different tables per different folders because they don’t follow a traditional partition format.
  2. Based on the structure of the file content, AWS Glue identifies the tables as having a single column of type array.
  3. CloudTrail logs have JSON attributes that use uppercase letters. According to the Best Practices When Using Athena with AWS Glue, it is recommended that you convert these to lowercase.

To have AWS Glue catalog all log files in a single table with all the columns describing each event, implement the following Lambda function:

from __future__ import print_function
import json
import urllib
import boto3
import gzip

s3 = boto3.resource('s3')
client = boto3.client('s3')

def convertColumntoLowwerCaps(obj):
    for key in obj.keys():
        new_key = key.lower()
        if new_key != key:
            obj[new_key] = obj[key]
            del obj[key]
    return obj


def lambda_handler(event, context):

    bucket = event['Records'][0]['s3']['bucket']['name']
    key = urllib.unquote_plus(event['Records'][0]['s3']['object']['key'].encode('utf8'))
    print(bucket)
    print(key)
    try:
        newKey = 'flatfiles/' + key.replace("/", "")
        client.download_file(bucket, key, '/tmp/file.json.gz')
        with gzip.open('/tmp/out.json.gz', 'w') as output, gzip.open('/tmp/file.json.gz', 'rb') as file:
            i = 0
            for line in file: 
                for record in json.loads(line,object_hook=convertColumntoLowwerCaps)['records']:
            		if i != 0:
            		    output.write("\n")
            		output.write(json.dumps(record))
            		i += 1
        client.upload_file('/tmp/out.json.gz', bucket,newKey)
        return "success"
    except Exception as e:
        print(e)
        print('Error processing object {} from bucket {}. Make sure they exist and your bucket is in the same region as this function.'.format(key, bucket))
        raise e

The function goes over each element of the records array, changes uppercase letters to lowercase in column names, and inserts each element of the array as a single line of a new file. The new file is saved inside a flatfiles folder created by the function without any subfolders in the S3 bucket.

The function should have a role containing a policy with at least the following permissions:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Action": [
                "s3:*"
            ],
            "Resource": [
                "arn:aws:s3:::cloudtraillfcaro/*",
                "arn:aws:s3:::cloudtraillfcaro"
            ],
            "Effect": "Allow"
        }
    ]
}

In this example, CloudTrail delivers logs to the cloudtraillfcaro bucket. Make sure that you replace this name with your bucket name in the policy. For more information about how to work with inline policies, see Working with Inline Policies.

After the Lambda function is created, you can set up the following trigger using the Triggers tab on the AWS Lambda console.

Choose Add trigger, and choose S3 as a source of the trigger.

After choosing the source, configure the following settings:

In the trigger, any file that is written to the path for the log files—which in this case is AWSLogs/119582755581/CloudTrail/—is processed. Make sure that the Enable trigger check box is selected and that the bucket and prefix parameters match your use case.

After you set up the function and receive log files, the bucket (in this case cloudtraillfcaro) should contain the processed files inside the flatfiles folder.

Catalog source data

Once the files are processed by the Lambda function, set up a crawler named cloudtrail to catalog them.

The crawler must point to the flatfiles folder.

All the crawlers and AWS Glue jobs created for this solution must have a role with the AWSGlueServiceRole managed policy and an inline policy with permissions to modify the S3 buckets used on the Lambda function. For more information, see Working with Managed Policies.

The role should look like the following:

In this example, the inline policy named s3perms contains the permissions to modify the S3 buckets.

After you choose the role, you can schedule the crawler to run on demand.

A new database is created, and the crawler is set to use it. In this case, the cloudtrail database is used for all the tables.

After the crawler runs, a single table should be created in the catalog with the following structure:

The table should contain the following columns:

Create and run the AWS Glue job

To convert all the CloudTrail logs to a columnar store in Parquet, set up an AWS Glue job by following these steps.

Upload the following script into a bucket in Amazon S3:

import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job
import boto3
import time

## @params: [JOB_NAME]
args = getResolvedOptions(sys.argv, ['JOB_NAME'])

sc = SparkContext()
glueContext = GlueContext(sc)
spark = glueContext.spark_session
job = Job(glueContext)
job.init(args['JOB_NAME'], args)

datasource0 = glueContext.create_dynamic_frame.from_catalog(database = "cloudtrail", table_name = "flatfiles", transformation_ctx = "datasource0")
resolvechoice1 = ResolveChoice.apply(frame = datasource0, choice = "make_struct", transformation_ctx = "resolvechoice1")
relationalized1 = resolvechoice1.relationalize("trail", args["TempDir"]).select("trail")
datasink = glueContext.write_dynamic_frame.from_options(frame = relationalized1, connection_type = "s3", connection_options = {"path": "s3://cloudtraillfcaro/parquettrails"}, format = "parquet", transformation_ctx = "datasink4")
job.commit()

In the example, you load the script as a file named cloudtrailtoparquet.py. Make sure that you modify the script and update the “{"path": "s3://cloudtraillfcaro/parquettrails"}” with the destination in which you want to store your results.

After uploading the script, add a new AWS Glue job. Choose a name and role for the job, and choose the option of running the job from An existing script that you provide.

To avoid processing the same data twice, enable the Job bookmark setting in the Advanced properties section of the job properties.

Choose Next twice, and then choose Finish.

If logs are already in the flatfiles folder, you can run the job on demand to generate the first set of results.

Once the job starts running, wait for it to complete.

When the job is finished, its Run status should be Succeeded. After that, you can verify that the Parquet files are written to the Amazon S3 location.

Catalog results

To be able to process results from Athena, you can use an AWS Glue crawler to catalog the results of the AWS Glue job.

In this example, the crawler is set to use the same database as the source named cloudtrail.

You can run the crawler using the console. When the crawler finishes running and has processed the Parquet results, a new table should be created in the AWS Glue Data Catalog. In this example, it’s named parquettrails.

The table should have the classification set to parquet.

It should have the same columns as the flatfiles table, with the exception of the struct type columns, which should be relationalized into several columns:

In this example, notice how the requestparameters column, which was a struct in the original table (flatfiles), was transformed to several columns—one for each key value inside it. This is done using a transformation native to AWS Glue called relationalize.

Query results with Athena

After crawling the results, you can query them using Athena. For example, to query what events took place in the time frame between 2017-10-23t12:00:00 and 2017-10-23t13:00, use the following select statement:

select *
from cloudtrail.parquettrails
where eventtime > '2017-10-23T12:00:00Z' AND eventtime < '2017-10-23T13:00:00Z'
order by eventtime asc;

Be sure to replace cloudtrail.parquettrails with the names of your database and table that references the Parquet results. Replace the datetimes with an hour when your account had activity and was processed by the AWS Glue job.

Visualize results using Amazon QuickSight

Once you can query the data using Athena, you can visualize it using Amazon QuickSight. Before connecting Amazon QuickSight to Athena, be sure to grant QuickSight access to Athena and the associated S3 buckets in your account. For more information, see Managing Amazon QuickSight Permissions to AWS Resources. You can then create a new data set in Amazon QuickSight based on the Athena table that you created.

After setting up permissions, you can create a new analysis in Amazon QuickSight by choosing New analysis.

Then add a new data set.

Choose Athena as the source.

Give the data source a name (in this case, I named it cloudtrail).

Choose the name of the database and the table referencing the Parquet results.

Then choose Visualize.

After that, you should see the following screen:

Now you can create some visualizations. First, search for the sourceipaddress column, and drag it to the AutoGraph section.

You can see a list of the IP addresses that you have used to interact with AWS. To review whether these IP addresses have been used from IAM users, internal AWS services, or roles, use the type value that is inside the useridentity field of the original log files. Thanks to the relationalize transformation, this value is available as the useridentity.type column. After the column is added into the Group/Color box, the visualization should look like the following:

You can now see and distinguish the most used IPs and whether they are used from roles, AWS services, or IAM users.

After following all these steps, you can use Amazon QuickSight to add different columns from CloudTrail and perform different types of visualizations. You can build operational dashboards that continuously monitor AWS infrastructure usage and access. You can share those dashboards with others in your organization who might need to see this data.

Summary

In this post, you saw how you can use a simple Lambda function and an AWS Glue script to convert text files into Parquet to improve Athena query performance and data compression. The post also demonstrated how to use AWS Lambda to preprocess files in Amazon S3 and transform them into a format that is recognizable by AWS Glue crawlers.

This example, used AWS CloudTrail logs, but you can apply the proposed solution to any set of files that after preprocessing, can be cataloged by AWS Glue.


Additional Reading

Learn how to Harmonize, Query, and Visualize Data from Various Providers using AWS Glue, Amazon Athena, and Amazon QuickSight.


About the Authors

Luis Caro is a Big Data Consultant for AWS Professional Services. He works with our customers to provide guidance and technical assistance on big data projects, helping them improving the value of their solutions when using AWS.

 

 

 

Me on the Equifax Breach

Post Syndicated from Bruce Schneier original https://www.schneier.com/blog/archives/2017/11/me_on_the_equif.html

Testimony and Statement for the Record of Bruce Schneier
Fellow and Lecturer, Belfer Center for Science and International Affairs, Harvard Kennedy School
Fellow, Berkman Center for Internet and Society at Harvard Law School

Hearing on “Securing Consumers’ Credit Data in the Age of Digital Commerce”

Before the

Subcommittee on Digital Commerce and Consumer Protection
Committee on Energy and Commerce
United States House of Representatives

1 November 2017
2125 Rayburn House Office Building
Washington, DC 20515

Mister Chairman and Members of the Committee, thank you for the opportunity to testify today concerning the security of credit data. My name is Bruce Schneier, and I am a security technologist. For over 30 years I have studied the technologies of security and privacy. I have authored 13 books on these subjects, including Data and Goliath: The Hidden Battles to Collect Your Data and Control Your World (Norton, 2015). My popular newsletter CryptoGram and my blog Schneier on Security are read by over 250,000 people.

Additionally, I am a Fellow and Lecturer at the Harvard Kennedy School of Government –where I teach Internet security policy — and a Fellow at the Berkman-Klein Center for Internet and Society at Harvard Law School. I am a board member of the Electronic Frontier Foundation, AccessNow, and the Tor Project; and an advisory board member of Electronic Privacy Information Center and VerifiedVoting.org. I am also a special advisor to IBM Security and the Chief Technology Officer of IBM Resilient.

I am here representing none of those organizations, and speak only for myself based on my own expertise and experience.

I have eleven main points:

1. The Equifax breach was a serious security breach that puts millions of Americans at risk.

Equifax reported that 145.5 million US customers, about 44% of the population, were impacted by the breach. (That’s the original 143 million plus the additional 2.5 million disclosed a month later.) The attackers got access to full names, Social Security numbers, birth dates, addresses, and driver’s license numbers.

This is exactly the sort of information criminals can use to impersonate victims to banks, credit card companies, insurance companies, cell phone companies and other businesses vulnerable to fraud. As a result, all 143 million US victims are at greater risk of identity theft, and will remain at risk for years to come. And those who suffer identify theft will have problems for months, if not years, as they work to clean up their name and credit rating.

2. Equifax was solely at fault.

This was not a sophisticated attack. The security breach was a result of a vulnerability in the software for their websites: a program called Apache Struts. The particular vulnerability was fixed by Apache in a security patch that was made available on March 6, 2017. This was not a minor vulnerability; the computer press at the time called it “critical.” Within days, it was being used by attackers to break into web servers. Equifax was notified by Apache, US CERT, and the Department of Homeland Security about the vulnerability, and was provided instructions to make the fix.

Two months later, Equifax had still failed to patch its systems. It eventually got around to it on July 29. The attackers used the vulnerability to access the company’s databases and steal consumer information on May 13, over two months after Equifax should have patched the vulnerability.

The company’s incident response after the breach was similarly damaging. It waited nearly six weeks before informing victims that their personal information had been stolen and they were at increased risk of identity theft. Equifax opened a website to help aid customers, but the poor security around that — the site was at a domain separate from the Equifax domain — invited fraudulent imitators and even more damage to victims. At one point, the official Equifax communications even directed people to that fraudulent site.

This is not the first time Equifax failed to take computer security seriously. It confessed to another data leak in January 2017. In May 2016, one of its websites was hacked, resulting in 430,000 people having their personal information stolen. Also in 2016, a security researcher found and reported a basic security vulnerability in its main website. And in 2014, the company reported yet another security breach of consumer information. There are more.

3. There are thousands of data brokers with similarly intimate information, similarly at risk.

Equifax is more than a credit reporting agency. It’s a data broker. It collects information about all of us, analyzes it all, and then sells those insights. It might be one of the biggest, but there are 2,500 to 4,000 other data brokers that are collecting, storing, and selling information about us — almost all of them companies you’ve never heard of and have no business relationship with.

The breadth and depth of information that data brokers have is astonishing. Data brokers collect and store billions of data elements covering nearly every US consumer. Just one of the data brokers studied holds information on more than 1.4 billion consumer transactions and 700 billion data elements, and another adds more than 3 billion new data points to its database each month.

These brokers collect demographic information: names, addresses, telephone numbers, e-mail addresses, gender, age, marital status, presence and ages of children in household, education level, profession, income level, political affiliation, cars driven, and information about homes and other property. They collect lists of things we’ve purchased, when we’ve purchased them, and how we paid for them. They keep track of deaths, divorces, and diseases in our families. They collect everything about what we do on the Internet.

4. These data brokers deliberately hide their actions, and make it difficult for consumers to learn about or control their data.

If there were a dozen people who stood behind us and took notes of everything we purchased, read, searched for, or said, we would be alarmed at the privacy invasion. But because these companies operate in secret, inside our browsers and financial transactions, we don’t see them and we don’t know they’re there.

Regarding Equifax, few consumers have any idea what the company knows about them, who they sell personal data to or why. If anyone knows about them at all, it’s about their business as a credit bureau, not their business as a data broker. Their website lists 57 different offerings for business: products for industries like automotive, education, health care, insurance, and restaurants.

In general, options to “opt-out” don’t work with data brokers. It’s a confusing process, and doesn’t result in your data being deleted. Data brokers will still collect data about consumers who opt out. It will still be in those companies’ databases, and will still be vulnerable. It just don’t be included individually when they sell data to their customers.

5. The existing regulatory structure is inadequate.

Right now, there is no way for consumers to protect themselves. Their data has been harvested and analyzed by these companies without their knowledge or consent. They cannot improve the security of their personal data, and have no control over how vulnerable it is. They only learn about data breaches when the companies announce them — which can be months after the breaches occur — and at that point the onus is on them to obtain credit monitoring services or credit freezes. And even those only protect consumers from some of the harms, and only those suffered after Equifax admitted to the breach.

Right now, the press is reporting “dozens” of lawsuits against Equifax from shareholders, consumers, and banks. Massachusetts has sued Equifax for violating state consumer protection and privacy laws. Other states may follow suit.

If any of these plaintiffs win in the court, it will be a rare victory for victims of privacy breaches against the companies that have our personal information. Current law is too narrowly focused on people who have suffered financial losses directly traceable to a specific breach. Proving this is difficult. If you are the victim of identity theft in the next month, is it because of Equifax or does the blame belong to another of the thousands of companies who have your personal data? As long as one can’t prove it one way or the other, data brokers remain blameless and liability free.

Additionally, much of this market in our personal data falls outside the protections of the Fair Credit Reporting Act. And in order for the Federal Trade Commission to levy a fine against Equifax, it needs to have a consent order and then a subsequent violation. Any fines will be limited to credit information, which is a small portion of the enormous amount of information these companies know about us. In reality, this is not an effective enforcement regime.

Although the FTC is investigating Equifax, it is unclear if it has a viable case.

6. The market cannot fix this because we are not the customers of data brokers.

The customers of these companies are people and organizations who want to buy information: banks looking to lend you money, landlords deciding whether to rent you an apartment, employers deciding whether to hire you, companies trying to figure out whether you’d be a profitable customer — everyone who wants to sell you something, even governments.

Markets work because buyers choose from a choice of sellers, and sellers compete for buyers. None of us are Equifax’s customers. None of us are the customers of any of these data brokers. We can’t refuse to do business with the companies. We can’t remove our data from their databases. With few limited exceptions, we can’t even see what data these companies have about us or correct any mistakes.

We are the product that these companies sell to their customers: those who want to use our personal information to understand us, categorize us, make decisions about us, and persuade us.

Worse, the financial markets reward bad security. Given the choice between increasing their cybersecurity budget by 5%, or saving that money and taking the chance, a rational CEO chooses to save the money. Wall Street rewards those whose balance sheets look good, not those who are secure. And if senior management gets unlucky and the a public breach happens, they end up okay. Equifax’s CEO didn’t get his $5.2 million severance pay, but he did keep his $18.4 million pension. Any company that spends more on security than absolutely necessary is immediately penalized by shareholders when its profits decrease.

Even the negative PR that Equifax is currently suffering will fade. Unless we expect data brokers to put public interest ahead of profits, the security of this industry will never improve without government regulation.

7. We need effective regulation of data brokers.

In 2014, the Federal Trade Commission recommended that Congress require data brokers be more transparent and give consumers more control over their personal information. That report contains good suggestions on how to regulate this industry.

First, Congress should help plaintiffs in data breach cases by authorizing and funding empirical research on the harm individuals receive from these breaches.

Specifically, Congress should move forward legislative proposals that establish a nationwide “credit freeze” — which is better described as changing the default for disclosure from opt-out to opt-in — and free lifetime credit monitoring services. By this I do not mean giving customers free credit-freeze options, a proposal by Senators Warren and Schatz, but that the default should be a credit freeze.

The credit card industry routinely notifies consumers when there are suspicious charges. It is obvious that credit reporting agencies should have a similar obligation to notify consumers when there is suspicious activity concerning their credit report.

On the technology side, more could be done to limit the amount of personal data companies are allowed to collect. Increasingly, privacy safeguards impose “data minimization” requirements to ensure that only the data that is actually needed is collected. On the other hand, Congress should not create a new national identifier to replace the Social Security Numbers. That would make the system of identification even more brittle. Better is to reduce dependence on systems of identification and to create contextual identification where necessary.

Finally, Congress needs to give the Federal Trade Commission the authority to set minimum security standards for data brokers and to give consumers more control over their personal information. This is essential as long as consumers are these companies’ products and not their customers.

8. Resist complaints from the industry that this is “too hard.”

The credit bureaus and data brokers, and their lobbyists and trade-association representatives, will claim that many of these measures are too hard. They’re not telling you the truth.

Take one example: credit freezes. This is an effective security measure that protects consumers, but the process of getting one and of temporarily unfreezing credit is made deliberately onerous by the credit bureaus. Why isn’t there a smartphone app that alerts me when someone wants to access my credit rating, and lets me freeze and unfreeze my credit at the touch of the screen? Too hard? Today, you can have an app on your phone that does something similar if you try to log into a computer network, or if someone tries to use your credit card at a physical location different from where you are.

Moreover, any credit bureau or data broker operating in Europe is already obligated to follow the more rigorous EU privacy laws. The EU General Data Protection Regulation will come into force, requiring even more security and privacy controls for companies collecting storing the personal data of EU citizens. Those companies have already demonstrated that they can comply with those more stringent regulations.

Credit bureaus, and data brokers in general, are deliberately not implementing these 21st-century security solutions, because they want their services to be as easy and useful as possible for their actual customers: those who are buying your information. Similarly, companies that use this personal information to open accounts are not implementing more stringent security because they want their services to be as easy-to-use and convenient as possible.

9. This has foreign trade implications.

The Canadian Broadcast Corporation reported that 100,000 Canadians had their data stolen in the Equifax breach. The British Broadcasting Corporation originally reported that 400,000 UK consumers were affected; Equifax has since revised that to 15.2 million.

Many American Internet companies have significant numbers of European users and customers, and rely on negotiated safe harbor agreements to legally collect and store personal data of EU citizens.

The European Union is in the middle of a massive regulatory shift in its privacy laws, and those agreements are coming under renewed scrutiny. Breaches such as Equifax give these European regulators a powerful argument that US privacy regulations are inadequate to protect their citizens’ data, and that they should require that data to remain in Europe. This could significantly harm American Internet companies.

10. This has national security implications.

Although it is still unknown who compromised the Equifax database, it could easily have been a foreign adversary that routinely attacks the servers of US companies and US federal agencies with the goal of exploiting security vulnerabilities and obtaining personal data.

When the Fair Credit Reporting Act was passed in 1970, the concern was that the credit bureaus might misuse our data. That is still a concern, but the world has changed since then. Credit bureaus and data brokers have far more intimate data about all of us. And it is valuable not only to companies wanting to advertise to us, but foreign governments as well. In 2015, the Chinese breached the database of the Office of Personal Management and stole the detailed security clearance information of 21 million Americans. North Korea routinely engages in cybercrime as way to fund its other activities. In a world where foreign governments use cyber capabilities to attack US assets, requiring data brokers to limit collection of personal data, securely store the data they collect, and delete data about consumers when it is no longer needed is a matter of national security.

11. We need to do something about it.

Yes, this breach is a huge black eye and a temporary stock dip for Equifax — this month. Soon, another company will have suffered a massive data breach and few will remember Equifax’s problem. Does anyone remember last year when Yahoo admitted that it exposed personal information of a billion users in 2013 and another half billion in 2014?

Unless Congress acts to protect consumer information in the digital age, these breaches will continue.

Thank you for the opportunity to testify today. I will be pleased to answer your questions.

AWS Online Tech Talks – November 2017

Post Syndicated from Sara Rodas original https://aws.amazon.com/blogs/aws/aws-online-tech-talks-november-2017/

Leaves are crunching under my boots, Halloween is tomorrow, and pumpkin is having its annual moment in the sun – it’s fall everybody! And just in time to celebrate, we have whipped up a fresh batch of pumpkin spice Tech Talks. Grab your planner (Outlook calendar) and pencil these puppies in. This month we are covering re:Invent, serverless, and everything in between.

November 2017 – Schedule

Noted below are the upcoming scheduled live, online technical sessions being held during the month of November. Make sure to register ahead of time so you won’t miss out on these free talks conducted by AWS subject matter experts.

Webinars featured this month are:

Monday, November 6

Compute

9:00 – 9:40 AM PDT: Set it and Forget it: Auto Scaling Target Tracking Policies

Tuesday, November 7

Big Data

9:00 – 9:40 AM PDT: Real-time Application Monitoring with Amazon Kinesis and Amazon CloudWatch

Compute

10:30 – 11:10 AM PDT: Simplify Microsoft Windows Server Management with Amazon Lightsail

Mobile

12:00 – 12:40 PM PDT: Deep Dive on Amazon SES What’s New

Wednesday, November 8

Databases

10:30 – 11:10 AM PDT: Migrating Your Oracle Database to PostgreSQL

Compute

12:00 – 12:40 PM PDT: Run Your CI/CD Pipeline at Scale for a Fraction of the Cost

Thursday, November 9

Databases

10:30 – 11:10 AM PDT: Migrating Your Oracle Database to PostgreSQL

Containers

9:00 – 9:40 AM PDT: Managing Container Images with Amazon ECR

Big Data

12:00 – 12:40 PM PDT: Amazon Elasticsearch Service Security Deep Dive

Monday, November 13

re:Invent

10:30 – 11:10 AM PDT: AWS re:Invent 2017: Know Before You Go

5:00 – 5:40 PM PDT: AWS re:Invent 2017: Know Before You Go

Tuesday, November 14

AI

9:00 – 9:40 AM PDT: Sentiment Analysis Using Apache MXNet and Gluon

10:30 – 11:10 AM PDT: Bringing Characters to Life with Amazon Polly Text-to-Speech

IoT

12:00 – 12:40 PM PDT: Essential Capabilities of an IoT Cloud Platform

Enterprise

2:00 – 2:40 PM PDT: Everything you wanted to know about licensing Windows workloads on AWS, but were afraid to ask

Wednesday, November 15

Security & Identity

9:00 – 9:40 AM PDT: How to Integrate AWS Directory Service with Office365

Storage

10:30 – 11:10 AM PDT: Disaster Recovery Options with AWS

Hands on Lab

12:30 – 2:00 PM PDT: Hands on Lab: Windows Workloads

Thursday, November 16

Serverless

9:00 – 9:40 AM PDT: Building Serverless Websites with [email protected]

Hands on Lab

12:30 – 2:00 PM PDT: Hands on Lab: Deploy .NET Code to AWS from Visual Studio

– Sara

Security updates for Monday

Post Syndicated from jake original https://lwn.net/Articles/737764/rss

Security updates have been issued by Arch Linux (apr, apr-util, chromium, and wget), CentOS (tomcat and tomcat6), Debian (curl, git-annex, golang, shadowsocks-libev, and wget), Fedora (libextractor and sssd), Gentoo (apache, asterisk, jython, oracle-jdk-bin, and xorg-server), openSUSE (chromium, curl, gcc48, GraphicsMagick, hostapd, kernel, libjpeg-turbo, libvirt, mysql-community-server, openvpn, SDL2, tcpdump, and wget), Oracle (tomcat and tomcat6), Red Hat (chromium-browser, tomcat, and tomcat6), Scientific Linux (tomcat and tomcat6), Slackware (php and wget), SUSE (firefox, mozilla-nss, kernel, wget, and xen), and Ubuntu (mysql-5.5, poppler, and wget).

New – Amazon EC2 Instances with Up to 8 NVIDIA Tesla V100 GPUs (P3)

Post Syndicated from Jeff Barr original https://aws.amazon.com/blogs/aws/new-amazon-ec2-instances-with-up-to-8-nvidia-tesla-v100-gpus-p3/

Driven by customer demand and made possible by on-going advances in the state-of-the-art, we’ve come a long way since the original m1.small instance that we launched in 2006, with instances that are emphasize compute power, burstable performance, memory size, local storage, and accelerated computing.

The New P3
Today we are making the next generation of GPU-powered EC2 instances available in four AWS regions. Powered by up to eight NVIDIA Tesla V100 GPUs, the P3 instances are designed to handle compute-intensive machine learning, deep learning, computational fluid dynamics, computational finance, seismic analysis, molecular modeling, and genomics workloads.

P3 instances use customized Intel Xeon E5-2686v4 processors running at up to 2.7 GHz. They are available in three sizes (all VPC-only and EBS-only):

Model NVIDIA Tesla V100 GPUs GPU Memory NVIDIA NVLink vCPUs Main Memory Network Bandwidth EBS Bandwidth
p3.2xlarge 1 16 GiB n/a 8 61 GiB Up to 10 Gbps 1.5 Gbps
p3.8xlarge 4 64 GiB 200 GBps 32 244 GiB 10 Gbps 7 Gbps
p3.16xlarge 8 128 GiB 300 GBps 64 488 GiB 25 Gbps 14 Gbps

Each of the NVIDIA GPUs is packed with 5,120 CUDA cores and another 640 Tensor cores and can deliver up to 125 TFLOPS of mixed-precision floating point, 15.7 TFLOPS of single-precision floating point, and 7.8 TFLOPS of double-precision floating point. On the two larger sizes, the GPUs are connected together via NVIDIA NVLink 2.0 running at a total data rate of up to 300 GBps. This allows the GPUs to exchange intermediate results and other data at high speed, without having to move it through the CPU or the PCI-Express fabric.

What’s a Tensor Core?
I had not heard the term Tensor core before starting to write this post. According to this very helpful post on the NVIDIA Blog, Tensor cores are designed to speed up the training and inference of large, deep neural networks. Each core is able to quickly and efficiently multiply a pair of 4×4 half-precision (also known as FP16) matrices together, add the resulting 4×4 matrix to another half or single-precision (FP32) matrix, and store the resulting 4×4 matrix in either half or single-precision form. Here’s a diagram from NVIDIA’s blog post:

This operation is in the innermost loop of the training process for a deep neural network, and is an excellent example of how today’s NVIDIA GPU hardware is purpose-built to address a very specific market need. By the way, the mixed-precision qualifier on the Tensor core performance means that it is flexible enough to work with with a combination of 16-bit and 32-bit floating point values.

Performance in Perspective
I always like to put raw performance numbers into a real-world perspective so that they are easier to relate to and (hopefully) more meaningful. This turned out to be surprisingly difficult, given that the eight NVIDIA Tesla V100 GPUs on a single p3.16xlarge can do 125 trillion single-precision floating point multiplications per second.

Let’s go back to the dawn of the microprocessor era, and consider the Intel 8080A chip that powered the MITS Altair that I bought in the summer of 1977. With a 2 MHz clock, it was able to do about 832 multiplications per second (I used this data and corrected it for the faster clock speed). The p3.16xlarge is roughly 150 billion times faster. However, just 1.2 billion seconds have gone by since that summer. In other words, I can do 100x more calculations today in one second than my Altair could have done in the last 40 years!

What about the innovative 8087 math coprocessor that was an optional accessory for the IBM PC that was announced in the summer of 1981? With a 5 MHz clock and purpose-built hardware, it was able to do about 52,632 multiplications per second. 1.14 billion seconds have elapsed since then, p3.16xlarge is 2.37 billion times faster, so the poor little PC would be barely halfway through a calculation that would run for 1 second today.

Ok, how about a Cray-1? First delivered in 1976, this supercomputer was able to perform vector operations at 160 MFLOPS, making the p3.x16xlarge 781,000 times faster. It could have iterated on some interesting problem 1500 times over the years since it was introduced.

Comparisons between the P3 and today’s scale-out supercomputers are harder to make, given that you can think of the P3 as a step-and-repeat component of a supercomputer that you can launch on as as-needed basis.

Run One Today
In order to take full advantage of the NVIDIA Tesla V100 GPUs and the Tensor cores, you will need to use CUDA 9 and cuDNN7. These drivers and libraries have already been added to the newest versions of the Windows AMIs and will be included in an updated Amazon Linux AMI that is scheduled for release on November 7th. New packages are already available in our repos if you want to to install them on your existing Amazon Linux AMI.

The newest AWS Deep Learning AMIs come preinstalled with the latest releases of Apache MxNet, Caffe2, and Tensorflow (each with support for the NVIDIA Tesla V100 GPUs), and will be updated to support P3 instances with other machine learning frameworks such as Microsoft Cognitive Toolkit and PyTorch as soon as these frameworks release support for the NVIDIA Tesla V100 GPUs. You can also use the NVIDIA Volta Deep Learning AMI for NGC.

P3 instances are available in the US East (Northern Virginia), US West (Oregon), EU (Ireland), and Asia Pacific (Tokyo) Regions in On-Demand, Spot, Reserved Instance, and Dedicated Host form.

Jeff;

 

Bringing Datacenter-Scale Hardware-Software Co-design to the Cloud with FireSim and Amazon EC2 F1 Instances

Post Syndicated from Mia Champion original https://aws.amazon.com/blogs/compute/bringing-datacenter-scale-hardware-software-co-design-to-the-cloud-with-firesim-and-amazon-ec2-f1-instances/

The recent addition of Xilinx FPGAs to AWS Cloud compute offerings is one way that AWS is enabling global growth in the areas of advanced analytics, deep learning and AI. The customized F1 servers use pooled accelerators, enabling interconnectivity of up to 8 FPGAs, each one including 64 GiB DDR4 ECC protected memory, with a dedicated PCIe x16 connection. That makes this a powerful engine with the capacity to process advanced analytical applications at scale, at a significantly faster rate. For example, AWS commercial partner Edico Genome is able to achieve an approximately 30X speedup in analyzing whole genome sequencing datasets using their DRAGEN platform powered with F1 instances.

While the availability of FPGA F1 compute on-demand provides clear accessibility and cost advantages, many mainstream users are still finding that the “threshold to entry” in developing or running FPGA-accelerated simulations is too high. Researchers at the UC Berkeley RISE Lab have developed “FireSim”, powered by Amazon FPGA F1 instances as an open-source resource, FireSim lowers that entry bar and makes it easier for everyone to leverage the power of an FPGA-accelerated compute environment. Whether you are part of a small start-up development team or working at a large datacenter scale, hardware-software co-design enables faster time-to-deployment, lower costs, and more predictable performance. We are excited to feature FireSim in this post from Sagar Karandikar and his colleagues at UC-Berkeley.

―Mia Champion, Sr. Data Scientist, AWS

Mapping an 8-node FireSim cluster simulation to Amazon EC2 F1

As traditional hardware scaling nears its end, the data centers of tomorrow are trending towards heterogeneity, employing custom hardware accelerators and increasingly high-performance interconnects. Prototyping new hardware at scale has traditionally been either extremely expensive, or very slow. In this post, I introduce FireSim, a new hardware simulation platform under development in the computer architecture research group at UC Berkeley that enables fast, scalable hardware simulation using Amazon EC2 F1 instances.

FireSim benefits both hardware and software developers working on new rack-scale systems: software developers can use the simulated nodes with new hardware features as they would use a real machine, while hardware developers have full control over the hardware being simulated and can run real software stacks while hardware is still under development. In conjunction with this post, we’re releasing the first public demo of FireSim, which lets you deploy your own 8-node simulated cluster on an F1 Instance and run benchmarks against it. This demo simulates a pre-built “vanilla” cluster, but demonstrates FireSim’s high performance and usability.

Why FireSim + F1?

FPGA-accelerated hardware simulation is by no means a new concept. However, previous attempts to use FPGAs for simulation have been fraught with usability, scalability, and cost issues. FireSim takes advantage of EC2 F1 and open-source hardware to address the traditional problems with FPGA-accelerated simulation:
Problem #1: FPGA-based simulations have traditionally been expensive, difficult to deploy, and difficult to reproduce.
FireSim uses public-cloud infrastructure like F1, which means no upfront cost to purchase and deploy FPGAs. Developers and researchers can distribute pre-built AMIs and AFIs, as in this public demo (more details later in this post), to make experiments easy to reproduce. FireSim also automates most of the work involved in deploying an FPGA simulation, essentially enabling one-click conversion from new RTL to deploying on an FPGA cluster.

Problem #2: FPGA-based simulations have traditionally been difficult (and expensive) to scale.
Because FireSim uses F1, users can scale out experiments by spinning up additional EC2 instances, rather than spending hundreds of thousands of dollars on large FPGA clusters.

Problem #3: Finding open hardware to simulate has traditionally been difficult. Finding open hardware that can run real software stacks is even harder.
FireSim simulates RocketChip, an open, silicon-proven, RISC-V-based processor platform, and adds peripherals like a NIC and disk device to build up a realistic system. Processors that implement RISC-V automatically support real operating systems (such as Linux) and even support applications like Apache and Memcached. We provide a custom Buildroot-based FireSim Linux distribution that runs on our simulated nodes and includes many popular developer tools.

Problem #4: Writing hardware in traditional HDLs is time-consuming.
Both FireSim and RocketChip use the Chisel HDL, which brings modern programming paradigms to hardware description languages. Chisel greatly simplifies the process of building large, highly parameterized hardware components.

How to use FireSim for hardware/software co-design

FireSim drastically improves the process of co-designing hardware and software by acting as a push-button interface for collaboration between hardware developers and systems software developers. The following diagram describes the workflows that hardware and software developers use when working with FireSim.

Figure 2. The FireSim custom hardware development workflow.

The hardware developer’s view:

  1. Write custom RTL for your accelerator, peripheral, or processor modification in a productive language like Chisel.
  2. Run a software simulation of your hardware design in standard gate-level simulation tools for early-stage debugging.
  3. Run FireSim build scripts, which automatically build your simulation, run it through the Vivado toolchain/AWS shell scripts, and publish an AFI.
  4. Deploy your simulation on EC2 F1 using the generated simulation driver and AFI
  5. Run real software builds released by software developers to benchmark your hardware

The software developer’s view:

  1. Deploy the AMI/AFI generated by the hardware developer on an F1 instance to simulate a cluster of nodes (or scale out to many F1 nodes for larger simulated core-counts).
  2. Connect using SSH into the simulated nodes in the cluster and boot the Linux distribution included with FireSim. This distribution is easy to customize, and already supports many standard software packages.
  3. Directly prototype your software using the same exact interfaces that the software will see when deployed on the real future system you’re prototyping, with the same performance characteristics as observed from software, even at scale.

FireSim demo v1.0

Figure 3. Cluster topology simulated by FireSim demo v1.0.

This first public demo of FireSim focuses on the aforementioned “software-developer’s view” of the custom hardware development cycle. The demo simulates a cluster of 1 to 8 RocketChip-based nodes, interconnected by a functional network simulation. The simulated nodes work just like “real” machines:  they boot Linux, you can connect to them using SSH, and you can run real applications on top. The nodes can see each other (and the EC2 F1 instance on which they’re deployed) on the network and communicate with one another. While the demo currently simulates a pre-built “vanilla” cluster, the entire hardware configuration of these simulated nodes can be modified after FireSim is open-sourced.

In this post, I walk through bringing up a single-node FireSim simulation for experienced EC2 F1 users. For more detailed instructions for new users and instructions for running a larger 8-node simulation, see FireSim Demo v1.0 on Amazon EC2 F1. Both demos walk you through setting up an instance from a demo AMI/AFI and booting Linux on the simulated nodes. The full demo instructions also walk you through an example workload, running Memcached on the simulated nodes, with YCSB as a load generator to demonstrate network functionality.

Deploying the demo on F1

In this release, we provide pre-built binaries for driving simulation from the host and a pre-built AFI that contains the FPGA infrastructure necessary to simulate a RocketChip-based node.

Starting your F1 instances

First, launch an instance using the free FireSim Demo v1.0 product available on the AWS Marketplace on an f1.2xlarge instance. After your instance has booted, log in using the user name centos. On the first login, you should see the message “FireSim network config completed.” This sets up the necessary tap interfaces and bridge on the EC2 instance to enable communicating with the simulated nodes.

AMI contents

The AMI contains a variety of tools to help you run simulations and build software for RISC-V systems, including the riscv64 toolchain, a Buildroot-based Linux distribution that runs on the simulated nodes, and the simulation driver program. For more details, see the AMI Contents section on the FireSim website.

Single-node demo

First, you need to flash the FPGA with the FireSim AFI. To do so, run:

[[email protected]_ADDR ~]$ sudo fpga-load-local-image -S 0 -I agfi-00a74c2d615134b21

To start a simulation, run the following at the command line:

[[email protected]_ADDR ~]$ boot-firesim-singlenode

This automatically calls the simulation driver, telling it to load the Linux kernel image and root filesystem for the Linux distro. This produces output similar to the following:

Simulations Started. You can use the UART console of each simulated node by attaching to the following screens:

There is a screen on:

2492.fsim0      (Detached)

1 Socket in /var/run/screen/S-centos.

You could connect to the simulated UART console by connecting to this screen, but instead opt to use SSH to access the node instead.

First, ping the node to make sure it has come online. This is currently required because nodes may get stuck at Linux boot if the NIC does not receive any network traffic. For more information, see Troubleshooting/Errata. The node is always assigned the IP address 192.168.1.10:

[[email protected]_ADDR ~]$ ping 192.168.1.10

This should eventually produce the following output:

PING 192.168.1.10 (192.168.1.10) 56(84) bytes of data.

From 192.168.1.1 icmp_seq=1 Destination Host Unreachable

64 bytes from 192.168.1.10: icmp_seq=1 ttl=64 time=2017 ms

64 bytes from 192.168.1.10: icmp_seq=2 ttl=64 time=1018 ms

64 bytes from 192.168.1.10: icmp_seq=3 ttl=64 time=19.0 ms

At this point, you know that the simulated node is online. You can connect to it using SSH with the user name root and password firesim. It is also convenient to make sure that your TERM variable is set correctly. In this case, the simulation expects TERM=linux, so provide that:

[[email protected]_ADDR ~]$ TERM=linux ssh [email protected]

The authenticity of host ‘192.168.1.10 (192.168.1.10)’ can’t be established.

ECDSA key fingerprint is 63:e9:66:d0:5c:06:2c:1d:5c:95:33:c8:36:92:30:49.

Are you sure you want to continue connecting (yes/no)? yes

Warning: Permanently added ‘192.168.1.10’ (ECDSA) to the list of known hosts.

[email protected]’s password:

#

At this point, you’re connected to the simulated node. Run uname -a as an example. You should see the following output, indicating that you’re connected to a RISC-V system:

# uname -a

Linux buildroot 4.12.0-rc2 #1 Fri Aug 4 03:44:55 UTC 2017 riscv64 GNU/Linux

Now you can run programs on the simulated node, as you would with a real machine. For an example workload (running YCSB against Memcached on the simulated node) or to run a larger 8-node simulation, see the full FireSim Demo v1.0 on Amazon EC2 F1 demo instructions.

Finally, when you are finished, you can shut down the simulated node by running the following command from within the simulated node:

# poweroff

You can confirm that the simulation has ended by running screen -ls, which should now report that there are no detached screens.

Future plans

At Berkeley, we’re planning to keep improving the FireSim platform to enable our own research in future data center architectures, like FireBox. The FireSim platform will eventually support more sophisticated processors, custom accelerators (such as Hwacha), network models, and peripherals, in addition to scaling to larger numbers of FPGAs. In the future, we’ll open source the entire platform, including Midas, the tool used to transform RTL into FPGA simulators, allowing users to modify any part of the hardware/software stack. Follow @firesimproject on Twitter to stay tuned to future FireSim updates.

Acknowledgements

FireSim is the joint work of many students and faculty at Berkeley: Sagar Karandikar, Donggyu Kim, Howard Mao, David Biancolin, Jack Koenig, Jonathan Bachrach, and Krste Asanović. This work is partially funded by AWS through the RISE Lab, by the Intel Science and Technology Center for Agile HW Design, and by ASPIRE Lab sponsors and affiliates Intel, Google, HPE, Huawei, NVIDIA, and SK hynix.

Apache OpenOffice 4.1.4 released

Post Syndicated from corbet original https://lwn.net/Articles/736898/rss

The OpenOffice
4.1.4 release
is finally available; see this article for some background on this
release. The announcement is all bright and sunny, but a look at the
August 16 Apache board minutes
shows concern about the state of
the project. Indeed, the OpenOffice project management committee was,
according to these minutes, supposed to post an announcement about the
state of the project; it would appear that has not yet happened.

Implementing Default Directory Indexes in Amazon S3-backed Amazon CloudFront Origins Using [email protected]

Post Syndicated from Ronnie Eichler original https://aws.amazon.com/blogs/compute/implementing-default-directory-indexes-in-amazon-s3-backed-amazon-cloudfront-origins-using-lambdaedge/

With the recent launch of [email protected], it’s now possible for you to provide even more robust functionality to your static websites. Amazon CloudFront is a content distribution network service. In this post, I show how you can use [email protected] along with the CloudFront origin access identity (OAI) for Amazon S3 and still provide simple URLs (such as www.example.com/about/ instead of www.example.com/about/index.html).

Background

Amazon S3 is a great platform for hosting a static website. You don’t need to worry about managing servers or underlying infrastructure—you just publish your static to content to an S3 bucket. S3 provides a DNS name such as <bucket-name>.s3-website-<AWS-region>.amazonaws.com. Use this name for your website by creating a CNAME record in your domain’s DNS environment (or Amazon Route 53) as follows:

www.example.com -> <bucket-name>.s3-website-<AWS-region>.amazonaws.com

You can also put CloudFront in front of S3 to further scale the performance of your site and cache the content closer to your users. CloudFront can enable HTTPS-hosted sites, by either using a custom Secure Sockets Layer (SSL) certificate or a managed certificate from AWS Certificate Manager. In addition, CloudFront also offers integration with AWS WAF, a web application firewall. As you can see, it’s possible to achieve some robust functionality by using S3, CloudFront, and other managed services and not have to worry about maintaining underlying infrastructure.

One of the key concerns that you might have when implementing any type of WAF or CDN is that you want to force your users to go through the CDN. If you implement CloudFront in front of S3, you can achieve this by using an OAI. However, in order to do this, you cannot use the HTTP endpoint that is exposed by S3’s static website hosting feature. Instead, CloudFront must use the S3 REST endpoint to fetch content from your origin so that the request can be authenticated using the OAI. This presents some challenges in that the REST endpoint does not support redirection to a default index page.

CloudFront does allow you to specify a default root object (index.html), but it only works on the root of the website (such as http://www.example.com > http://www.example.com/index.html). It does not work on any subdirectory (such as http://www.example.com/about/). If you were to attempt to request this URL through CloudFront, CloudFront would do a S3 GetObject API call against a key that does not exist.

Of course, it is a bad user experience to expect users to always type index.html at the end of every URL (or even know that it should be there). Until now, there has not been an easy way to provide these simpler URLs (equivalent to the DirectoryIndex Directive in an Apache Web Server configuration) to users through CloudFront. Not if you still want to be able to restrict access to the S3 origin using an OAI. However, with the release of [email protected], you can use a JavaScript function running on the CloudFront edge nodes to look for these patterns and request the appropriate object key from the S3 origin.

Solution

In this example, you use the compute power at the CloudFront edge to inspect the request as it’s coming in from the client. Then re-write the request so that CloudFront requests a default index object (index.html in this case) for any request URI that ends in ‘/’.

When a request is made against a web server, the client specifies the object to obtain in the request. You can use this URI and apply a regular expression to it so that these URIs get resolved to a default index object before CloudFront requests the object from the origin. Use the following code:

'use strict';
exports.handler = (event, context, callback) => {
    
    // Extract the request from the CloudFront event that is sent to [email protected] 
    var request = event.Records[0].cf.request;

    // Extract the URI from the request
    var olduri = request.uri;

    // Match any '/' that occurs at the end of a URI. Replace it with a default index
    var newuri = olduri.replace(/\/$/, '\/index.html');
    
    // Log the URI as received by CloudFront and the new URI to be used to fetch from origin
    console.log("Old URI: " + olduri);
    console.log("New URI: " + newuri);
    
    // Replace the received URI with the URI that includes the index page
    request.uri = newuri;
    
    // Return to CloudFront
    return callback(null, request);

};

To get started, create an S3 bucket to be the origin for CloudFront:

Create bucket

On the other screens, you can just accept the defaults for the purposes of this walkthrough. If this were a production implementation, I would recommend enabling bucket logging and specifying an existing S3 bucket as the destination for access logs. These logs can be useful if you need to troubleshoot issues with your S3 access.

Now, put some content into your S3 bucket. For this walkthrough, create two simple webpages to demonstrate the functionality:  A page that resides at the website root, and another that is in a subdirectory.

<s3bucketname>/index.html

<!doctype html>
<html>
    <head>
        <meta charset="utf-8">
        <title>Root home page</title>
    </head>
    <body>
        <p>Hello, this page resides in the root directory.</p>
    </body>
</html>

<s3bucketname>/subdirectory/index.html

<!doctype html>
<html>
    <head>
        <meta charset="utf-8">
        <title>Subdirectory home page</title>
    </head>
    <body>
        <p>Hello, this page resides in the /subdirectory/ directory.</p>
    </body>
</html>

When uploading the files into S3, you can accept the defaults. You add a bucket policy as part of the CloudFront distribution creation that allows CloudFront to access the S3 origin. You should now have an S3 bucket that looks like the following:

Root of bucket

Subdirectory in bucket

Next, create a CloudFront distribution that your users will use to access the content. Open the CloudFront console, and choose Create Distribution. For Select a delivery method for your content, under Web, choose Get Started.

On the next screen, you set up the distribution. Below are the options to configure:

  • Origin Domain Name:  Select the S3 bucket that you created earlier.
  • Restrict Bucket Access: Choose Yes.
  • Origin Access Identity: Create a new identity.
  • Grant Read Permissions on Bucket: Choose Yes, Update Bucket Policy.
  • Object Caching: Choose Customize (I am changing the behavior to avoid having CloudFront cache objects, as this could affect your ability to troubleshoot while implementing the Lambda code).
    • Minimum TTL: 0
    • Maximum TTL: 0
    • Default TTL: 0

You can accept all of the other defaults. Again, this is a proof-of-concept exercise. After you are comfortable that the CloudFront distribution is working properly with the origin and Lambda code, you can re-visit the preceding values and make changes before implementing it in production.

CloudFront distributions can take several minutes to deploy (because the changes have to propagate out to all of the edge locations). After that’s done, test the functionality of the S3-backed static website. Looking at the distribution, you can see that CloudFront assigns a domain name:

CloudFront Distribution Settings

Try to access the website using a combination of various URLs:

http://<domainname>/:  Works

› curl -v http://d3gt20ea1hllb.cloudfront.net/
*   Trying 54.192.192.214...
* TCP_NODELAY set
* Connected to d3gt20ea1hllb.cloudfront.net (54.192.192.214) port 80 (#0)
> GET / HTTP/1.1
> Host: d3gt20ea1hllb.cloudfront.net
> User-Agent: curl/7.51.0
> Accept: */*
>
< HTTP/1.1 200 OK
< ETag: "cb7e2634fe66c1fd395cf868087dd3b9"
< Accept-Ranges: bytes
< Server: AmazonS3
< X-Cache: Miss from cloudfront
< X-Amz-Cf-Id: -D2FSRwzfcwyKZKFZr6DqYFkIf4t7HdGw2MkUF5sE6YFDxRJgi0R1g==
< Content-Length: 209
< Content-Type: text/html
< Last-Modified: Wed, 19 Jul 2017 19:21:16 GMT
< Via: 1.1 6419ba8f3bd94b651d416054d9416f1e.cloudfront.net (CloudFront), 1.1 iad6-proxy-3.amazon.com:80 (Cisco-WSA/9.1.2-010)
< Connection: keep-alive
<
<!doctype html>
<html>
    <head>
        <meta charset="utf-8">
        <title>Root home page</title>
    </head>
    <body>
        <p>Hello, this page resides in the root directory.</p>
    </body>
</html>
* Curl_http_done: called premature == 0
* Connection #0 to host d3gt20ea1hllb.cloudfront.net left intact

This is because CloudFront is configured to request a default root object (index.html) from the origin.

http://<domainname>/subdirectory/:  Doesn’t work

› curl -v http://d3gt20ea1hllb.cloudfront.net/subdirectory/
*   Trying 54.192.192.214...
* TCP_NODELAY set
* Connected to d3gt20ea1hllb.cloudfront.net (54.192.192.214) port 80 (#0)
> GET /subdirectory/ HTTP/1.1
> Host: d3gt20ea1hllb.cloudfront.net
> User-Agent: curl/7.51.0
> Accept: */*
>
< HTTP/1.1 200 OK
< ETag: "d41d8cd98f00b204e9800998ecf8427e"
< x-amz-server-side-encryption: AES256
< Accept-Ranges: bytes
< Server: AmazonS3
< X-Cache: Miss from cloudfront
< X-Amz-Cf-Id: Iqf0Gy8hJLiW-9tOAdSFPkL7vCWBrgm3-1ly5tBeY_izU82ftipodA==
< Content-Length: 0
< Content-Type: application/x-directory
< Last-Modified: Wed, 19 Jul 2017 19:21:24 GMT
< Via: 1.1 6419ba8f3bd94b651d416054d9416f1e.cloudfront.net (CloudFront), 1.1 iad6-proxy-3.amazon.com:80 (Cisco-WSA/9.1.2-010)
< Connection: keep-alive
<
* Curl_http_done: called premature == 0
* Connection #0 to host d3gt20ea1hllb.cloudfront.net left intact

If you use a tool such like cURL to test this, you notice that CloudFront and S3 are returning a blank response. The reason for this is that the subdirectory does exist, but it does not resolve to an S3 object. Keep in mind that S3 is an object store, so there are no real directories. User interfaces such as the S3 console present a hierarchical view of a bucket with folders based on the presence of forward slashes, but behind the scenes the bucket is just a collection of keys that represent stored objects.

http://<domainname>/subdirectory/index.html:  Works

› curl -v http://d3gt20ea1hllb.cloudfront.net/subdirectory/index.html
*   Trying 54.192.192.130...
* TCP_NODELAY set
* Connected to d3gt20ea1hllb.cloudfront.net (54.192.192.130) port 80 (#0)
> GET /subdirectory/index.html HTTP/1.1
> Host: d3gt20ea1hllb.cloudfront.net
> User-Agent: curl/7.51.0
> Accept: */*
>
< HTTP/1.1 200 OK
< Date: Thu, 20 Jul 2017 20:35:15 GMT
< ETag: "ddf87c487acf7cef9d50418f0f8f8dae"
< Accept-Ranges: bytes
< Server: AmazonS3
< X-Cache: RefreshHit from cloudfront
< X-Amz-Cf-Id: bkh6opXdpw8pUomqG3Qr3UcjnZL8axxOH82Lh0OOcx48uJKc_Dc3Cg==
< Content-Length: 227
< Content-Type: text/html
< Last-Modified: Wed, 19 Jul 2017 19:21:45 GMT
< Via: 1.1 3f2788d309d30f41de96da6f931d4ede.cloudfront.net (CloudFront), 1.1 iad6-proxy-3.amazon.com:80 (Cisco-WSA/9.1.2-010)
< Connection: keep-alive
<
<!doctype html>
<html>
    <head>
        <meta charset="utf-8">
        <title>Subdirectory home page</title>
    </head>
    <body>
        <p>Hello, this page resides in the /subdirectory/ directory.</p>
    </body>
</html>
* Curl_http_done: called premature == 0
* Connection #0 to host d3gt20ea1hllb.cloudfront.net left intact

This request works as expected because you are referencing the object directly. Now, you implement the [email protected] function to return the default index.html page for any subdirectory. Looking at the example JavaScript code, here’s where the magic happens:

var newuri = olduri.replace(/\/$/, '\/index.html');

You are going to use a JavaScript regular expression to match any ‘/’ that occurs at the end of the URI and replace it with ‘/index.html’. This is the equivalent to what S3 does on its own with static website hosting. However, as I mentioned earlier, you can’t rely on this if you want to use a policy on the bucket to restrict it so that users must access the bucket through CloudFront. That way, all requests to the S3 bucket must be authenticated using the S3 REST API. Because of this, you implement a [email protected] function that takes any client request ending in ‘/’ and append a default ‘index.html’ to the request before requesting the object from the origin.

In the Lambda console, choose Create function. On the next screen, skip the blueprint selection and choose Author from scratch, as you’ll use the sample code provided.

Next, configure the trigger. Choosing the empty box shows a list of available triggers. Choose CloudFront and select your CloudFront distribution ID (created earlier). For this example, leave Cache Behavior as * and CloudFront Event as Origin Request. Select the Enable trigger and replicate box and choose Next.

Lambda Trigger

Next, give the function a name and a description. Then, copy and paste the following code:

'use strict';
exports.handler = (event, context, callback) => {
    
    // Extract the request from the CloudFront event that is sent to [email protected] 
    var request = event.Records[0].cf.request;

    // Extract the URI from the request
    var olduri = request.uri;

    // Match any '/' that occurs at the end of a URI. Replace it with a default index
    var newuri = olduri.replace(/\/$/, '\/index.html');
    
    // Log the URI as received by CloudFront and the new URI to be used to fetch from origin
    console.log("Old URI: " + olduri);
    console.log("New URI: " + newuri);
    
    // Replace the received URI with the URI that includes the index page
    request.uri = newuri;
    
    // Return to CloudFront
    return callback(null, request);

};

Next, define a role that grants permissions to the Lambda function. For this example, choose Create new role from template, Basic Edge Lambda permissions. This creates a new IAM role for the Lambda function and grants the following permissions:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "logs:CreateLogGroup",
                "logs:CreateLogStream",
                "logs:PutLogEvents"
            ],
            "Resource": [
                "arn:aws:logs:*:*:*"
            ]
        }
    ]
}

In a nutshell, these are the permissions that the function needs to create the necessary CloudWatch log group and log stream, and to put the log events so that the function is able to write logs when it executes.

After the function has been created, you can go back to the browser (or cURL) and re-run the test for the subdirectory request that failed previously:

› curl -v http://d3gt20ea1hllb.cloudfront.net/subdirectory/
*   Trying 54.192.192.202...
* TCP_NODELAY set
* Connected to d3gt20ea1hllb.cloudfront.net (54.192.192.202) port 80 (#0)
> GET /subdirectory/ HTTP/1.1
> Host: d3gt20ea1hllb.cloudfront.net
> User-Agent: curl/7.51.0
> Accept: */*
>
< HTTP/1.1 200 OK
< Date: Thu, 20 Jul 2017 21:18:44 GMT
< ETag: "ddf87c487acf7cef9d50418f0f8f8dae"
< Accept-Ranges: bytes
< Server: AmazonS3
< X-Cache: Miss from cloudfront
< X-Amz-Cf-Id: rwFN7yHE70bT9xckBpceTsAPcmaadqWB9omPBv2P6WkIfQqdjTk_4w==
< Content-Length: 227
< Content-Type: text/html
< Last-Modified: Wed, 19 Jul 2017 19:21:45 GMT
< Via: 1.1 3572de112011f1b625bb77410b0c5cca.cloudfront.net (CloudFront), 1.1 iad6-proxy-3.amazon.com:80 (Cisco-WSA/9.1.2-010)
< Connection: keep-alive
<
<!doctype html>
<html>
    <head>
        <meta charset="utf-8">
        <title>Subdirectory home page</title>
    </head>
    <body>
        <p>Hello, this page resides in the /subdirectory/ directory.</p>
    </body>
</html>
* Curl_http_done: called premature == 0
* Connection #0 to host d3gt20ea1hllb.cloudfront.net left intact

You have now configured a way for CloudFront to return a default index page for subdirectories in S3!

Summary

In this post, you used [email protected] to be able to use CloudFront with an S3 origin access identity and serve a default root object on subdirectory URLs. To find out some more about this use-case, see [email protected] integration with CloudFront in our documentation.

If you have questions or suggestions, feel free to comment below. For troubleshooting or implementation help, check out the Lambda forum.

ACME Support in Apache HTTP Server Project

Post Syndicated from ris original https://lwn.net/Articles/736668/rss

Let’s Encrypt has announced
that Automatic Certificate Management Environment (ACME) protocol support
is being integrated into the Apache HTTP Server (httpd). “ACME support being built in to one of the world’s most popular Web servers, Apache httpd, is great because it means that deploying HTTPS will be even easier for millions of websites. It’s a huge step towards delivering the ideal certificate issuance and management experience to as many people as possible.

ACME Support in Apache HTTP Server Project

Post Syndicated from Let's Encrypt - Free SSL/TLS Certificates original https://letsencrypt.org//2017/10/17/acme-support-in-apache-httpd.html

We’re excited that support for getting and managing TLS certificates via the ACME protocol is coming to the Apache HTTP Server Project (httpd). ACME is the protocol used by Let’s Encrypt, and hopefully other Certificate Authorities in the future. We anticipate this feature will significantly aid the adoption of HTTPS for new and existing websites.

We created Let’s Encrypt in order to make getting and managing TLS certificates as simple as possible. For Let’s Encrypt subscribers, this usually means obtaining an ACME client and executing some simple commands. Ultimately though, we’d like for most Let’s Encrypt subscribers to have ACME clients built in to their server software so that obtaining an additional piece of software is not necessary. The less work people have to do to deploy HTTPS the better!

ACME support being built in to one of the world’s most popular Web servers, Apache httpd, is great because it means that deploying HTTPS will be even easier for millions of websites. It’s a huge step towards delivering the ideal certificate issuance and management experience to as many people as possible.

The Apache httpd ACME module is called mod_md. It’s currently in the development version of httpd and a plan is being formulated to backport it to an httpd 2.4.x stable release. The mod_md code is also available on GitHub.

It’s also worth mentioning that the development version of Apache httpd now includes support for an SSLPolicy directive. Properly configuring TLS has traditionally involved making a large number of complex choices. With the SSLPolicy directive, admins simply select a modern, intermediate, or old TLS configuration, and sensible choices will be made for them.

Development of mod_md and the SSLPolicy directive has been funded by Mozilla and carried out primarily by Stefan Eissing of greenbytes. Thank you Mozilla and Stefan!

Let’s Encrypt is currently providing certificates for more than 55 million websites. We look forward to being able to serve even more websites as efforts like this make deploying HTTPS with Let’s Encrypt even easier. If you’re as excited about the potential for a 100% HTTPS Web as we are, please consider getting involved, making a donation, or sponsoring Let’s Encrypt.

Predict Billboard Top 10 Hits Using RStudio, H2O and Amazon Athena

Post Syndicated from Gopal Wunnava original https://aws.amazon.com/blogs/big-data/predict-billboard-top-10-hits-using-rstudio-h2o-and-amazon-athena/

Success in the popular music industry is typically measured in terms of the number of Top 10 hits artists have to their credit. The music industry is a highly competitive multi-billion dollar business, and record labels incur various costs in exchange for a percentage of the profits from sales and concert tickets.

Predicting the success of an artist’s release in the popular music industry can be difficult. One release may be extremely popular, resulting in widespread play on TV, radio and social media, while another single may turn out quite unpopular, and therefore unprofitable. Record labels need to be selective in their decision making, and predictive analytics can help them with decision making around the type of songs and artists they need to promote.

In this walkthrough, you leverage H2O.ai, Amazon Athena, and RStudio to make predictions on whether a song might make it to the Top 10 Billboard charts. You explore the GLM, GBM, and deep learning modeling techniques using H2O’s rapid, distributed and easy-to-use open source parallel processing engine. RStudio is a popular IDE, licensed either commercially or under AGPLv3, for working with R. This is ideal if you don’t want to connect to a server via SSH and use code editors such as vi to do analytics. RStudio is available in a desktop version, or a server version that allows you to access R via a web browser. RStudio’s Notebooks feature is used to demonstrate the execution of code and output. In addition, this post showcases how you can leverage Athena for query and interactive analysis during the modeling phase. A working knowledge of statistics and machine learning would be helpful to interpret the analysis being performed in this post.

Walkthrough

Your goal is to predict whether a song will make it to the Top 10 Billboard charts. For this purpose, you will be using multiple modeling techniques―namely GLM, GBM and deep learning―and choose the model that is the best fit.

This solution involves the following steps:

  • Install and configure RStudio with Athena
  • Log in to RStudio
  • Install R packages
  • Connect to Athena
  • Create a dataset
  • Create models

Install and configure RStudio with Athena

Use the following AWS CloudFormation stack to install, configure, and connect RStudio on an Amazon EC2 instance with Athena.

Launching this stack creates all required resources and prerequisites:

  • Amazon EC2 instance with Amazon Linux (minimum size of t2.large is recommended)
  • Provisioning of the EC2 instance in an existing VPC and public subnet
  • Installation of Java 8
  • Assignment of an IAM role to the EC2 instance with the required permissions for accessing Athena and Amazon S3
  • Security group allowing access to the RStudio and SSH ports from the internet (I recommend restricting access to these ports)
  • S3 staging bucket required for Athena (referenced within RStudio as ATHENABUCKET)
  • RStudio username and password
  • Setup logs in Amazon CloudWatch Logs (if needed for additional troubleshooting)
  • Amazon EC2 Systems Manager agent, which makes it easy to manage and patch

All AWS resources are created in the US-East-1 Region. To avoid cross-region data transfer fees, launch the CloudFormation stack in the same region. To check the availability of Athena in other regions, see Region Table.

Log in to RStudio

The instance security group has been automatically configured to allow incoming connections on the RStudio port 8787 from any source internet address. You can edit the security group to restrict source IP access. If you have trouble connecting, ensure that port 8787 isn’t blocked by subnet network ACLS or by your outgoing proxy/firewall.

  1. In the CloudFormation stack, choose Outputs, Value, and then open the RStudio URL. You might need to wait for a few minutes until the instance has been launched.
  2. Log in to RStudio with the and password you provided during setup.

Install R packages

Next, install the required R packages from the RStudio console. You can download the R notebook file containing just the code.

#install pacman – a handy package manager for managing installs
if("pacman" %in% rownames(installed.packages()) == FALSE)
{install.packages("pacman")}  
library(pacman)
p_load(h2o,rJava,RJDBC,awsjavasdk)
h2o.init(nthreads = -1)
##  Connection successful!
## 
## R is connected to the H2O cluster: 
##     H2O cluster uptime:         2 hours 42 minutes 
##     H2O cluster version:        3.10.4.6 
##     H2O cluster version age:    4 months and 4 days !!! 
##     H2O cluster name:           H2O_started_from_R_rstudio_hjx881 
##     H2O cluster total nodes:    1 
##     H2O cluster total memory:   3.30 GB 
##     H2O cluster total cores:    4 
##     H2O cluster allowed cores:  4 
##     H2O cluster healthy:        TRUE 
##     H2O Connection ip:          localhost 
##     H2O Connection port:        54321 
##     H2O Connection proxy:       NA 
##     H2O Internal Security:      FALSE 
##     R Version:                  R version 3.3.3 (2017-03-06)
## Warning in h2o.clusterInfo(): 
## Your H2O cluster version is too old (4 months and 4 days)!
## Please download and install the latest version from http://h2o.ai/download/
#install aws sdk if not present (pre-requisite for using Athena with an IAM role)
if (!aws_sdk_present()) {
  install_aws_sdk()
}

load_sdk()
## NULL

Connect to Athena

Next, establish a connection to Athena from RStudio, using an IAM role associated with your EC2 instance. Use ATHENABUCKET to specify the S3 staging directory.

URL <- 'https://s3.amazonaws.com/athena-downloads/drivers/AthenaJDBC41-1.0.1.jar'
fil <- basename(URL)
#download the file into current working directory
if (!file.exists(fil)) download.file(URL, fil)
#verify that the file has been downloaded successfully
list.files()
## [1] "AthenaJDBC41-1.0.1.jar"
drv <- JDBC(driverClass="com.amazonaws.athena.jdbc.AthenaDriver", fil, identifier.quote="'")

con <- jdbcConnection <- dbConnect(drv, 'jdbc:awsathena://athena.us-east-1.amazonaws.com:443/',
                                   s3_staging_dir=Sys.getenv("ATHENABUCKET"),
                                   aws_credentials_provider_class="com.amazonaws.auth.DefaultAWSCredentialsProviderChain")

Verify the connection. The results returned depend on your specific Athena setup.

con
## <JDBCConnection>
dbListTables(con)
##  [1] "gdelt"               "wikistats"           "elb_logs_raw_native"
##  [4] "twitter"             "twitter2"            "usermovieratings"   
##  [7] "eventcodes"          "events"              "billboard"          
## [10] "billboardtop10"      "elb_logs"            "gdelthist"          
## [13] "gdeltmaster"         "twitter"             "twitter3"

Create a dataset

For this analysis, you use a sample dataset combining information from Billboard and Wikipedia with Echo Nest data in the Million Songs Dataset. Upload this dataset into your own S3 bucket. The table below provides a description of the fields used in this dataset.

Field Description
year Year that song was released
songtitle Title of the song
artistname Name of the song artist
songid Unique identifier for the song
artistid Unique identifier for the song artist
timesignature Variable estimating the time signature of the song
timesignature_confidence Confidence in the estimate for the timesignature
loudness Continuous variable indicating the average amplitude of the audio in decibels
tempo Variable indicating the estimated beats per minute of the song
tempo_confidence Confidence in the estimate for tempo
key Variable with twelve levels indicating the estimated key of the song (C, C#, B)
key_confidence Confidence in the estimate for key
energy Variable that represents the overall acoustic energy of the song, using a mix of features such as loudness
pitch Continuous variable that indicates the pitch of the song
timbre_0_min thru timbre_11_min Variables that indicate the minimum values over all segments for each of the twelve values in the timbre vector
timbre_0_max thru timbre_11_max Variables that indicate the maximum values over all segments for each of the twelve values in the timbre vector
top10 Indicator for whether or not the song made it to the Top 10 of the Billboard charts (1 if it was in the top 10, and 0 if not)

Create an Athena table based on the dataset

In the Athena console, select the default database, sampled, or create a new database.

Run the following create table statement.

create external table if not exists billboard
(
year int,
songtitle string,
artistname string,
songID string,
artistID string,
timesignature int,
timesignature_confidence double,
loudness double,
tempo double,
tempo_confidence double,
key int,
key_confidence double,
energy double,
pitch double,
timbre_0_min double,
timbre_0_max double,
timbre_1_min double,
timbre_1_max double,
timbre_2_min double,
timbre_2_max double,
timbre_3_min double,
timbre_3_max double,
timbre_4_min double,
timbre_4_max double,
timbre_5_min double,
timbre_5_max double,
timbre_6_min double,
timbre_6_max double,
timbre_7_min double,
timbre_7_max double,
timbre_8_min double,
timbre_8_max double,
timbre_9_min double,
timbre_9_max double,
timbre_10_min double,
timbre_10_max double,
timbre_11_min double,
timbre_11_max double,
Top10 int
)
ROW FORMAT DELIMITED
FIELDS TERMINATED BY ','
STORED AS TEXTFILE
LOCATION 's3://aws-bigdata-blog/artifacts/predict-billboard/data'
;

Inspect the table definition for the ‘billboard’ table that you have created. If you chose a database other than sampledb, replace that value with your choice.

dbGetQuery(con, "show create table sampledb.billboard")
##                                      createtab_stmt
## 1       CREATE EXTERNAL TABLE `sampledb.billboard`(
## 2                                       `year` int,
## 3                               `songtitle` string,
## 4                              `artistname` string,
## 5                                  `songid` string,
## 6                                `artistid` string,
## 7                              `timesignature` int,
## 8                `timesignature_confidence` double,
## 9                                `loudness` double,
## 10                                  `tempo` double,
## 11                       `tempo_confidence` double,
## 12                                       `key` int,
## 13                         `key_confidence` double,
## 14                                 `energy` double,
## 15                                  `pitch` double,
## 16                           `timbre_0_min` double,
## 17                           `timbre_0_max` double,
## 18                           `timbre_1_min` double,
## 19                           `timbre_1_max` double,
## 20                           `timbre_2_min` double,
## 21                           `timbre_2_max` double,
## 22                           `timbre_3_min` double,
## 23                           `timbre_3_max` double,
## 24                           `timbre_4_min` double,
## 25                           `timbre_4_max` double,
## 26                           `timbre_5_min` double,
## 27                           `timbre_5_max` double,
## 28                           `timbre_6_min` double,
## 29                           `timbre_6_max` double,
## 30                           `timbre_7_min` double,
## 31                           `timbre_7_max` double,
## 32                           `timbre_8_min` double,
## 33                           `timbre_8_max` double,
## 34                           `timbre_9_min` double,
## 35                           `timbre_9_max` double,
## 36                          `timbre_10_min` double,
## 37                          `timbre_10_max` double,
## 38                          `timbre_11_min` double,
## 39                          `timbre_11_max` double,
## 40                                     `top10` int)
## 41                             ROW FORMAT DELIMITED 
## 42                         FIELDS TERMINATED BY ',' 
## 43                            STORED AS INPUTFORMAT 
## 44       'org.apache.hadoop.mapred.TextInputFormat' 
## 45                                     OUTPUTFORMAT 
## 46  'org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat'
## 47                                        LOCATION
## 48    's3://aws-bigdata-blog/artifacts/predict-billboard/data'
## 49                                  TBLPROPERTIES (
## 50            'transient_lastDdlTime'='1505484133')

Run a sample query

Next, run a sample query to obtain a list of all songs from Janet Jackson that made it to the Billboard Top 10 charts.

dbGetQuery(con, " SELECT songtitle,artistname,top10   FROM sampledb.billboard WHERE lower(artistname) =     'janet jackson' AND top10 = 1")
##                       songtitle    artistname top10
## 1                       Runaway Janet Jackson     1
## 2               Because Of Love Janet Jackson     1
## 3                         Again Janet Jackson     1
## 4                            If Janet Jackson     1
## 5  Love Will Never Do (Without You) Janet Jackson 1
## 6                     Black Cat Janet Jackson     1
## 7               Come Back To Me Janet Jackson     1
## 8                       Alright Janet Jackson     1
## 9                      Escapade Janet Jackson     1
## 10                Rhythm Nation Janet Jackson     1

Determine how many songs in this dataset are specifically from the year 2010.

dbGetQuery(con, " SELECT count(*)   FROM sampledb.billboard WHERE year = 2010")
##   _col0
## 1   373

The sample dataset provides certain song properties of interest that can be analyzed to gauge the impact to the song’s overall popularity. Look at one such property, timesignature, and determine the value that is the most frequent among songs in the database. Timesignature is a measure of the number of beats and the type of note involved.

Running the query directly may result in an error, as shown in the commented lines below. This error is a result of trying to retrieve a large result set over a JDBC connection, which can cause out-of-memory issues at the client level. To address this, reduce the fetch size and run again.

#t<-dbGetQuery(con, " SELECT timesignature FROM sampledb.billboard")
#Note:  Running the preceding query results in the following error: 
#Error in .jcall(rp, "I", "fetch", stride, block): java.sql.SQLException: The requested #fetchSize is more than the allowed value in Athena. Please reduce the fetchSize and try #again. Refer to the Athena documentation for valid fetchSize values.
# Use the dbSendQuery function, reduce the fetch size, and run again
r <- dbSendQuery(con, " SELECT timesignature     FROM sampledb.billboard")
dftimesignature<- fetch(r, n=-1, block=100)
dbClearResult(r)
## [1] TRUE
table(dftimesignature)
## dftimesignature
##    0    1    3    4    5    7 
##   10  143  503 6787  112   19
nrow(dftimesignature)
## [1] 7574

From the results, observe that 6787 songs have a timesignature of 4.

Next, determine the song with the highest tempo.

dbGetQuery(con, " SELECT songtitle,artistname,tempo   FROM sampledb.billboard WHERE tempo = (SELECT max(tempo) FROM sampledb.billboard) ")
##                   songtitle      artistname   tempo
## 1 Wanna Be Startin' Somethin' Michael Jackson 244.307

Create the training dataset

Your model needs to be trained such that it can learn and make accurate predictions. Split the data into training and test datasets, and create the training dataset first.  This dataset contains all observations from the year 2009 and earlier. You may face the same JDBC connection issue pointed out earlier, so this query uses a fetch size.

#BillboardTrain <- dbGetQuery(con, "SELECT * FROM sampledb.billboard WHERE year <= 2009")
#Running the preceding query results in the following error:-
#Error in .verify.JDBC.result(r, "Unable to retrieve JDBC result set for ", : Unable to retrieve #JDBC result set for SELECT * FROM sampledb.billboard WHERE year <= 2009 (Internal error)
#Follow the same approach as before to address this issue.

r <- dbSendQuery(con, "SELECT * FROM sampledb.billboard WHERE year <= 2009")
BillboardTrain <- fetch(r, n=-1, block=100)
dbClearResult(r)
## [1] TRUE
BillboardTrain[1:2,c(1:3,6:10)]
##   year           songtitle artistname timesignature
## 1 2009 The Awkward Goodbye    Athlete             3
## 2 2009        Rubik's Cube    Athlete             3
##   timesignature_confidence loudness   tempo tempo_confidence
## 1                    0.732   -6.320  89.614   0.652
## 2                    0.906   -9.541 117.742   0.542
nrow(BillboardTrain)
## [1] 7201

Create the test dataset

BillboardTest <- dbGetQuery(con, "SELECT * FROM sampledb.billboard where year = 2010")
BillboardTest[1:2,c(1:3,11:15)]
##   year              songtitle        artistname key
## 1 2010 This Is the House That Doubt Built A Day to Remember  11
## 2 2010        Sticks & Bricks A Day to Remember  10
##   key_confidence    energy pitch timbre_0_min
## 1          0.453 0.9666556 0.024        0.002
## 2          0.469 0.9847095 0.025        0.000
nrow(BillboardTest)
## [1] 373

Convert the training and test datasets into H2O dataframes

train.h2o <- as.h2o(BillboardTrain)
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |=================================================================| 100%
test.h2o <- as.h2o(BillboardTest)
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |=================================================================| 100%

Inspect the column names in your H2O dataframes.

colnames(train.h2o)
##  [1] "year"                     "songtitle"               
##  [3] "artistname"               "songid"                  
##  [5] "artistid"                 "timesignature"           
##  [7] "timesignature_confidence" "loudness"                
##  [9] "tempo"                    "tempo_confidence"        
## [11] "key"                      "key_confidence"          
## [13] "energy"                   "pitch"                   
## [15] "timbre_0_min"             "timbre_0_max"            
## [17] "timbre_1_min"             "timbre_1_max"            
## [19] "timbre_2_min"             "timbre_2_max"            
## [21] "timbre_3_min"             "timbre_3_max"            
## [23] "timbre_4_min"             "timbre_4_max"            
## [25] "timbre_5_min"             "timbre_5_max"            
## [27] "timbre_6_min"             "timbre_6_max"            
## [29] "timbre_7_min"             "timbre_7_max"            
## [31] "timbre_8_min"             "timbre_8_max"            
## [33] "timbre_9_min"             "timbre_9_max"            
## [35] "timbre_10_min"            "timbre_10_max"           
## [37] "timbre_11_min"            "timbre_11_max"           
## [39] "top10"

Create models

You need to designate the independent and dependent variables prior to applying your modeling algorithms. Because you’re trying to predict the ‘top10’ field, this would be your dependent variable and everything else would be independent.

Create your first model using GLM. Because GLM works best with numeric data, you create your model by dropping non-numeric variables. You only use the variables in the dataset that describe the numerical attributes of the song in the logistic regression model. You won’t use these variables:  “year”, “songtitle”, “artistname”, “songid”, or “artistid”.

y.dep <- 39
x.indep <- c(6:38)
x.indep
##  [1]  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
## [24] 29 30 31 32 33 34 35 36 37 38

Create Model 1: All numeric variables

Create Model 1 with the training dataset, using GLM as the modeling algorithm and H2O’s built-in h2o.glm function.

modelh1 <- h2o.glm( y = y.dep, x = x.indep, training_frame = train.h2o, family = "binomial")
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |=====                                                            |   8%
  |                                                                       
  |=================================================================| 100%

Measure the performance of Model 1, using H2O’s built-in performance function.

h2o.performance(model=modelh1,newdata=test.h2o)
## H2OBinomialMetrics: glm
## 
## MSE:  0.09924684
## RMSE:  0.3150347
## LogLoss:  0.3220267
## Mean Per-Class Error:  0.2380168
## AUC:  0.8431394
## Gini:  0.6862787
## R^2:  0.254663
## Null Deviance:  326.0801
## Residual Deviance:  240.2319
## AIC:  308.2319
## 
## Confusion Matrix (vertical: actual; across: predicted) for F1-optimal threshold:
##          0   1    Error     Rate
## 0      255  59 0.187898  =59/314
## 1       17  42 0.288136   =17/59
## Totals 272 101 0.203753  =76/373
## 
## Maximum Metrics: Maximum metrics at their respective thresholds
##                         metric threshold    value idx
## 1                       max f1  0.192772 0.525000 100
## 2                       max f2  0.124912 0.650510 155
## 3                 max f0point5  0.416258 0.612903  23
## 4                 max accuracy  0.416258 0.879357  23
## 5                max precision  0.813396 1.000000   0
## 6                   max recall  0.037579 1.000000 282
## 7              max specificity  0.813396 1.000000   0
## 8             max absolute_mcc  0.416258 0.455251  23
## 9   max min_per_class_accuracy  0.161402 0.738854 125
## 10 max mean_per_class_accuracy  0.124912 0.765006 155
## 
## Gains/Lift Table: Extract with `h2o.gainsLift(<model>, <data>)` or ` 
h2o.auc(h2o.performance(modelh1,test.h2o)) 
## [1] 0.8431394

The AUC metric provides insight into how well the classifier is able to separate the two classes. In this case, the value of 0.8431394 indicates that the classification is good. (A value of 0.5 indicates a worthless test, while a value of 1.0 indicates a perfect test.)

Next, inspect the coefficients of the variables in the dataset.

dfmodelh1 <- as.data.frame(h2o.varimp(modelh1))
dfmodelh1
##                       names coefficients sign
## 1              timbre_0_max  1.290938663  NEG
## 2                  loudness  1.262941934  POS
## 3                     pitch  0.616995941  NEG
## 4              timbre_1_min  0.422323735  POS
## 5              timbre_6_min  0.349016024  NEG
## 6                    energy  0.348092062  NEG
## 7             timbre_11_min  0.307331997  NEG
## 8              timbre_3_max  0.302225619  NEG
## 9             timbre_11_max  0.243632060  POS
## 10             timbre_4_min  0.224233951  POS
## 11             timbre_4_max  0.204134342  POS
## 12             timbre_5_min  0.199149324  NEG
## 13             timbre_0_min  0.195147119  POS
## 14 timesignature_confidence  0.179973904  POS
## 15         tempo_confidence  0.144242598  POS
## 16            timbre_10_max  0.137644568  POS
## 17             timbre_7_min  0.126995955  NEG
## 18            timbre_10_min  0.123851179  POS
## 19             timbre_7_max  0.100031481  NEG
## 20             timbre_2_min  0.096127636  NEG
## 21           key_confidence  0.083115820  POS
## 22             timbre_6_max  0.073712419  POS
## 23            timesignature  0.067241917  POS
## 24             timbre_8_min  0.061301881  POS
## 25             timbre_8_max  0.060041698  POS
## 26                      key  0.056158445  POS
## 27             timbre_3_min  0.050825116  POS
## 28             timbre_9_max  0.033733561  POS
## 29             timbre_2_max  0.030939072  POS
## 30             timbre_9_min  0.020708113  POS
## 31             timbre_1_max  0.014228818  NEG
## 32                    tempo  0.008199861  POS
## 33             timbre_5_max  0.004837870  POS
## 34                                    NA <NA>

Typically, songs with heavier instrumentation tend to be louder (have higher values in the variable “loudness”) and more energetic (have higher values in the variable “energy”). This knowledge is helpful for interpreting the modeling results.

You can make the following observations from the results:

  • The coefficient estimates for the confidence values associated with the time signature, key, and tempo variables are positive. This suggests that higher confidence leads to a higher predicted probability of a Top 10 hit.
  • The coefficient estimate for loudness is positive, meaning that mainstream listeners prefer louder songs with heavier instrumentation.
  • The coefficient estimate for energy is negative, meaning that mainstream listeners prefer songs that are less energetic, which are those songs with light instrumentation.

These coefficients lead to contradictory conclusions for Model 1. This could be due to multicollinearity issues. Inspect the correlation between the variables “loudness” and “energy” in the training set.

cor(train.h2o$loudness,train.h2o$energy)
## [1] 0.7399067

This number indicates that these two variables are highly correlated, and Model 1 does indeed suffer from multicollinearity. Typically, you associate a value of -1.0 to -0.5 or 1.0 to 0.5 to indicate strong correlation, and a value of 0.1 to 0.1 to indicate weak correlation. To avoid this correlation issue, omit one of these two variables and re-create the models.

You build two variations of the original model:

  • Model 2, in which you keep “energy” and omit “loudness”
  • Model 3, in which you keep “loudness” and omit “energy”

You compare these two models and choose the model with a better fit for this use case.

Create Model 2: Keep energy and omit loudness

colnames(train.h2o)
##  [1] "year"                     "songtitle"               
##  [3] "artistname"               "songid"                  
##  [5] "artistid"                 "timesignature"           
##  [7] "timesignature_confidence" "loudness"                
##  [9] "tempo"                    "tempo_confidence"        
## [11] "key"                      "key_confidence"          
## [13] "energy"                   "pitch"                   
## [15] "timbre_0_min"             "timbre_0_max"            
## [17] "timbre_1_min"             "timbre_1_max"            
## [19] "timbre_2_min"             "timbre_2_max"            
## [21] "timbre_3_min"             "timbre_3_max"            
## [23] "timbre_4_min"             "timbre_4_max"            
## [25] "timbre_5_min"             "timbre_5_max"            
## [27] "timbre_6_min"             "timbre_6_max"            
## [29] "timbre_7_min"             "timbre_7_max"            
## [31] "timbre_8_min"             "timbre_8_max"            
## [33] "timbre_9_min"             "timbre_9_max"            
## [35] "timbre_10_min"            "timbre_10_max"           
## [37] "timbre_11_min"            "timbre_11_max"           
## [39] "top10"
y.dep <- 39
x.indep <- c(6:7,9:38)
x.indep
##  [1]  6  7  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
## [24] 30 31 32 33 34 35 36 37 38
modelh2 <- h2o.glm( y = y.dep, x = x.indep, training_frame = train.h2o, family = "binomial")
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |=======                                                          |  10%
  |                                                                       
  |=================================================================| 100%

Measure the performance of Model 2.

h2o.performance(model=modelh2,newdata=test.h2o)
## H2OBinomialMetrics: glm
## 
## MSE:  0.09922606
## RMSE:  0.3150017
## LogLoss:  0.3228213
## Mean Per-Class Error:  0.2490554
## AUC:  0.8431933
## Gini:  0.6863867
## R^2:  0.2548191
## Null Deviance:  326.0801
## Residual Deviance:  240.8247
## AIC:  306.8247
## 
## Confusion Matrix (vertical: actual; across: predicted) for F1-optimal threshold:
##          0  1    Error     Rate
## 0      280 34 0.108280  =34/314
## 1       23 36 0.389831   =23/59
## Totals 303 70 0.152815  =57/373
## 
## Maximum Metrics: Maximum metrics at their respective thresholds
##                         metric threshold    value idx
## 1                       max f1  0.254391 0.558140  69
## 2                       max f2  0.113031 0.647208 157
## 3                 max f0point5  0.413999 0.596026  22
## 4                 max accuracy  0.446250 0.876676  18
## 5                max precision  0.811739 1.000000   0
## 6                   max recall  0.037682 1.000000 283
## 7              max specificity  0.811739 1.000000   0
## 8             max absolute_mcc  0.254391 0.469060  69
## 9   max min_per_class_accuracy  0.141051 0.716561 131
## 10 max mean_per_class_accuracy  0.113031 0.761821 157
## 
## Gains/Lift Table: Extract with `h2o.gainsLift(<model>, <data>)` or `h2o.gainsLift(<model>, valid=<T/F>, xval=<T/F>)`
dfmodelh2 <- as.data.frame(h2o.varimp(modelh2))
dfmodelh2
##                       names coefficients sign
## 1                     pitch  0.700331511  NEG
## 2              timbre_1_min  0.510270513  POS
## 3              timbre_0_max  0.402059546  NEG
## 4              timbre_6_min  0.333316236  NEG
## 5             timbre_11_min  0.331647383  NEG
## 6              timbre_3_max  0.252425901  NEG
## 7             timbre_11_max  0.227500308  POS
## 8              timbre_4_max  0.210663865  POS
## 9              timbre_0_min  0.208516163  POS
## 10             timbre_5_min  0.202748055  NEG
## 11             timbre_4_min  0.197246582  POS
## 12            timbre_10_max  0.172729619  POS
## 13         tempo_confidence  0.167523934  POS
## 14 timesignature_confidence  0.167398830  POS
## 15             timbre_7_min  0.142450727  NEG
## 16             timbre_8_max  0.093377516  POS
## 17            timbre_10_min  0.090333426  POS
## 18            timesignature  0.085851625  POS
## 19             timbre_7_max  0.083948442  NEG
## 20           key_confidence  0.079657073  POS
## 21             timbre_6_max  0.076426046  POS
## 22             timbre_2_min  0.071957831  NEG
## 23             timbre_9_max  0.071393189  POS
## 24             timbre_8_min  0.070225578  POS
## 25                      key  0.061394702  POS
## 26             timbre_3_min  0.048384697  POS
## 27             timbre_1_max  0.044721121  NEG
## 28                   energy  0.039698433  POS
## 29             timbre_5_max  0.039469064  POS
## 30             timbre_2_max  0.018461133  POS
## 31                    tempo  0.013279926  POS
## 32             timbre_9_min  0.005282143  NEG
## 33                                    NA <NA>

h2o.auc(h2o.performance(modelh2,test.h2o)) 
## [1] 0.8431933

You can make the following observations:

  • The AUC metric is 0.8431933.
  • Inspecting the coefficient of the variable energy, Model 2 suggests that songs with high energy levels tend to be more popular. This is as per expectation.
  • As H2O orders variables by significance, the variable energy is not significant in this model.

You can conclude that Model 2 is not ideal for this use , as energy is not significant.

CreateModel 3: Keep loudness but omit energy

colnames(train.h2o)
##  [1] "year"                     "songtitle"               
##  [3] "artistname"               "songid"                  
##  [5] "artistid"                 "timesignature"           
##  [7] "timesignature_confidence" "loudness"                
##  [9] "tempo"                    "tempo_confidence"        
## [11] "key"                      "key_confidence"          
## [13] "energy"                   "pitch"                   
## [15] "timbre_0_min"             "timbre_0_max"            
## [17] "timbre_1_min"             "timbre_1_max"            
## [19] "timbre_2_min"             "timbre_2_max"            
## [21] "timbre_3_min"             "timbre_3_max"            
## [23] "timbre_4_min"             "timbre_4_max"            
## [25] "timbre_5_min"             "timbre_5_max"            
## [27] "timbre_6_min"             "timbre_6_max"            
## [29] "timbre_7_min"             "timbre_7_max"            
## [31] "timbre_8_min"             "timbre_8_max"            
## [33] "timbre_9_min"             "timbre_9_max"            
## [35] "timbre_10_min"            "timbre_10_max"           
## [37] "timbre_11_min"            "timbre_11_max"           
## [39] "top10"
y.dep <- 39
x.indep <- c(6:12,14:38)
x.indep
##  [1]  6  7  8  9 10 11 12 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
## [24] 30 31 32 33 34 35 36 37 38
modelh3 <- h2o.glm( y = y.dep, x = x.indep, training_frame = train.h2o, family = "binomial")
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |========                                                         |  12%
  |                                                                       
  |=================================================================| 100%
perfh3<-h2o.performance(model=modelh3,newdata=test.h2o)
perfh3
## H2OBinomialMetrics: glm
## 
## MSE:  0.0978859
## RMSE:  0.3128672
## LogLoss:  0.3178367
## Mean Per-Class Error:  0.264925
## AUC:  0.8492389
## Gini:  0.6984778
## R^2:  0.2648836
## Null Deviance:  326.0801
## Residual Deviance:  237.1062
## AIC:  303.1062
## 
## Confusion Matrix (vertical: actual; across: predicted) for F1-optimal threshold:
##          0  1    Error     Rate
## 0      286 28 0.089172  =28/314
## 1       26 33 0.440678   =26/59
## Totals 312 61 0.144772  =54/373
## 
## Maximum Metrics: Maximum metrics at their respective thresholds
##                         metric threshold    value idx
## 1                       max f1  0.273799 0.550000  60
## 2                       max f2  0.125503 0.663265 155
## 3                 max f0point5  0.435479 0.628931  24
## 4                 max accuracy  0.435479 0.882038  24
## 5                max precision  0.821606 1.000000   0
## 6                   max recall  0.038328 1.000000 280
## 7              max specificity  0.821606 1.000000   0
## 8             max absolute_mcc  0.435479 0.471426  24
## 9   max min_per_class_accuracy  0.173693 0.745763 120
## 10 max mean_per_class_accuracy  0.125503 0.775073 155
## 
## Gains/Lift Table: Extract with `h2o.gainsLift(<model>, <data>)` or `h2o.gainsLift(<model>, valid=<T/F>, xval=<T/F>)`
dfmodelh3 <- as.data.frame(h2o.varimp(modelh3))
dfmodelh3
##                       names coefficients sign
## 1              timbre_0_max 1.216621e+00  NEG
## 2                  loudness 9.780973e-01  POS
## 3                     pitch 7.249788e-01  NEG
## 4              timbre_1_min 3.891197e-01  POS
## 5              timbre_6_min 3.689193e-01  NEG
## 6             timbre_11_min 3.086673e-01  NEG
## 7              timbre_3_max 3.025593e-01  NEG
## 8             timbre_11_max 2.459081e-01  POS
## 9              timbre_4_min 2.379749e-01  POS
## 10             timbre_4_max 2.157627e-01  POS
## 11             timbre_0_min 1.859531e-01  POS
## 12             timbre_5_min 1.846128e-01  NEG
## 13 timesignature_confidence 1.729658e-01  POS
## 14             timbre_7_min 1.431871e-01  NEG
## 15            timbre_10_max 1.366703e-01  POS
## 16            timbre_10_min 1.215954e-01  POS
## 17         tempo_confidence 1.183698e-01  POS
## 18             timbre_2_min 1.019149e-01  NEG
## 19           key_confidence 9.109701e-02  POS
## 20             timbre_7_max 8.987908e-02  NEG
## 21             timbre_6_max 6.935132e-02  POS
## 22             timbre_8_max 6.878241e-02  POS
## 23            timesignature 6.120105e-02  POS
## 24                      key 5.814805e-02  POS
## 25             timbre_8_min 5.759228e-02  POS
## 26             timbre_1_max 2.930285e-02  NEG
## 27             timbre_9_max 2.843755e-02  POS
## 28             timbre_3_min 2.380245e-02  POS
## 29             timbre_2_max 1.917035e-02  POS
## 30             timbre_5_max 1.715813e-02  POS
## 31                    tempo 1.364418e-02  NEG
## 32             timbre_9_min 8.463143e-05  NEG
## 33                                    NA <NA>
h2o.sensitivity(perfh3,0.5)
## Warning in h2o.find_row_by_threshold(object, t): Could not find exact
## threshold: 0.5 for this set of metrics; using closest threshold found:
## 0.501855569251422. Run `h2o.predict` and apply your desired threshold on a
## probability column.
## [[1]]
## [1] 0.2033898
h2o.auc(perfh3)
## [1] 0.8492389

You can make the following observations:

  • The AUC metric is 0.8492389.
  • From the confusion matrix, the model correctly predicts that 33 songs will be top 10 hits (true positives). However, it has 26 false positives (songs that the model predicted would be Top 10 hits, but ended up not being Top 10 hits).
  • Loudness has a positive coefficient estimate, meaning that this model predicts that songs with heavier instrumentation tend to be more popular. This is the same conclusion from Model 2.
  • Loudness is significant in this model.

Overall, Model 3 predicts a higher number of top 10 hits with an accuracy rate that is acceptable. To choose the best fit for production runs, record labels should consider the following factors:

  • Desired model accuracy at a given threshold
  • Number of correct predictions for top10 hits
  • Tolerable number of false positives or false negatives

Next, make predictions using Model 3 on the test dataset.

predict.regh <- h2o.predict(modelh3, test.h2o)
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |=================================================================| 100%
print(predict.regh)
##   predict        p0          p1
## 1       0 0.9654739 0.034526052
## 2       0 0.9654748 0.034525236
## 3       0 0.9635547 0.036445318
## 4       0 0.9343579 0.065642149
## 5       0 0.9978334 0.002166601
## 6       0 0.9779949 0.022005078
## 
## [373 rows x 3 columns]
predict.regh$predict
##   predict
## 1       0
## 2       0
## 3       0
## 4       0
## 5       0
## 6       0
## 
## [373 rows x 1 column]
dpr<-as.data.frame(predict.regh)
#Rename the predicted column 
colnames(dpr)[colnames(dpr) == 'predict'] <- 'predict_top10'
table(dpr$predict_top10)
## 
##   0   1 
## 312  61

The first set of output results specifies the probabilities associated with each predicted observation.  For example, observation 1 is 96.54739% likely to not be a Top 10 hit, and 3.4526052% likely to be a Top 10 hit (predict=1 indicates Top 10 hit and predict=0 indicates not a Top 10 hit).  The second set of results list the actual predictions made.  From the third set of results, this model predicts that 61 songs will be top 10 hits.

Compute the baseline accuracy, by assuming that the baseline predicts the most frequent outcome, which is that most songs are not Top 10 hits.

table(BillboardTest$top10)
## 
##   0   1 
## 314  59

Now observe that the baseline model would get 314 observations correct, and 59 wrong, for an accuracy of 314/(314+59) = 0.8418231.

It seems that Model 3, with an accuracy of 0.8552, provides you with a small improvement over the baseline model. But is this model useful for record labels?

View the two models from an investment perspective:

  • A production company is interested in investing in songs that are more likely to make it to the Top 10. The company’s objective is to minimize the risk of financial losses attributed to investing in songs that end up unpopular.
  • How many songs does Model 3 correctly predict as a Top 10 hit in 2010? Looking at the confusion matrix, you see that it predicts 33 top 10 hits correctly at an optimal threshold, which is more than half the number
  • It will be more useful to the record label if you can provide the production company with a list of songs that are highly likely to end up in the Top 10.
  • The baseline model is not useful, as it simply does not label any song as a hit.

Considering the three models built so far, you can conclude that Model 3 proves to be the best investment choice for the record label.

GBM model

H2O provides you with the ability to explore other learning models, such as GBM and deep learning. Explore building a model using the GBM technique, using the built-in h2o.gbm function.

Before you do this, you need to convert the target variable to a factor for multinomial classification techniques.

train.h2o$top10=as.factor(train.h2o$top10)
gbm.modelh <- h2o.gbm(y=y.dep, x=x.indep, training_frame = train.h2o, ntrees = 500, max_depth = 4, learn_rate = 0.01, seed = 1122,distribution="multinomial")
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |===                                                              |   5%
  |                                                                       
  |=====                                                            |   7%
  |                                                                       
  |======                                                           |   9%
  |                                                                       
  |=======                                                          |  10%
  |                                                                       
  |======================                                           |  33%
  |                                                                       
  |=====================================                            |  56%
  |                                                                       
  |====================================================             |  79%
  |                                                                       
  |================================================================ |  98%
  |                                                                       
  |=================================================================| 100%
perf.gbmh<-h2o.performance(gbm.modelh,test.h2o)
perf.gbmh
## H2OBinomialMetrics: gbm
## 
## MSE:  0.09860778
## RMSE:  0.3140188
## LogLoss:  0.3206876
## Mean Per-Class Error:  0.2120263
## AUC:  0.8630573
## Gini:  0.7261146
## 
## Confusion Matrix (vertical: actual; across: predicted) for F1-optimal threshold:
##          0  1    Error     Rate
## 0      266 48 0.152866  =48/314
## 1       16 43 0.271186   =16/59
## Totals 282 91 0.171582  =64/373
## 
## Maximum Metrics: Maximum metrics at their respective thresholds
##                       metric threshold    value idx
## 1                     max f1  0.189757 0.573333  90
## 2                     max f2  0.130895 0.693717 145
## 3               max f0point5  0.327346 0.598802  26
## 4               max accuracy  0.442757 0.876676  14
## 5              max precision  0.802184 1.000000   0
## 6                 max recall  0.049990 1.000000 284
## 7            max specificity  0.802184 1.000000   0
## 8           max absolute_mcc  0.169135 0.496486 104
## 9 max min_per_class_accuracy  0.169135 0.796610 104
## 10 max mean_per_class_accuracy  0.169135 0.805948 104
## 
## Gains/Lift Table: Extract with `h2o.gainsLift(<model>, <data>)` or `
h2o.sensitivity(perf.gbmh,0.5)
## Warning in h2o.find_row_by_threshold(object, t): Could not find exact
## threshold: 0.5 for this set of metrics; using closest threshold found:
## 0.501205344484314. Run `h2o.predict` and apply your desired threshold on a
## probability column.
## [[1]]
## [1] 0.1355932
h2o.auc(perf.gbmh)
## [1] 0.8630573

This model correctly predicts 43 top 10 hits, which is 10 more than the number predicted by Model 3. Moreover, the AUC metric is higher than the one obtained from Model 3.

As seen above, H2O’s API provides the ability to obtain key statistical measures required to analyze the models easily, using several built-in functions. The record label can experiment with different parameters to arrive at the model that predicts the maximum number of Top 10 hits at the desired level of accuracy and threshold.

H2O also allows you to experiment with deep learning models. Deep learning models have the ability to learn features implicitly, but can be more expensive computationally.

Now, create a deep learning model with the h2o.deeplearning function, using the same training and test datasets created before. The time taken to run this model depends on the type of EC2 instance chosen for this purpose.  For models that require more computation, consider using accelerated computing instances such as the P2 instance type.

system.time(
  dlearning.modelh <- h2o.deeplearning(y = y.dep,
                                      x = x.indep,
                                      training_frame = train.h2o,
                                      epoch = 250,
                                      hidden = c(250,250),
                                      activation = "Rectifier",
                                      seed = 1122,
                                      distribution="multinomial"
  )
)
## 
  |                                                                       
  |                                                                 |   0%
  |                                                                       
  |===                                                              |   4%
  |                                                                       
  |=====                                                            |   8%
  |                                                                       
  |========                                                         |  12%
  |                                                                       
  |==========                                                       |  16%
  |                                                                       
  |=============                                                    |  20%
  |                                                                       
  |================                                                 |  24%
  |                                                                       
  |==================                                               |  28%
  |                                                                       
  |=====================                                            |  32%
  |                                                                       
  |=======================                                          |  36%
  |                                                                       
  |==========================                                       |  40%
  |                                                                       
  |=============================                                    |  44%
  |                                                                       
  |===============================                                  |  48%
  |                                                                       
  |==================================                               |  52%
  |                                                                       
  |====================================                             |  56%
  |                                                                       
  |=======================================                          |  60%
  |                                                                       
  |==========================================                       |  64%
  |                                                                       
  |============================================                     |  68%
  |                                                                       
  |===============================================                  |  72%
  |                                                                       
  |=================================================                |  76%
  |                                                                       
  |====================================================             |  80%
  |                                                                       
  |=======================================================          |  84%
  |                                                                       
  |=========================================================        |  88%
  |                                                                       
  |============================================================     |  92%
  |                                                                       
  |==============================================================   |  96%
  |                                                                       
  |=================================================================| 100%
##    user  system elapsed 
##   1.216   0.020 166.508
perf.dl<-h2o.performance(model=dlearning.modelh,newdata=test.h2o)
perf.dl
## H2OBinomialMetrics: deeplearning
## 
## MSE:  0.1678359
## RMSE:  0.4096778
## LogLoss:  1.86509
## Mean Per-Class Error:  0.3433013
## AUC:  0.7568822
## Gini:  0.5137644
## 
## Confusion Matrix (vertical: actual; across: predicted) for F1-optimal threshold:
##          0  1    Error     Rate
## 0      290 24 0.076433  =24/314
## 1       36 23 0.610169   =36/59
## Totals 326 47 0.160858  =60/373
## 
## Maximum Metrics: Maximum metrics at their respective thresholds
##                       metric threshold    value idx
## 1                     max f1  0.826267 0.433962  46
## 2                     max f2  0.000000 0.588235 239
## 3               max f0point5  0.999929 0.511811  16
## 4               max accuracy  0.999999 0.865952  10
## 5              max precision  1.000000 1.000000   0
## 6                 max recall  0.000000 1.000000 326
## 7            max specificity  1.000000 1.000000   0
## 8           max absolute_mcc  0.999929 0.363219  16
## 9 max min_per_class_accuracy  0.000004 0.662420 145
## 10 max mean_per_class_accuracy  0.000000 0.685334 224
## 
## Gains/Lift Table: Extract with `h2o.gainsLift(<model>, <data>)` or `h2o.gainsLift(<model>, valid=<T/F>, xval=<T/F>)`
h2o.sensitivity(perf.dl,0.5)
## Warning in h2o.find_row_by_threshold(object, t): Could not find exact
## threshold: 0.5 for this set of metrics; using closest threshold found:
## 0.496293348880151. Run `h2o.predict` and apply your desired threshold on a
## probability column.
## [[1]]
## [1] 0.3898305
h2o.auc(perf.dl)
## [1] 0.7568822

The AUC metric for this model is 0.7568822, which is less than what you got from the earlier models. I recommend further experimentation using different hyper parameters, such as the learning rate, epoch or the number of hidden layers.

H2O’s built-in functions provide many key statistical measures that can help measure model performance. Here are some of these key terms.

Metric Description
Sensitivity Measures the proportion of positives that have been correctly identified. It is also called the true positive rate, or recall.
Specificity Measures the proportion of negatives that have been correctly identified. It is also called the true negative rate.
Threshold Cutoff point that maximizes specificity and sensitivity. While the model may not provide the highest prediction at this point, it would not be biased towards positives or negatives.
Precision The fraction of the documents retrieved that are relevant to the information needed, for example, how many of the positively classified are relevant
AUC

Provides insight into how well the classifier is able to separate the two classes. The implicit goal is to deal with situations where the sample distribution is highly skewed, with a tendency to overfit to a single class.

0.90 – 1 = excellent (A)

0.8 – 0.9 = good (B)

0.7 – 0.8 = fair (C)

.6 – 0.7 = poor (D)

0.5 – 0.5 = fail (F)

Here’s a summary of the metrics generated from H2O’s built-in functions for the three models that produced useful results.

Metric Model 3 GBM Model Deep Learning Model

Accuracy

(max)

0.882038

(t=0.435479)

0.876676

(t=0.442757)

0.865952

(t=0.999999)

Precision

(max)

1.0

(t=0.821606)

1.0

(t=0802184)

1.0

(t=1.0)

Recall

(max)

1.0 1.0

1.0

(t=0)

Specificity

(max)

1.0 1.0

1.0

(t=1)

Sensitivity

 

0.2033898 0.1355932

0.3898305

(t=0.5)

AUC 0.8492389 0.8630573 0.756882

Note: ‘t’ denotes threshold.

Your options at this point could be narrowed down to Model 3 and the GBM model, based on the AUC and accuracy metrics observed earlier.  If the slightly lower accuracy of the GBM model is deemed acceptable, the record label can choose to go to production with the GBM model, as it can predict a higher number of Top 10 hits.  The AUC metric for the GBM model is also higher than that of Model 3.

Record labels can experiment with different learning techniques and parameters before arriving at a model that proves to be the best fit for their business. Because deep learning models can be computationally expensive, record labels can choose more powerful EC2 instances on AWS to run their experiments faster.

Conclusion

In this post, I showed how the popular music industry can use analytics to predict the type of songs that make the Top 10 Billboard charts. By running H2O’s scalable machine learning platform on AWS, data scientists can easily experiment with multiple modeling techniques and interactively query the data using Amazon Athena, without having to manage the underlying infrastructure. This helps record labels make critical decisions on the type of artists and songs to promote in a timely fashion, thereby increasing sales and revenue.

If you have questions or suggestions, please comment below.


Additional Reading

Learn how to build and explore a simple geospita simple GEOINT application using SparkR.


About the Authors

gopalGopal Wunnava is a Partner Solution Architect with the AWS GSI Team. He works with partners and customers on big data engagements, and is passionate about building analytical solutions that drive business capabilities and decision making. In his spare time, he loves all things sports and movies related and is fond of old classics like Asterix, Obelix comics and Hitchcock movies.

 

 

Bob Strahan, a Senior Consultant with AWS Professional Services, contributed to this post.

 

 

Introducing Gluon: a new library for machine learning from AWS and Microsoft

Post Syndicated from Ana Visneski original https://aws.amazon.com/blogs/aws/introducing-gluon-a-new-library-for-machine-learning-from-aws-and-microsoft/

Post by Dr. Matt Wood

Today, AWS and Microsoft announced Gluon, a new open source deep learning interface which allows developers to more easily and quickly build machine learning models, without compromising performance.

Gluon Logo

Gluon provides a clear, concise API for defining machine learning models using a collection of pre-built, optimized neural network components. Developers who are new to machine learning will find this interface more familiar to traditional code, since machine learning models can be defined and manipulated just like any other data structure. More seasoned data scientists and researchers will value the ability to build prototypes quickly and utilize dynamic neural network graphs for entirely new model architectures, all without sacrificing training speed.

Gluon is available in Apache MXNet today, a forthcoming Microsoft Cognitive Toolkit release, and in more frameworks over time.

Neural Networks vs Developers
Machine learning with neural networks (including ‘deep learning’) has three main components: data for training; a neural network model, and an algorithm which trains the neural network. You can think of the neural network in a similar way to a directed graph; it has a series of inputs (which represent the data), which connect to a series of outputs (the prediction), through a series of connected layers and weights. During training, the algorithm adjusts the weights in the network based on the error in the network output. This is the process by which the network learns; it is a memory and compute intensive process which can take days.

Deep learning frameworks such as Caffe2, Cognitive Toolkit, TensorFlow, and Apache MXNet are, in part, an answer to the question ‘how can we speed this process up? Just like query optimizers in databases, the more a training engine knows about the network and the algorithm, the more optimizations it can make to the training process (for example, it can infer what needs to be re-computed on the graph based on what else has changed, and skip the unaffected weights to speed things up). These frameworks also provide parallelization to distribute the computation process, and reduce the overall training time.

However, in order to achieve these optimizations, most frameworks require the developer to do some extra work: specifically, by providing a formal definition of the network graph, up-front, and then ‘freezing’ the graph, and just adjusting the weights.

The network definition, which can be large and complex with millions of connections, usually has to be constructed by hand. Not only are deep learning networks unwieldy, but they can be difficult to debug and it’s hard to re-use the code between projects.

The result of this complexity can be difficult for beginners and is a time-consuming task for more experienced researchers. At AWS, we’ve been experimenting with some ideas in MXNet around new, flexible, more approachable ways to define and train neural networks. Microsoft is also a contributor to the open source MXNet project, and were interested in some of these same ideas. Based on this, we got talking, and found we had a similar vision: to use these techniques to reduce the complexity of machine learning, making it accessible to more developers.

Enter Gluon: dynamic graphs, rapid iteration, scalable training
Gluon introduces four key innovations.

  1. Friendly API: Gluon networks can be defined using a simple, clear, concise code – this is easier for developers to learn, and much easier to understand than some of the more arcane and formal ways of defining networks and their associated weighted scoring functions.
  2. Dynamic networks: the network definition in Gluon is dynamic: it can bend and flex just like any other data structure. This is in contrast to the more common, formal, symbolic definition of a network which the deep learning framework has to effectively carve into stone in order to be able to effectively optimizing computation during training. Dynamic networks are easier to manage, and with Gluon, developers can easily ‘hybridize’ between these fast symbolic representations and the more friendly, dynamic ‘imperative’ definitions of the network and algorithms.
  3. The algorithm can define the network: the model and the training algorithm are brought much closer together. Instead of separate definitions, the algorithm can adjust the network dynamically during definition and training. Not only does this mean that developers can use standard programming loops, and conditionals to create these networks, but researchers can now define even more sophisticated algorithms and models which were not possible before. They are all easier to create, change, and debug.
  4. High performance operators for training: which makes it possible to have a friendly, concise API and dynamic graphs, without sacrificing training speed. This is a huge step forward in machine learning. Some frameworks bring a friendly API or dynamic graphs to deep learning, but these previous methods all incur a cost in terms of training speed. As with other areas of software, abstraction can slow down computation since it needs to be negotiated and interpreted at run time. Gluon can efficiently blend together a concise API with the formal definition under the hood, without the developer having to know about the specific details or to accommodate the compiler optimizations manually.

The team here at AWS, and our collaborators at Microsoft, couldn’t be more excited to bring these improvements to developers through Gluon. We’re already seeing quite a bit of excitement from developers and researchers alike.

Getting started with Gluon
Gluon is available today in Apache MXNet, with support coming for the Microsoft Cognitive Toolkit in a future release. We’re also publishing the front-end interface and the low-level API specifications so it can be included in other frameworks in the fullness of time.

You can get started with Gluon today. Fire up the AWS Deep Learning AMI with a single click and jump into one of 50 fully worked, notebook examples. If you’re a contributor to a machine learning framework, check out the interface specs on GitHub.

-Dr. Matt Wood