Tag Archives: metric

Monitoring your Amazon SNS message filtering activity with Amazon CloudWatch

Post Syndicated from Rachel Richardson original https://aws.amazon.com/blogs/compute/monitoring-your-amazon-sns-message-filtering-activity-with-amazon-cloudwatch/

This post is courtesy of Otavio Ferreira, Manager, Amazon SNS, AWS Messaging.

Amazon SNS message filtering provides a set of string and numeric matching operators that allow each subscription to receive only the messages of interest. Hence, SNS message filtering can simplify your pub/sub messaging architecture by offloading the message filtering logic from your subscriber systems, as well as the message routing logic from your publisher systems.

After you set the subscription attribute that defines a filter policy, the subscribing endpoint receives only the messages that carry attributes matching this filter policy. Other messages published to the topic are filtered out for this subscription. In this way, the native integration between SNS and Amazon CloudWatch provides visibility into the number of messages delivered, as well as the number of messages filtered out.

CloudWatch metrics are captured automatically for you. To get started with SNS message filtering, see Filtering Messages with Amazon SNS.

Message Filtering Metrics

The following six CloudWatch metrics are relevant to understanding your SNS message filtering activity:

  • NumberOfMessagesPublished – Inbound traffic to SNS. This metric tracks all the messages that have been published to the topic.
  • NumberOfNotificationsDelivered – Outbound traffic from SNS. This metric tracks all the messages that have been successfully delivered to endpoints subscribed to the topic. A delivery takes place either when the incoming message attributes match a subscription filter policy, or when the subscription has no filter policy at all, which results in a catch-all behavior.
  • NumberOfNotificationsFilteredOut – This metric tracks all the messages that were filtered out because they carried attributes that didn’t match the subscription filter policy.
  • NumberOfNotificationsFilteredOut-NoMessageAttributes – This metric tracks all the messages that were filtered out because they didn’t carry any attributes at all and, consequently, didn’t match the subscription filter policy.
  • NumberOfNotificationsFilteredOut-InvalidAttributes – This metric keeps track of messages that were filtered out because they carried invalid or malformed attributes and, thus, didn’t match the subscription filter policy.
  • NumberOfNotificationsFailed – This last metric tracks all the messages that failed to be delivered to subscribing endpoints, regardless of whether a filter policy had been set for the endpoint. This metric is emitted after the message delivery retry policy is exhausted, and SNS stops attempting to deliver the message. At that moment, the subscribing endpoint is likely no longer reachable. For example, the subscribing SQS queue or Lambda function has been deleted by its owner. You may want to closely monitor this metric to address message delivery issues quickly.

Message filtering graphs

Through the AWS Management Console, you can compose graphs to display your SNS message filtering activity. The graph shows the number of messages published, delivered, and filtered out within the timeframe you specify (1h, 3h, 12h, 1d, 3d, 1w, or custom).

SNS message filtering for CloudWatch Metrics

To compose an SNS message filtering graph with CloudWatch:

  1. Open the CloudWatch console.
  2. Choose Metrics, SNS, All Metrics, and Topic Metrics.
  3. Select all metrics to add to the graph, such as:
    • NumberOfMessagesPublished
    • NumberOfNotificationsDelivered
    • NumberOfNotificationsFilteredOut
  4. Choose Graphed metrics.
  5. In the Statistic column, switch from Average to Sum.
  6. Title your graph with a descriptive name, such as “SNS Message Filtering”

After you have your graph set up, you may want to copy the graph link for bookmarking, emailing, or sharing with co-workers. You may also want to add your graph to a CloudWatch dashboard for easy access in the future. Both actions are available to you on the Actions menu, which is found above the graph.

Summary

SNS message filtering defines how SNS topics behave in terms of message delivery. By using CloudWatch metrics, you gain visibility into the number of messages published, delivered, and filtered out. This enables you to validate the operation of filter policies and more easily troubleshoot during development phases.

SNS message filtering can be implemented easily with existing AWS SDKs by applying message and subscription attributes across all SNS supported protocols (Amazon SQS, AWS Lambda, HTTP, SMS, email, and mobile push). CloudWatch metrics for SNS message filtering is available now, in all AWS Regions.

For information about pricing, see the CloudWatch pricing page.

For more information, see:

Monitoring for Everyone

Post Syndicated from Blogs on Grafana Labs Blog original https://grafana.com/blog/2018/05/23/monitoring-for-everyone/

Øredev – Carl Bergquist – Monitoring for Everyone What is monitoring?
What do the terms log, metric, and distributed tracing actually mean?
What makes a good alert?
Why should I care?
At a recent developer conference in Malmö, Sweden, I gave a presentation on monitoring and observability to discuss the high level concepts and common tools that are out there.
Monitoring and observability can easily become quite complex, but at the heart of it, we simply want to know how our systems are performing, and when performance drops – be able to find out why.

LC4: Another Pen-and-Paper Cipher

Post Syndicated from Bruce Schneier original https://www.schneier.com/blog/archives/2018/05/lc4_another_pen.html

Interesting symmetric cipher: LC4:

Abstract: ElsieFour (LC4) is a low-tech cipher that can be computed by hand; but unlike many historical ciphers, LC4 is designed to be hard to break. LC4 is intended for encrypted communication between humans only, and therefore it encrypts and decrypts plaintexts and ciphertexts consisting only of the English letters A through Z plus a few other characters. LC4 uses a nonce in addition to the secret key, and requires that different messages use unique nonces. LC4 performs authenticated encryption, and optional header data can be included in the authentication. This paper defines the LC4 encryption and decryption algorithms, analyzes LC4’s security, and describes a simple appliance for computing LC4 by hand.

Almost two decades ago I designed Solitaire, a pen-and-paper cipher that uses a deck of playing cards to store the cipher’s state. This algorithm uses specialized tiles. This gives the cipher designer more options, but it can be incriminating in a way that regular playing cards are not.

Still, I like seeing more designs like this.

Hacker News thread.

Grafana v5.1 Released

Post Syndicated from Blogs on Grafana Labs Blog original https://grafana.com/blog/2018/04/26/grafana-v5.1-released/

v5.1 Stable Release

The recent 5.0 major release contained a lot of new features so the Grafana 5.1 release is focused on smoothing out the rough edges and iterating over some of the new features.

Download Grafana 5.1 Now

Release Highlights

There are two new features included, Heatmap Support for Prometheus and a new core data source for Microsoft SQL Server.

Another highlight is the revamp of the Grafana docker container that makes it easier to run and control but be aware there is a breaking change to file permissions that will affect existing containers with data volumes.

We got tons of useful improvement suggestions, bug reports and Pull Requests from our amazing community. Thank you all! See the full changelog for more details.

Improved Scrolling Experience

In Grafana v5.0 we introduced a new scrollbar component. Unfortunately this introduced a lot of issues and in some scenarios removed
the native scrolling functionality. Grafana v5.1 ships with a native scrollbar for all pages together with a scrollbar component for
the dashboard grid and panels that does not override the native scrolling functionality. We hope that these changes and improvements should
make the Grafana user experience much better!

Improved Docker Image

Grafana v5.1 brings an improved official docker image which should make it easier to run and use the Grafana docker image and at the same time give more control to the user how to use/run it.

We have switched the id of the grafana user running Grafana inside a docker container. Unfortunately this means that files created prior to 5.1 will not have the correct permissions for later versions and thereby introduces a breaking change. We made this change so that it would be easier for you to control what user Grafana is executed as.

Please read the updated documentation which includes migration instructions and more information.

Heatmap Support for Prometheus

The Prometheus datasource now supports transforming Prometheus histograms to the heatmap panel. The Prometheus histogram is a powerful feature, and we’re
really happy to finally allow our users to render those as heatmaps. The Heatmap panel documentation
contains more information on how to use it.

Another improvement is that the Prometheus query editor now supports autocomplete for template variables. More information in the Prometheus data source documentation.

Microsoft SQL Server

Grafana v5.1 now ships with a built-in Microsoft SQL Server (MSSQL) data source plugin that allows you to query and visualize data from any
Microsoft SQL Server 2005 or newer, including Microsoft Azure SQL Database. Do you have metric or log data in MSSQL? You can now visualize
that data and define alert rules on it as with any of Grafana’s other core datasources.

The using Microsoft SQL Server in Grafana documentation has more detailed information on how to get started.

Adding New Panels to Dashboards

The control for adding new panels to dashboards now includes panel search and it is also now possible to copy and paste panels between dashboards.

By copying a panel in a dashboard it will be displayed in the Paste tab. When you switch to a new dashboard you can paste the
copied panel.

Align Zero-Line for Right and Left Y-axes

The feature request to align the zero-line for right and left Y-axes on the Graph panel is more than 3 years old. It has finally been implemented – more information in the Graph panel documentation.

Other Highlights

  • Table Panel: New enhancements includes support for mapping a numeric value/range to text and additional units. More information in the Table panel documentation.
  • New variable interpolation syntax: We now support a new option for rendering variables that gives the user full control of how the value(s) should be rendered. More details in the in the Variables documentation.
  • Improved workflow for provisioned dashboards. More details here.

Changelog

Checkout the CHANGELOG.md file for a complete list
of new features, changes, and bug fixes.

Two NSA Algorithms Rejected by the ISO

Post Syndicated from Bruce Schneier original https://www.schneier.com/blog/archives/2018/04/two_nsa_algorit.html

The ISO has rejected two symmetric encryption algorithms: SIMON and SPECK. These algorithms were both designed by the NSA and made public in 2013. They are optimized for small and low-cost processors like IoT devices.

The risk of using NSA-designed ciphers, of course, is that they include NSA-designed backdoors. Personally, I doubt that they’re backdoored. And I always like seeing NSA-designed cryptography (particularly its key schedules). It’s like examining alien technology.

The End of Google Cloud Messaging, and What it Means for Your Apps

Post Syndicated from Zach Barbitta original https://aws.amazon.com/blogs/messaging-and-targeting/the-end-of-google-cloud-messaging-and-what-it-means-for-your-apps/

On April 10, 2018, Google announced the deprecation of its Google Cloud Messaging (GCM) platform. Specifically, the GCM server and client APIs are deprecated and will be removed as soon as April 11, 2019.  What does this mean for you and your applications that use Amazon Simple Notification Service (Amazon SNS) or Amazon Pinpoint?

First, nothing will break now or after April 11, 2019. GCM device tokens are completely interchangeable with the newer Firebase Cloud Messaging (FCM) device tokens. If you have existing GCM tokens, you’ll still be able to use them to send notifications. This statement is also true for GCM tokens that you generate in the future.

On the back end, we’ve already migrated Amazon SNS and Amazon Pinpoint to the server endpoint for FCM (https://fcm.googleapis.com/fcm/send). As a developer, you don’t need to make any changes as a result of this deprecation.

We created the following mini-FAQ to address some of the questions you may have as a developer who uses Amazon SNS or Amazon Pinpoint.

If I migrate to FCM from GCM, can I still use Amazon Pinpoint and Amazon SNS?

Yes. Your ability to connect to your applications and send messages through both Amazon SNS and Amazon Pinpoint doesn’t change. We’ll update the documentation for Amazon SNS and Amazon Pinpoint soon to reflect these changes.

If I don’t migrate to FCM from GCM, can I still use Amazon Pinpoint and Amazon SNS?

Yes. If you do nothing, your existing credentials and GCM tokens will still be valid. All applications that you previously set up to use Amazon Pinpoint or Amazon SNS will continue to work normally. When you call the API for Amazon Pinpoint or Amazon SNS, we initiate a request to the FCM server endpoint directly.

What are the differences between Amazon SNS and Amazon Pinpoint?

Amazon SNS makes it easy for developers to set up, operate, and send notifications at scale, affordably and with a high degree of flexibility. Amazon Pinpoint has many of the same messaging capabilities as Amazon SNS, with the same levels of scalability and flexibility.

The main difference between the two services is that Amazon Pinpoint provides both transactional and targeted messaging capabilities. By using Amazon Pinpoint, marketers and developers can not only send transactional messages to their customers, but can also segment their audiences, create campaigns, and analyze both application and message metrics.

How do I migrate from GCM to FCM?

For more information about migrating from GCM to FCM, see Migrate a GCM Client App for Android to Firebase Cloud Messaging on the Google Developers site.

If you have any questions, please post them in the comments section, or in the Amazon Pinpoint or Amazon SNS forums.

Implementing safe AWS Lambda deployments with AWS CodeDeploy

Post Syndicated from Chris Munns original https://aws.amazon.com/blogs/compute/implementing-safe-aws-lambda-deployments-with-aws-codedeploy/

This post courtesy of George Mao, AWS Senior Serverless Specialist – Solutions Architect

AWS Lambda and AWS CodeDeploy recently made it possible to automatically shift incoming traffic between two function versions based on a preconfigured rollout strategy. This new feature allows you to gradually shift traffic to the new function. If there are any issues with the new code, you can quickly rollback and control the impact to your application.

Previously, you had to manually move 100% of traffic from the old version to the new version. Now, you can have CodeDeploy automatically execute pre- or post-deployment tests and automate a gradual rollout strategy. Traffic shifting is built right into the AWS Serverless Application Model (SAM), making it easy to define and deploy your traffic shifting capabilities. SAM is an extension of AWS CloudFormation that provides a simplified way of defining serverless applications.

In this post, I show you how to use SAM, CloudFormation, and CodeDeploy to accomplish an automated rollout strategy for safe Lambda deployments.

Scenario

For this walkthrough, you write a Lambda application that returns a count of the S3 buckets that you own. You deploy it and use it in production. Later on, you receive requirements that tell you that you need to change your Lambda application to count only buckets that begin with the letter “a”.

Before you make the change, you need to be sure that your new Lambda application works as expected. If it does have issues, you want to minimize the number of impacted users and roll back easily. To accomplish this, you create a deployment process that publishes the new Lambda function, but does not send any traffic to it. You use CodeDeploy to execute a PreTraffic test to ensure that your new function works as expected. After the test succeeds, CodeDeploy automatically shifts traffic gradually to the new version of the Lambda function.

Your Lambda function is exposed as a REST service via an Amazon API Gateway deployment. This makes it easy to test and integrate.

Prerequisites

To execute the SAM and CloudFormation deployment, you must have the following IAM permissions:

  • cloudformation:*
  • lambda:*
  • codedeploy:*
  • iam:create*

You may use the AWS SAM Local CLI or the AWS CLI to package and deploy your Lambda application. If you choose to use SAM Local, be sure to install it onto your system. For more information, see AWS SAM Local Installation.

All of the code used in this post can be found in this GitHub repository: https://github.com/aws-samples/aws-safe-lambda-deployments.

Walkthrough

For this post, use SAM to define your resources because it comes with built-in CodeDeploy support for safe Lambda deployments.  The deployment is handled and automated by CloudFormation.

SAM allows you to define your Serverless applications in a simple and concise fashion, because it automatically creates all necessary resources behind the scenes. For example, if you do not define an execution role for a Lambda function, SAM automatically creates one. SAM also creates the CodeDeploy application necessary to drive the traffic shifting, as well as the IAM service role that CodeDeploy uses to execute all actions.

Create a SAM template

To get started, write your SAM template and call it template.yaml.

AWSTemplateFormatVersion : '2010-09-09'
Transform: AWS::Serverless-2016-10-31
Description: An example SAM template for Lambda Safe Deployments.

Resources:

  returnS3Buckets:
    Type: AWS::Serverless::Function
    Properties:
      Handler: returnS3Buckets.handler
      Runtime: nodejs6.10
      AutoPublishAlias: live
      Policies:
        - Version: "2012-10-17"
          Statement: 
          - Effect: "Allow"
            Action: 
              - "s3:ListAllMyBuckets"
            Resource: '*'
      DeploymentPreference:
          Type: Linear10PercentEvery1Minute
          Hooks:
            PreTraffic: !Ref preTrafficHook
      Events:
        Api:
          Type: Api
          Properties:
            Path: /test
            Method: get

  preTrafficHook:
    Type: AWS::Serverless::Function
    Properties:
      Handler: preTrafficHook.handler
      Policies:
        - Version: "2012-10-17"
          Statement: 
          - Effect: "Allow"
            Action: 
              - "codedeploy:PutLifecycleEventHookExecutionStatus"
            Resource:
              !Sub 'arn:aws:codedeploy:${AWS::Region}:${AWS::AccountId}:deploymentgroup:${ServerlessDeploymentApplication}/*'
        - Version: "2012-10-17"
          Statement: 
          - Effect: "Allow"
            Action: 
              - "lambda:InvokeFunction"
            Resource: !Ref returnS3Buckets.Version
      Runtime: nodejs6.10
      FunctionName: 'CodeDeployHook_preTrafficHook'
      DeploymentPreference:
        Enabled: false
      Timeout: 5
      Environment:
        Variables:
          NewVersion: !Ref returnS3Buckets.Version

This template creates two functions:

  • returnS3Buckets
  • preTrafficHook

The returnS3Buckets function is where your application logic lives. It’s a simple piece of code that uses the AWS SDK for JavaScript in Node.JS to call the Amazon S3 listBuckets API action and return the number of buckets.

'use strict';

var AWS = require('aws-sdk');
var s3 = new AWS.S3();

exports.handler = (event, context, callback) => {
	console.log("I am here! " + context.functionName  +  ":"  +  context.functionVersion);

	s3.listBuckets(function (err, data){
		if(err){
			console.log(err, err.stack);
			callback(null, {
				statusCode: 500,
				body: "Failed!"
			});
		}
		else{
			var allBuckets = data.Buckets;

			console.log("Total buckets: " + allBuckets.length);
			callback(null, {
				statusCode: 200,
				body: allBuckets.length
			});
		}
	});	
}

Review the key parts of the SAM template that defines returnS3Buckets:

  • The AutoPublishAlias attribute instructs SAM to automatically publish a new version of the Lambda function for each new deployment and link it to the live alias.
  • The Policies attribute specifies additional policy statements that SAM adds onto the automatically generated IAM role for this function. The first statement provides the function with permission to call listBuckets.
  • The DeploymentPreference attribute configures the type of rollout pattern to use. In this case, you are shifting traffic in a linear fashion, moving 10% of traffic every minute to the new version. For more information about supported patterns, see Serverless Application Model: Traffic Shifting Configurations.
  • The Hooks attribute specifies that you want to execute the preTrafficHook Lambda function before CodeDeploy automatically begins shifting traffic. This function should perform validation testing on the newly deployed Lambda version. This function invokes the new Lambda function and checks the results. If you’re satisfied with the tests, instruct CodeDeploy to proceed with the rollout via an API call to: codedeploy.putLifecycleEventHookExecutionStatus.
  • The Events attribute defines an API-based event source that can trigger this function. It accepts requests on the /test path using an HTTP GET method.
'use strict';

const AWS = require('aws-sdk');
const codedeploy = new AWS.CodeDeploy({apiVersion: '2014-10-06'});
var lambda = new AWS.Lambda();

exports.handler = (event, context, callback) => {

	console.log("Entering PreTraffic Hook!");
	
	// Read the DeploymentId & LifecycleEventHookExecutionId from the event payload
    var deploymentId = event.DeploymentId;
	var lifecycleEventHookExecutionId = event.LifecycleEventHookExecutionId;

	var functionToTest = process.env.NewVersion;
	console.log("Testing new function version: " + functionToTest);

	// Perform validation of the newly deployed Lambda version
	var lambdaParams = {
		FunctionName: functionToTest,
		InvocationType: "RequestResponse"
	};

	var lambdaResult = "Failed";
	lambda.invoke(lambdaParams, function(err, data) {
		if (err){	// an error occurred
			console.log(err, err.stack);
			lambdaResult = "Failed";
		}
		else{	// successful response
			var result = JSON.parse(data.Payload);
			console.log("Result: " +  JSON.stringify(result));

			// Check the response for valid results
			// The response will be a JSON payload with statusCode and body properties. ie:
			// {
			//		"statusCode": 200,
			//		"body": 51
			// }
			if(result.body == 9){	
				lambdaResult = "Succeeded";
				console.log ("Validation testing succeeded!");
			}
			else{
				lambdaResult = "Failed";
				console.log ("Validation testing failed!");
			}

			// Complete the PreTraffic Hook by sending CodeDeploy the validation status
			var params = {
				deploymentId: deploymentId,
				lifecycleEventHookExecutionId: lifecycleEventHookExecutionId,
				status: lambdaResult // status can be 'Succeeded' or 'Failed'
			};
			
			// Pass AWS CodeDeploy the prepared validation test results.
			codedeploy.putLifecycleEventHookExecutionStatus(params, function(err, data) {
				if (err) {
					// Validation failed.
					console.log('CodeDeploy Status update failed');
					console.log(err, err.stack);
					callback("CodeDeploy Status update failed");
				} else {
					// Validation succeeded.
					console.log('Codedeploy status updated successfully');
					callback(null, 'Codedeploy status updated successfully');
				}
			});
		}  
	});
}

The hook is hardcoded to check that the number of S3 buckets returned is 9.

Review the key parts of the SAM template that defines preTrafficHook:

  • The Policies attribute specifies additional policy statements that SAM adds onto the automatically generated IAM role for this function. The first statement provides permissions to call the CodeDeploy PutLifecycleEventHookExecutionStatus API action. The second statement provides permissions to invoke the specific version of the returnS3Buckets function to test
  • This function has traffic shifting features disabled by setting the DeploymentPreference option to false.
  • The FunctionName attribute explicitly tells CloudFormation what to name the function. Otherwise, CloudFormation creates the function with the default naming convention: [stackName]-[FunctionName]-[uniqueID].  Name the function with the “CodeDeployHook_” prefix because the CodeDeployServiceRole role only allows InvokeFunction on functions named with that prefix.
  • Set the Timeout attribute to allow enough time to complete your validation tests.
  • Use an environment variable to inject the ARN of the newest deployed version of the returnS3Buckets function. The ARN allows the function to know the specific version to invoke and perform validation testing on.

Deploy the function

Your SAM template is all set and the code is written—you’re ready to deploy the function for the first time. Here’s how to do it via the SAM CLI. Replace “sam” with “cloudformation” to use CloudFormation instead.

First, package the function. This command returns a CloudFormation importable file, packaged.yaml.

sam package –template-file template.yaml –s3-bucket mybucket –output-template-file packaged.yaml

Now deploy everything:

sam deploy –template-file packaged.yaml –stack-name mySafeDeployStack –capabilities CAPABILITY_IAM

At this point, both Lambda functions have been deployed within the CloudFormation stack mySafeDeployStack. The returnS3Buckets has been deployed as Version 1:

SAM automatically created a few things, including the CodeDeploy application, with the deployment pattern that you specified (Linear10PercentEvery1Minute). There is currently one deployment group, with no action, because no deployments have occurred. SAM also created the IAM service role that this CodeDeploy application uses:

There is a single managed policy attached to this role, which allows CodeDeploy to invoke any Lambda function that begins with “CodeDeployHook_”.

An API has been set up called safeDeployStack. It targets your Lambda function with the /test resource using the GET method. When you test the endpoint, API Gateway executes the returnS3Buckets function and it returns the number of S3 buckets that you own. In this case, it’s 51.

Publish a new Lambda function version

Now implement the requirements change, which is to make returnS3Buckets count only buckets that begin with the letter “a”. The code now looks like the following (see returnS3BucketsNew.js in GitHub):

'use strict';

var AWS = require('aws-sdk');
var s3 = new AWS.S3();

exports.handler = (event, context, callback) => {
	console.log("I am here! " + context.functionName  +  ":"  +  context.functionVersion);

	s3.listBuckets(function (err, data){
		if(err){
			console.log(err, err.stack);
			callback(null, {
				statusCode: 500,
				body: "Failed!"
			});
		}
		else{
			var allBuckets = data.Buckets;

			console.log("Total buckets: " + allBuckets.length);
			//callback(null, allBuckets.length);

			//  New Code begins here
			var counter=0;
			for(var i  in allBuckets){
				if(allBuckets[i].Name[0] === "a")
					counter++;
			}
			console.log("Total buckets starting with a: " + counter);

			callback(null, {
				statusCode: 200,
				body: counter
			});
			
		}
	});	
}

Repackage and redeploy with the same two commands as earlier:

sam package –template-file template.yaml –s3-bucket mybucket –output-template-file packaged.yaml
	
sam deploy –template-file packaged.yaml –stack-name mySafeDeployStack –capabilities CAPABILITY_IAM

CloudFormation understands that this is a stack update instead of an entirely new stack. You can see that reflected in the CloudFormation console:

During the update, CloudFormation deploys the new Lambda function as version 2 and adds it to the “live” alias. There is no traffic routing there yet. CodeDeploy now takes over to begin the safe deployment process.

The first thing CodeDeploy does is invoke the preTrafficHook function. Verify that this happened by reviewing the Lambda logs and metrics:

The function should progress successfully, invoke Version 2 of returnS3Buckets, and finally invoke the CodeDeploy API with a success code. After this occurs, CodeDeploy begins the predefined rollout strategy. Open the CodeDeploy console to review the deployment progress (Linear10PercentEvery1Minute):

Verify the traffic shift

During the deployment, verify that the traffic shift has started to occur by running the test periodically. As the deployment shifts towards the new version, a larger percentage of the responses return 9 instead of 51. These numbers match the S3 buckets.

A minute later, you see 10% more traffic shifting to the new version. The whole process takes 10 minutes to complete. After completion, open the Lambda console and verify that the “live” alias now points to version 2:

After 10 minutes, the deployment is complete and CodeDeploy signals success to CloudFormation and completes the stack update.

Check the results

If you invoke the function alias manually, you see the results of the new implementation.

aws lambda invoke –function [lambda arn to live alias] out.txt

You can also execute the prod stage of your API and verify the results by issuing an HTTP GET to the invoke URL:

Summary

This post has shown you how you can safely automate your Lambda deployments using the Lambda traffic shifting feature. You used the Serverless Application Model (SAM) to define your Lambda functions and configured CodeDeploy to manage your deployment patterns. Finally, you used CloudFormation to automate the deployment and updates to your function and PreTraffic hook.

Now that you know all about this new feature, you’re ready to begin automating Lambda deployments with confidence that things will work as designed. I look forward to hearing about what you’ve built with the AWS Serverless Platform.

Achieving Major Stability and Performance Improvements in Yahoo Mail with a Novel Redux Architecture

Post Syndicated from mikesefanov original https://yahooeng.tumblr.com/post/173062946866

yahoodevelopers:

By Mohit Goenka, Gnanavel Shanmugam, and Lance Welsh

At Yahoo Mail, we’re constantly striving to upgrade our product experience. We do this not only by adding new features based on our members’ feedback, but also by providing the best technical solutions to power the most engaging experiences. As such, we’ve recently introduced a number of novel and unique revisions to the way in which we use Redux that have resulted in significant stability and performance improvements. Developers may find our methods useful in achieving similar results in their apps.

Improvements to product metrics

Last year Yahoo Mail implemented a brand new architecture using Redux. Since then, we have transformed the overall architecture to reduce latencies in various operations, reduce JavaScript exceptions, and better synchronized states. As a result, the product is much faster and more stable.

Stability improvements:

  • when checking for new emails – 20%
  • when reading emails – 30%
  • when sending emails – 20%

Performance improvements:

  • 10% improvement in page load performance
  • 40% improvement in frame rendering time

We have also reduced API calls by approximately 20%.

How we use Redux in Yahoo Mail

Redux architecture is reliant on one large store that represents the application state. In a Redux cycle, action creators dispatch actions to change the state of the store. React Components then respond to those state changes. We’ve made some modifications on top of this architecture that are atypical in the React-Redux community.

For instance, when fetching data over the network, the traditional methodology is to use Thunk middleware. Yahoo Mail fetches data over the network from our API. Thunks would create an unnecessary and undesirable dependency between the action creators and our API. If and when the API changes, the action creators must then also change. To keep these concerns separate we dispatch the action payload from the action creator to store them in the Redux state for later processing by “action syncers”. Action syncers use the payload information from the store to make requests to the API and process responses. In other words, the action syncers form an API layer by interacting with the store. An additional benefit to keeping the concerns separate is that the API layer can change as the backend changes, thereby preventing such changes from bubbling back up into the action creators and components. This also allowed us to optimize the API calls by batching, deduping, and processing the requests only when the network is available. We applied similar strategies for handling other side effects like route handling and instrumentation. Overall, action syncers helped us to reduce our API calls by ~20% and bring down API errors by 20-30%.

Another change to the normal Redux architecture was made to avoid unnecessary props. The React-Redux community has learned to avoid passing unnecessary props from high-level components through multiple layers down to lower-level components (prop drilling) for rendering. We have introduced action enhancers middleware to avoid passing additional unnecessary props that are purely used when dispatching actions. Action enhancers add data to the action payload so that data does not have to come from the component when dispatching the action. This avoids the component from having to receive that data through props and has improved frame rendering by ~40%. The use of action enhancers also avoids writing utility functions to add commonly-used data to each action from action creators.

image

In our new architecture, the store reducers accept the dispatched action via action enhancers to update the state. The store then updates the UI, completing the action cycle. Action syncers then initiate the call to the backend APIs to synchronize local changes.

Conclusion

Our novel use of Redux in Yahoo Mail has led to significant user-facing benefits through a more performant application. It has also reduced development cycles for new features due to its simplified architecture. We’re excited to share our work with the community and would love to hear from anyone interested in learning more.

AWS AppSync – Production-Ready with Six New Features

Post Syndicated from Jeff Barr original https://aws.amazon.com/blogs/aws/aws-appsync-production-ready-with-six-new-features/

If you build (or want to build) data-driven web and mobile apps and need real-time updates and the ability to work offline, you should take a look at AWS AppSync. Announced in preview form at AWS re:Invent 2017 and described in depth here, AWS AppSync is designed for use in iOS, Android, JavaScript, and React Native apps. AWS AppSync is built around GraphQL, an open, standardized query language that makes it easy for your applications to request the precise data that they need from the cloud.

I’m happy to announce that the preview period is over and that AWS AppSync is now generally available and production-ready, with six new features that will simplify and streamline your application development process:

Console Log Access – You can now see the CloudWatch Logs entries that are created when you test your GraphQL queries, mutations, and subscriptions from within the AWS AppSync Console.

Console Testing with Mock Data – You can now create and use mock context objects in the console for testing purposes.

Subscription Resolvers – You can now create resolvers for AWS AppSync subscription requests, just as you can already do for query and mutate requests.

Batch GraphQL Operations for DynamoDB – You can now make use of DynamoDB’s batch operations (BatchGetItem and BatchWriteItem) across one or more tables. in your resolver functions.

CloudWatch Support – You can now use Amazon CloudWatch Metrics and CloudWatch Logs to monitor calls to the AWS AppSync APIs.

CloudFormation Support – You can now define your schemas, data sources, and resolvers using AWS CloudFormation templates.

A Brief AppSync Review
Before diving in to the new features, let’s review the process of creating an AWS AppSync API, starting from the console. I click Create API to begin:

I enter a name for my API and (for demo purposes) choose to use the Sample schema:

The schema defines a collection of GraphQL object types. Each object type has a set of fields, with optional arguments:

If I was creating an API of my own I would enter my schema at this point. Since I am using the sample, I don’t need to do this. Either way, I click on Create to proceed:

The GraphQL schema type defines the entry points for the operations on the data. All of the data stored on behalf of a particular schema must be accessible using a path that begins at one of these entry points. The console provides me with an endpoint and key for my API:

It also provides me with guidance and a set of fully functional sample apps that I can clone:

When I clicked Create, AWS AppSync created a pair of Amazon DynamoDB tables for me. I can click Data Sources to see them:

I can also see and modify my schema, issue queries, and modify an assortment of settings for my API.

Let’s take a quick look at each new feature…

Console Log Access
The AWS AppSync Console already allows me to issue queries and to see the results, and now provides access to relevant log entries.In order to see the entries, I must enable logs (as detailed below), open up the LOGS, and check the checkbox. Here’s a simple mutation query that adds a new event. I enter the query and click the arrow to test it:

I can click VIEW IN CLOUDWATCH for a more detailed view:

To learn more, read Test and Debug Resolvers.

Console Testing with Mock Data
You can now create a context object in the console where it will be passed to one of your resolvers for testing purposes. I’ll add a testResolver item to my schema:

Then I locate it on the right-hand side of the Schema page and click Attach:

I choose a data source (this is for testing and the actual source will not be accessed), and use the Put item mapping template:

Then I click Select test context, choose Create New Context, assign a name to my test content, and click Save (as you can see, the test context contains the arguments from the query along with values to be returned for each field of the result):

After I save the new Resolver, I click Test to see the request and the response:

Subscription Resolvers
Your AWS AppSync application can monitor changes to any data source using the @aws_subscribe GraphQL schema directive and defining a Subscription type. The AWS AppSync client SDK connects to AWS AppSync using MQTT over Websockets and the application is notified after each mutation. You can now attach resolvers (which convert GraphQL payloads into the protocol needed by the underlying storage system) to your subscription fields and perform authorization checks when clients attempt to connect. This allows you to perform the same fine grained authorization routines across queries, mutations, and subscriptions.

To learn more about this feature, read Real-Time Data.

Batch GraphQL Operations
Your resolvers can now make use of DynamoDB batch operations that span one or more tables in a region. This allows you to use a list of keys in a single query, read records multiple tables, write records in bulk to multiple tables, and conditionally write or delete related records across multiple tables.

In order to use this feature the IAM role that you use to access your tables must grant access to DynamoDB’s BatchGetItem and BatchPutItem functions.

To learn more, read the DynamoDB Batch Resolvers tutorial.

CloudWatch Logs Support
You can now tell AWS AppSync to log API requests to CloudWatch Logs. Click on Settings and Enable logs, then choose the IAM role and the log level:

CloudFormation Support
You can use the following CloudFormation resource types in your templates to define AWS AppSync resources:

AWS::AppSync::GraphQLApi – Defines an AppSync API in terms of a data source (an Amazon Elasticsearch Service domain or a DynamoDB table).

AWS::AppSync::ApiKey – Defines the access key needed to access the data source.

AWS::AppSync::GraphQLSchema – Defines a GraphQL schema.

AWS::AppSync::DataSource – Defines a data source.

AWS::AppSync::Resolver – Defines a resolver by referencing a schema and a data source, and includes a mapping template for requests.

Here’s a simple schema definition in YAML form:

  AppSyncSchema:
    Type: "AWS::AppSync::GraphQLSchema"
    DependsOn:
      - AppSyncGraphQLApi
    Properties:
      ApiId: !GetAtt AppSyncGraphQLApi.ApiId
      Definition: |
        schema {
          query: Query
          mutation: Mutation
        }
        type Query {
          singlePost(id: ID!): Post
          allPosts: [Post]
        }
        type Mutation {
          putPost(id: ID!, title: String!): Post
        }
        type Post {
          id: ID!
          title: String!
        }

Available Now
These new features are available now and you can start using them today! Here are a couple of blog posts and other resources that you might find to be of interest:

Jeff;

 

 

More power to your Pi

Post Syndicated from James Adams original https://www.raspberrypi.org/blog/pi-power-supply-chip/

It’s been just over three weeks since we launched the new Raspberry Pi 3 Model B+. Although the product is branded Raspberry Pi 3B+ and not Raspberry Pi 4, a serious amount of engineering was involved in creating it. The wireless networking, USB/Ethernet hub, on-board power supplies, and BCM2837 chip were all upgraded: together these represent almost all the circuitry on the board! Today, I’d like to tell you about the work that has gone into creating a custom power supply chip for our newest computer.

Raspberry Pi 3 Model B+, with custome power supply chip

The new Raspberry Pi 3B+, sporting a new, custom power supply chip (bottom left-hand corner)

Successful launch

The Raspberry Pi 3B+ has been well received, and we’ve enjoyed hearing feedback from the community as well as reading the various reviews and articles highlighting the solid improvements in wireless networking, Ethernet, CPU, and thermal performance of the new board. Gareth Halfacree’s post here has some particularly nice graphs showing the increased performance as well as how the Pi 3B+ keeps cool under load due to the new CPU package that incorporates a metal heat spreader. The Raspberry Pi production lines at the Sony UK Technology Centre are running at full speed, and it seems most people who want to get hold of the new board are able to find one in stock.

Powering your Pi

One of the most critical but often under-appreciated elements of any electronic product, particularly one such as Raspberry Pi with lots of complex on-board silicon (processor, networking, high-speed memory), is the power supply. In fact, the Raspberry Pi 3B+ has no fewer than six different voltage rails: two at 3.3V — one special ‘quiet’ one for audio, and one for everything else; 1.8V; 1.2V for the LPDDR2 memory; and 1.2V nominal for the CPU core. Note that the CPU voltage is actually raised and lowered on the fly as the speed of the CPU is increased and decreased depending on how hard the it is working. The sixth rail is 5V, which is the master supply that all the others are created from, and the output voltage for the four downstream USB ports; this is what the mains power adaptor is supplying through the micro USB power connector.

Power supply primer

There are two common classes of power supply circuits: linear regulators and switching regulators. Linear regulators work by creating a lower, regulated voltage from a higher one. In simple terms, they monitor the output voltage against an internally generated reference and continually change their own resistance to keep the output voltage constant. Switching regulators work in a different way: they ‘pump’ energy by first storing the energy coming from the source supply in a reactive component (usually an inductor, sometimes a capacitor) and then releasing it to the regulated output supply. The switches in switching regulators effect this energy transfer by first connecting the inductor (or capacitor) to store the source energy, and then switching the circuit so the energy is released to its destination.

Linear regulators produce smoother, less noisy output voltages, but they can only convert to a lower voltage, and have to dissipate energy to do so. The higher the output current and the voltage difference across them is, the more energy is lost as heat. On the other hand, switching supplies can, depending on their design, convert any voltage to any other voltage and can be much more efficient (efficiencies of 90% and above are not uncommon). However, they are more complex and generate noisier output voltages.

Designers use both types of regulators depending on the needs of the downstream circuit: for low-voltage drops, low current, or low noise, linear regulators are usually the right choice, while switching regulators are used for higher power or when efficiency of conversion is required. One of the simplest switching-mode power supply circuits is the buck converter, used to create a lower voltage from a higher one, and this is what we use on the Pi.

A history lesson

The BCM2835 processor chip (found on the original Raspberry Pi Model B and B+, as well as on the Zero products) has on-chip power supplies: one switch-mode regulator for the core voltage, as well as a linear one for the LPDDR2 memory supply. This meant that in addition to 5V, we only had to provide 3.3V and 1.8V on the board, which was relatively simple to do using cheap, off-the-shelf parts.

Pi Zero sporting a BCM2835 processor which only needs 2 external switchers (the components clustered behind the camera port)

When we moved to the BCM2836 for Raspberry Pi Model 2 (and subsequently to the BCM2837A1 and B0 for Raspberry Pi 3B and 3B+), the core supply and the on-chip LPDDR2 memory supply were not up to the job of supplying the extra processor cores and larger memory, so we removed them. (We also used the recovered chip area to help fit in the new quad-core ARM processors.) The upshot of this was that we had to supply these power rails externally for the Raspberry Pi 2 and models thereafter. Moreover, we also had to provide circuitry to sequence them correctly in order to control exactly when they power up compared to the other supplies on the board.

Power supply design is tricky (but critical)

Raspberry Pi boards take in 5V from the micro USB socket and have to generate the other required supplies from this. When 5V is first connected, each of these other supplies must ‘start up’, meaning go from ‘off’, or 0V, to their correct voltage in some short period of time. The order of the supplies starting up is often important: commonly, there are structures inside a chip that form diodes between supply rails, and bringing supplies up in the wrong order can sometimes ‘turn on’ these diodes, causing them to conduct, with undesirable consequences. Silicon chips come with a data sheet specifying what supplies (voltages and currents) are needed and whether they need to be low-noise, in what order they must power up (and in some cases down), and sometimes even the rate at which the voltages must power up and down.

A Pi3. Power supply components are clustered bottom left next to the micro USB, middle (above LPDDR2 chip which is on the bottom of the PCB) and above the A/V jack.

In designing the power chain for the Pi 2 and 3, the sequencing was fairly straightforward: power rails power up in order of voltage (5V, 3.3V, 1.8V, 1.2V). However, the supplies were all generated with individual, discrete devices. Therefore, I spent quite a lot of time designing circuitry to control the sequencing — even with some design tricks to reduce component count, quite a few sequencing components are required. More complex systems generally use a Power Management Integrated Circuit (PMIC) with multiple supplies on a single chip, and many different PMIC variants are made by various manufacturers. Since Raspberry Pi 2 days, I was looking for a suitable PMIC to simplify the Pi design, but invariably (and somewhat counter-intuitively) these were always too expensive compared to my discrete solution, usually because they came with more features than needed.

One device to rule them all

It was way back in May 2015 when I first chatted to Peter Coyle of Exar (Exar were bought by MaxLinear in 2017) about power supply products for Raspberry Pi. We didn’t find a product match then, but in June 2016 Peter, along with Tuomas Hollman and Trevor Latham, visited to pitch the possibility of building a custom power management solution for us.

I was initially sceptical that it could be made cheap enough. However, our discussion indicated that if we could tailor the solution to just what we needed, it could be cost-effective. Over the coming weeks and months, we honed a specification we agreed on from the initial sketches we’d made, and Exar thought they could build it for us at the target price.

The chip we designed would contain all the key supplies required for the Pi on one small device in a cheap QFN package, and it would also perform the required sequencing and voltage monitoring. Moreover, the chip would be flexible to allow adjustment of supply voltages from their default values via I2C; the largest supply would be capable of being adjusted quickly to perform the dynamic core voltage changes needed in order to reduce voltage to the processor when it is idling (to save power), and to boost voltage to the processor when running at maximum speed (1.4 GHz). The supplies on the chip would all be generously specified and could deliver significantly more power than those used on the Raspberry Pi 3. All in all, the chip would contain four switching-mode converters and one low-current linear regulator, this last one being low-noise for the audio circuitry.

The MXL7704 chip

The project was a great success: MaxLinear delivered working samples of first silicon at the end of May 2017 (almost exactly a year after we had kicked off the project), and followed through with production quantities in December 2017 in time for the Raspberry Pi 3B+ production ramp.

The team behind the power supply chip on the Raspberry Pi 3 Model B+ (group of six men, two of whom are holding Raspberry Pi boards)

Front row: Roger with the very first Pi 3B+ prototypes and James with a MXL7704 development board hacked to power a Pi 3. Back row left to right: Will Torgerson, Trevor Latham, Peter Coyle, Tuomas Hollman.

The MXL7704 device has been key to reducing Pi board complexity and therefore overall bill of materials cost. Furthermore, by being able to deliver more power when needed, it has also been essential to increasing the speed of the (newly packaged) BCM2837B0 processor on the 3B+ to 1.4GHz. The result is improvements to both the continuous output current to the CPU (from 3A to 4A) and to the transient performance (i.e. the chip has helped to reduce the ‘transient response’, which is the change in supply voltage due to a sudden current spike that occurs when the processor suddenly demands a large current in a few nanoseconds, as modern CPUs tend to do).

With the MXL7704, the power supply circuitry on the 3B+ is now a lot simpler than the Pi 3B design. This new supply also provides the LPDDR2 memory voltage directly from a switching regulator rather than using linear regulators like the Pi 3, thereby improving energy efficiency. This helps to somewhat offset the extra power that the faster Ethernet, wireless networking, and processor consume. A pleasing side effect of using the new chip is the symmetric board layout of the regulators — it’s easy to see the four switching-mode supplies, given away by four similar-looking blobs (three grey and one brownish), which are the inductors.

Close-up of the power supply chip on the Raspberry Pi 3 Model B+

The Pi 3B+ PMIC MXL7704 — pleasingly symmetric

Kudos

It takes a lot of effort to design a new chip from scratch and get it all the way through to production — we are very grateful to the team at MaxLinear for their hard work, dedication, and enthusiasm. We’re also proud to have created something that will not only power Raspberry Pis, but will also be useful for other product designs: it turns out when you have a low-cost and flexible device, it can be used for many things — something we’re fairly familiar with here at Raspberry Pi! For the curious, the product page (including the data sheet) for the MXL7704 chip is here. Particular thanks go to Peter Coyle, Tuomas Hollman, and Trevor Latham, and also to Jon Cronk, who has been our contact in the US and has had to get up early to attend all our conference calls!

The MXL7704 design team celebrating on Pi Day — it takes a lot of people to design a chip!

I hope you liked reading about some of the effort that has gone into creating the new Pi. It’s nice to finally have a chance to tell people about some of the (increasingly complex) technical work that makes building a $35 computer possible — we’re very pleased with the Raspberry Pi 3B+, and we hope you enjoy using it as much as we’ve enjoyed creating it!

The post More power to your Pi appeared first on Raspberry Pi.

A geometric Rust adventure

Post Syndicated from Eevee original https://eev.ee/blog/2018/03/30/a-geometric-rust-adventure/

Hi. Yes. Sorry. I’ve been trying to write this post for ages, but I’ve also been working on a huge writing project, and apparently I have a very limited amount of writing mana at my disposal. I think this is supposed to be a Patreon reward from January. My bad. I hope it’s super great to make up for the wait!

I recently ported some math code from C++ to Rust in an attempt to do a cool thing with Doom. Here is my story.

The problem

I presented it recently as a conundrum (spoilers: I solved it!), but most of those details are unimportant.

The short version is: I have some shapes. I want to find their intersection.

Really, I want more than that: I want to drop them all on a canvas, intersect everything with everything, and pluck out all the resulting polygons. The input is a set of cookie cutters, and I want to press them all down on the same sheet of dough and figure out what all the resulting contiguous pieces are. And I want to know which cookie cutter(s) each piece came from.

But intersection is a good start.

Example of the goal.  Given two squares that overlap at their corners, I want to find the small overlap piece, plus the two L-shaped pieces left over from each square

I’m carefully referring to the input as shapes rather than polygons, because each one could be a completely arbitrary collection of lines. Obviously there’s not much you can do with shapes that aren’t even closed, but at the very least, I need to handle concavity and multiple disconnected polygons that together are considered a single input.

This is a non-trivial problem with a lot of edge cases, and offhand I don’t know how to solve it robustly. I’m not too eager to go figure it out from scratch, so I went hunting for something I could build from.

(Infuriatingly enough, I can just dump all the shapes out in an SVG file and any SVG viewer can immediately solve the problem, but that doesn’t quite help me. Though I have had a few people suggest I just rasterize the whole damn problem, and after all this, I’m starting to think they may have a point.)

Alas, I couldn’t find a Rust library for doing this. I had a hard time finding any library for doing this that wasn’t a massive fully-featured geometry engine. (I could’ve used that, but I wanted to avoid non-Rust dependencies if possible, since distributing software is already enough of a nightmare.)

A Twitter follower directed me towards a paper that described how to do very nearly what I wanted and nothing else: “A simple algorithm for Boolean operations on polygons” by F. Martínez (2013). Being an academic paper, it’s trapped in paywall hell; sorry about that. (And as I understand it, none of the money you’d pay to get the paper would even go to the authors? Is that right? What a horrible and predatory system for discovering and disseminating knowledge.)

The paper isn’t especially long, but it does describe an awful lot of subtle details and is mostly written in terms of its own reference implementation. Rather than write my own implementation based solely on the paper, I decided to try porting the reference implementation from C++ to Rust.

And so I fell down the rabbit hole.

The basic algorithm

Thankfully, the author has published the sample code on his own website, if you want to follow along. (It’s the bottom link; the same author has, confusingly, published two papers on the same topic with similar titles, four years apart.)

If not, let me describe the algorithm and how the code is generally laid out. The algorithm itself is based on a sweep line, where a vertical line passes across the plane and ✨ does stuff ✨ as it encounters various objects. This implementation has no physical line; instead, it keeps track of which segments from the original polygon would be intersecting the sweep line, which is all we really care about.

A vertical line is passing rightwards over a couple intersecting shapes.  The line current intersects two of the shapes' sides, and these two sides are the "sweep list"

The code is all bundled inside a class with only a single public method, run, because… that’s… more object-oriented, I guess. There are several helper methods, and state is stored in some attributes. A rough outline of run is:

  1. Run through all the line segments in both input polygons. For each one, generate two SweepEvents (one for each endpoint) and add them to a std::deque for storage.

    Add pointers to the two SweepEvents to a std::priority_queue, the event queue. This queue uses a custom comparator to order the events from left to right, so the top element is always the leftmost endpoint.

  2. Loop over the event queue (where an “event” means the sweep line passed over the left or right end of a segment). Encountering a left endpoint means the sweep line is newly touching that segment, so add it to a std::set called the sweep list. An important point is that std::set is ordered, and the sweep list uses a comparator that keeps segments in order vertically.

    Encountering a right endpoint means the sweep line is leaving a segment, so that segment is removed from the sweep list.

  3. When a segment is added to the sweep list, it may have up to two neighbors: the segment above it and the segment below it. Call possibleIntersection to check whether it intersects either of those neighbors. (This is nearly sufficient to find all intersections, which is neat.)

  4. If possibleIntersection detects an intersection, it will split each segment into two pieces then and there. The old segment is shortened in-place to become the left part, and a new segment is created for the right part. The new endpoints at the point of intersection are added to the event queue.

  5. Some bookkeeping is done along the way to track which original polygons each segment is inside, and eventually the segments are reconstructed into new polygons.

Hopefully that’s enough to follow along. It took me an inordinately long time to tease this out. The comments aren’t especially helpful.

1
    std::deque<SweepEvent> eventHolder;    // It holds the events generated during the computation of the boolean operation

Syntax and basic semantics

The first step was to get something that rustc could at least parse, which meant translating C++ syntax to Rust syntax.

This was surprisingly straightforward! C++ classes become Rust structs. (There was no inheritance here, thankfully.) All the method declarations go away. Method implementations only need to be indented and wrapped in impl.

I did encounter some unnecessarily obtuse uses of the ternary operator:

1
(prevprev != sl.begin()) ? --prevprev : prevprev = sl.end();

Rust doesn’t have a ternary — you can use a regular if block as an expression — so I expanded these out.

C++ switch blocks become Rust match blocks, but otherwise function basically the same. Rust’s enums are scoped (hallelujah), so I had to explicitly spell out where enum values came from.

The only really annoying part was changing function signatures; C++ types don’t look much at all like Rust types, save for the use of angle brackets. Rust also doesn’t pass by implicit reference, so I needed to sprinkle a few &s around.

I would’ve had a much harder time here if this code had relied on any remotely esoteric C++ functionality, but thankfully it stuck to pretty vanilla features.

Language conventions

This is a geometry problem, so the sample code unsurprisingly has its own home-grown point type. Rather than port that type to Rust, I opted to use the popular euclid crate. Not only is it code I didn’t have to write, but it already does several things that the C++ code was doing by hand inline, like dot products and cross products. And all I had to do was add one line to Cargo.toml to use it! I have no idea how anyone writes C or C++ without a package manager.

The C++ code used getters, i.e. point.x (). I’m not a huge fan of getters, though I do still appreciate the need for them in lowish-level systems languages where you want to future-proof your API and the language wants to keep a clear distinction between attribute access and method calls. But this is a point, which is nothing more than two of the same numeric type glued together; what possible future logic might you add to an accessor? The euclid authors appear to side with me and leave the coordinates as public fields, so I took great joy in removing all the superfluous parentheses.

Polygons are represented with a Polygon class, which has some number of Contours. A contour is a single contiguous loop. Something you’d usually think of as a polygon would only have one, but a shape with a hole would have two: one for the outside, one for the inside. The weird part of this arrangement was that Polygon implemented nearly the entire STL container interface, then waffled between using it and not using it throughout the rest of the code. Rust lets anything in the same module access non-public fields, so I just skipped all that and used polygon.contours directly. Hell, I think I made contours public.

Finally, the SweepEvent type has a pol field that’s declared as an enum PolygonType (either SUBJECT or CLIPPING, to indicate which of the two inputs it is), but then some other code uses the same field as a numeric index into a polygon’s contours. Boy I sure do love static typing where everything’s a goddamn integer. I wanted to extend the algorithm to work on arbitrarily many input polygons anyway, so I scrapped the enum and this became a usize.


Then I got to all the uses of STL. I have only a passing familiarity with the C++ standard library, and this code actually made modest use of it, which caused some fun days-long misunderstandings.

As mentioned, the SweepEvents are stored in a std::deque, which is never read from. It took me a little thinking to realize that the deque was being used as an arena: it’s the canonical home for the structs so pointers to them can be tossed around freely. (It can’t be a std::vector, because that could reallocate and invalidate all the pointers; std::deque is probably a doubly-linked list, and guarantees no reallocation.)

Rust’s standard library does have a doubly-linked list type, but I knew I’d run into ownership hell here later anyway, so I think I replaced it with a Rust Vec to start with. It won’t compile either way, so whatever. We’ll get back to this in a moment.

The list of segments currently intersecting the sweep line is stored in a std::set. That type is explicitly ordered, which I’m very glad I knew already. Rust has two set types, HashSet and BTreeSet; unsurprisingly, the former is unordered and the latter is ordered. Dropping in BTreeSet and fixing some method names got me 90% of the way there.

Which brought me to the other 90%. See, the C++ code also relies on finding nodes adjacent to the node that was just inserted, via STL iterators.

1
2
3
next = prev = se->posSL = it = sl.insert(se).first;
(prev != sl.begin()) ? --prev : prev = sl.end();
++next;

I freely admit I’m bad at C++, but this seems like something that could’ve used… I don’t know, 1 comment. Or variable names more than two letters long. What it actually does is:

  1. Add the current sweep event (se) to the sweep list (sl), which returns a pair whose first element is an iterator pointing at the just-inserted event.

  2. Copies that iterator to several other variables, including prev and next.

  3. If the event was inserted at the beginning of the sweep list, set prev to the sweep list’s end iterator, which in C++ is a legal-but-invalid iterator meaning “the space after the end” or something. This is checked for in later code, to see if there is a previous event to look at. Otherwise, decrement prev, so it’s now pointing at the event immediately before the inserted one.

  4. Increment next normally. If the inserted event is last, then this will bump next to the end iterator anyway.

In other words, I need to get the previous and next elements from a BTreeSet. Rust does have bidirectional iterators, which BTreeSet supports… but BTreeSet::insert only returns a bool telling me whether or not anything was inserted, not the position. I came up with this:

1
2
3
let mut maybe_below = active_segments.range(..segment).last().map(|v| *v);
let mut maybe_above = active_segments.range(segment..).next().map(|v| *v);
active_segments.insert(segment);

The range method returns an iterator over a subset of the tree. The .. syntax makes a range (where the right endpoint is exclusive), so ..segment finds the part of the tree before the new segment, and segment.. finds the part of the tree after it. (The latter would start with the segment itself, except I haven’t inserted it yet, so it’s not actually there.)

Then the standard next() and last() methods on bidirectional iterators find me the element I actually want. But the iterator might be empty, so they both return an Option. Also, iterators tend to return references to their contents, but in this case the contents are already references, and I don’t want a double reference, so the map call dereferences one layer — but only if the Option contains a value. Phew!

This is slightly less efficient than the C++ code, since it has to look up where segment goes three times rather than just one. I might be able to get it down to two with some more clever finagling of the iterator, but microsopic performance considerations were a low priority here.

Finally, the event queue uses a std::priority_queue to keep events in a desired order and efficiently pop the next one off the top.

Except priority queues act like heaps, where the greatest (i.e., last) item is made accessible.

Sorting out sorting

C++ comparison functions return true to indicate that the first argument is less than the second argument. Sweep events occur from left to right. You generally implement sorts so that the first thing comes, erm, first.

But sweep events go in a priority queue, and priority queues surface the last item, not the first. This C++ code handled this minor wrinkle by implementing its comparison backwards.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
struct SweepEventComp : public std::binary_function<SweepEvent, SweepEvent, bool> { // for sorting sweep events
// Compare two sweep events
// Return true means that e1 is placed at the event queue after e2, i.e,, e1 is processed by the algorithm after e2
bool operator() (const SweepEvent* e1, const SweepEvent* e2)
{
    if (e1->point.x () > e2->point.x ()) // Different x-coordinate
        return true;
    if (e2->point.x () > e1->point.x ()) // Different x-coordinate
        return false;
    if (e1->point.y () != e2->point.y ()) // Different points, but same x-coordinate. The event with lower y-coordinate is processed first
        return e1->point.y () > e2->point.y ();
    if (e1->left != e2->left) // Same point, but one is a left endpoint and the other a right endpoint. The right endpoint is processed first
        return e1->left;
    // Same point, both events are left endpoints or both are right endpoints.
    if (signedArea (e1->point, e1->otherEvent->point, e2->otherEvent->point) != 0) // not collinear
        return e1->above (e2->otherEvent->point); // the event associate to the bottom segment is processed first
    return e1->pol > e2->pol;
}
};

Maybe it’s just me, but I had a hell of a time just figuring out what problem this was even trying to solve. I still have to reread it several times whenever I look at it, to make sure I’m getting the right things backwards.

Making this even more ridiculous is that there’s a second implementation of this same sort, with the same name, in another file — and that one’s implemented forwards. And doesn’t use a tiebreaker. I don’t entirely understand how this even compiles, but it does!

I painstakingly translated this forwards to Rust. Unlike the STL, Rust doesn’t take custom comparators for its containers, so I had to implement ordering on the types themselves (which makes sense, anyway). I wrapped everything in the priority queue in a Reverse, which does what it sounds like.

I’m fairly pleased with Rust’s ordering model. Most of the work is done in Ord, a trait with a cmp() method returning an Ordering (one of Less, Equal, and Greater). No magic numbers, no need to implement all six ordering methods! It’s incredible. Ordering even has some handy methods on it, so the usual case of “order by this, then by this” can be written as:

1
2
return self.point().x.cmp(&other.point().x)
    .then(self.point().y.cmp(&other.point().y));

Well. Just kidding! It’s not quite that easy. You see, the points here are composed of floats, and floats have the fun property that not all of them are comparable. Specifically, NaN is not less than, greater than, or equal to anything else, including itself. So IEEE 754 float ordering cannot be expressed with Ord. Unless you want to just make up an answer for NaN, but Rust doesn’t tend to do that.

Rust’s float types thus implement the weaker PartialOrd, whose method returns an Option<Ordering> instead. That makes the above example slightly uglier:

1
2
return self.point().x.partial_cmp(&other.point().x).unwrap()
    .then(self.point().y.partial_cmp(&other.point().y).unwrap())

Also, since I use unwrap() here, this code will panic and take the whole program down if the points are infinite or NaN. Don’t do that.

This caused some minor inconveniences in other places; for example, the general-purpose cmp::min() doesn’t work on floats, because it requires an Ord-erable type. Thankfully there’s a f64::min(), which handles a NaN by returning the other argument.

(Cool story: for the longest time I had this code using f32s. I’m used to translating int to “32 bits”, and apparently that instinct kicked in for floats as well, even floats spelled double.)

The only other sorting adventure was this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
// Due to overlapping edges the resultEvents array can be not wholly sorted
bool sorted = false;
while (!sorted) {
    sorted = true;
    for (unsigned int i = 0; i < resultEvents.size (); ++i) {
        if (i + 1 < resultEvents.size () && sec (resultEvents[i], resultEvents[i+1])) {
            std::swap (resultEvents[i], resultEvents[i+1]);
            sorted = false;
        }
    }
}

(I originally misread this comment as saying “the array cannot be wholly sorted” and had no idea why that would be the case, or why the author would then immediately attempt to bubble sort it.)

I’m still not sure why this uses an ad-hoc sort instead of std::sort. But I’m used to taking for granted that general-purpose sorting implementations are tuned to work well for almost-sorted data, like Python’s. Maybe C++ is untrustworthy here, for some reason. I replaced it with a call to .sort() and all seemed fine.

Phew! We’re getting there. Finally, my code appears to type-check.

But now I see storm clouds gathering on the horizon.

Ownership hell

I have a problem. I somehow run into this problem every single time I use Rust. The solutions are never especially satisfying, and all the hacks I might use if forced to write C++ turn out to be unsound, which is even more annoying because rustc is just sitting there with this smug “I told you so expression” and—

The problem is ownership, which Rust is fundamentally built on. Any given value must have exactly one owner, and Rust must be able to statically convince itself that:

  1. No reference to a value outlives that value.
  2. If a mutable reference to a value exists, no other references to that value exist at the same time.

This is the core of Rust. It guarantees at compile time that you cannot lose pointers to allocated memory, you cannot double-free, you cannot have dangling pointers.

It also completely thwarts a lot of approaches you might be inclined to take if you come from managed languages (where who cares, the GC will take care of it) or C++ (where you just throw pointers everywhere and hope for the best apparently).

For example, pointer loops are impossible. Rust’s understanding of ownership and lifetimes is hierarchical, and it simply cannot express loops. (Rust’s own doubly-linked list type uses raw pointers and unsafe code under the hood, where “unsafe” is an escape hatch for the usual ownership rules. Since I only recently realized that pointers to the inside of a mutable Vec are a bad idea, I figure I should probably not be writing unsafe code myself.)

This throws a few wrenches in the works.

Problem the first: pointer loops

I immediately ran into trouble with the SweepEvent struct itself. A SweepEvent pulls double duty: it represents one endpoint of a segment, but each left endpoint also handles bookkeeping for the segment itself — which means that most of the fields on a right endpoint are unused. Also, and more importantly, each SweepEvent has a pointer to the corresponding SweepEvent at the other end of the same segment. So a pair of SweepEvents point to each other.

Rust frowns upon this. In retrospect, I think I could’ve kept it working, but I also think I’m wrong about that.

My first step was to wrench SweepEvent apart. I moved all of the segment-stuff (which is virtually all of it) into a single SweepSegment type, and then populated the event queue with a SweepEndpoint tuple struct, similar to:

1
2
3
4
5
6
enum SegmentEnd {
    Left,
    Right,
}

struct SweepEndpoint<'a>(&'a SweepSegment, SegmentEnd);

This makes SweepEndpoint essentially a tuple with a name. The 'a is a lifetime and says, more or less, that a SweepEndpoint cannot outlive the SweepSegment it references. Makes sense.

Problem solved! I no longer have mutually referential pointers. But I do still have pointers (well, references), and they have to point to something.

Problem the second: where’s all the data

Which brings me to the problem I always run into with Rust. I have a bucket of things, and I need to refer to some of them multiple times.

I tried half a dozen different approaches here and don’t clearly remember all of them, but I think my core problem went as follows. I translated the C++ class to a Rust struct with some methods hanging off of it. A simplified version might look like this.

1
2
3
4
struct Algorithm {
    arena: LinkedList<SweepSegment>,
    event_queue: BinaryHeap<SweepEndpoint>,
}

Ah, hang on — SweepEndpoint needs to be annotated with a lifetime, so Rust can enforce that those endpoints don’t live longer than the segments they refer to. No problem?

1
2
3
4
struct Algorithm<'a> {
    arena: LinkedList<SweepSegment>,
    event_queue: BinaryHeap<SweepEndpoint<'a>>,
}

Okay! Now for some methods.

1
2
3
4
5
6
7
8
fn run(&mut self) {
    self.arena.push_back(SweepSegment{ data: 5 });
    self.event_queue.push(SweepEndpoint(self.arena.back().unwrap(), SegmentEnd::Left));
    self.event_queue.push(SweepEndpoint(self.arena.back().unwrap(), SegmentEnd::Right));
    for event in &self.event_queue {
        println!("{:?}", event)
    }
}

Aaand… this doesn’t work. Rust “cannot infer an appropriate lifetime for autoref due to conflicting requirements”. The trouble is that self.arena.back() takes a reference to self.arena, and then I put that reference in the event queue. But I promised that everything in the event queue has lifetime 'a, and I don’t actually know how long self lives here; I only know that it can’t outlive 'a, because that would invalidate the references it holds.

A little random guessing let me to change &mut self to &'a mut self — which is fine because the entire impl block this lives in is already parameterized by 'a — and that makes this compile! Hooray! I think that’s because I’m saying self itself has exactly the same lifetime as the references it holds onto, which is true, since it’s referring to itself.

Let’s get a little more ambitious and try having two segments.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
fn run(&'a mut self) {
    self.arena.push_back(SweepSegment{ data: 5 });
    self.event_queue.push(SweepEndpoint(self.arena.back().unwrap(), SegmentEnd::Left));
    self.event_queue.push(SweepEndpoint(self.arena.back().unwrap(), SegmentEnd::Right));
    self.arena.push_back(SweepSegment{ data: 17 });
    self.event_queue.push(SweepEndpoint(self.arena.back().unwrap(), SegmentEnd::Left));
    self.event_queue.push(SweepEndpoint(self.arena.back().unwrap(), SegmentEnd::Right));
    for event in &self.event_queue {
        println!("{:?}", event)
    }
}

Whoops! Rust complains that I’m trying to mutate self.arena while other stuff is referring to it. And, yes, that’s true — I have references to it in the event queue, and Rust is preventing me from potentially deleting everything from the queue when references to it still exist. I’m not actually deleting anything here, of course (though I could be if this were a Vec!), but Rust’s type system can’t encode that (and I dread the thought of a type system that can).

I struggled with this for a while, and rapidly encountered another complete showstopper:

1
2
3
4
5
6
fn run(&'a mut self) {
    self.mutate_something();
    self.mutate_something();
}

fn mutate_something(&'a mut self) {}

Rust objects that I’m trying to borrow self mutably, twice — once for the first call, once for the second.

But why? A borrow is supposed to end automatically once it’s no longer used, right? Maybe if I throw some braces around it for scope… nope, that doesn’t help either.

It’s true that borrows usually end automatically, but here I have explicitly told Rust that mutate_something() should borrow with the lifetime 'a, which is the same as the lifetime in run(). So the first call explicitly borrows self for at least the rest of the method. Removing the lifetime from mutate_something() does fix this error, but if that method tries to add new segments, I’m back to the original problem.

Oh no. The mutation in the C++ code is several calls deep. Porting it directly seems nearly impossible.

The typical solution here — at least, the first thing people suggest to me on Twitter — is to wrap basically everything everywhere in Rc<RefCell<T>>, which gives you something that’s reference-counted (avoiding questions of ownership) and defers borrow checks until runtime (avoiding questions of mutable borrows). But that seems pretty heavy-handed here — not only does RefCell add .borrow() noise anywhere you actually want to interact with the underlying value, but do I really need to refcount these tiny structs that only hold a handful of floats each?

I set out to find a middle ground.

Solution, kind of

I really, really didn’t want to perform serious surgery on this code just to get it to build. I still didn’t know if it worked at all, and now I had to rearrange it without being able to check if I was breaking it further. (This isn’t Rust’s fault; it’s a natural problem with porting between fairly different paradigms.)

So I kind of hacked it into working with minimal changes, producing a grotesque abomination which I’m ashamed to link to. Here’s how!

First, I got rid of the class. It turns out this makes lifetime juggling much easier right off the bat. I’m pretty sure Rust considers everything in a struct to be destroyed simultaneously (though in practice it guarantees it’ll destroy fields in order), which doesn’t leave much wiggle room. Locals within a function, on the other hand, can each have their own distinct lifetimes, which solves the problem of expressing that the borrows won’t outlive the arena.

Speaking of the arena, I solved the mutability problem there by switching to… an arena! The typed-arena crate (a port of a type used within Rust itself, I think) is an allocator — you give it a value, and it gives you back a reference, and the reference is guaranteed to be valid for as long as the arena exists. The method that does this is sneaky and takes &self rather than &mut self, so Rust doesn’t know you’re mutating the arena and won’t complain. (One drawback is that the arena will never free anything you give to it, but that’s not a big problem here.)


My next problem was with mutation. The main loop repeatedly calls possibleIntersection with pairs of segments, which can split either or both segment. Rust definitely doesn’t like that — I’d have to pass in two &muts, both of which are mutable references into the same arena, and I’d have a bunch of immutable references into that arena in the sweep list and elsewhere. This isn’t going to fly.

This is kind of a shame, and is one place where Rust seems a little overzealous. Something like this seems like it ought to be perfectly valid:

1
2
3
4
let mut v = vec![1u32, 2u32];
let a = &mut v[0];
let b = &mut v[1];
// do stuff with a, b

The trouble is, Rust only knows the type signature, which here is something like index_mut(&'a mut self, index: usize) -> &'a T. Nothing about that says that you’re borrowing distinct elements rather than some core part of the type — and, in fact, the above code is only safe because you’re borrowing distinct elements. In the general case, Rust can’t possibly know that. It seems obvious enough from the different indexes, but nothing about the type system even says that different indexes have to return different values. And what if one were borrowed as &mut v[1] and the other were borrowed with v.iter_mut().next().unwrap()?

Anyway, this is exactly where people start to turn to RefCell — if you’re very sure you know better than Rust, then a RefCell will skirt the borrow checker while still enforcing at runtime that you don’t have more than one mutable borrow at a time.

But half the lines in this algorithm examine the endpoints of a segment! I don’t want to wrap the whole thing in a RefCell, or I’ll have to say this everywhere:

1
if segment1.borrow().point.x < segment2.borrow().point.x { ... }

Gross.

But wait — this code only mutates the points themselves in one place. When a segment is split, the original segment becomes the left half, and a new segment is created to be the right half. There’s no compelling need for this; it saves an allocation for the left half, but it’s not critical to the algorithm.

Thus, I settled on a compromise. My segment type now looks like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
struct SegmentPacket {
    // a bunch of flags and whatnot used in the algorithm
}
struct SweepSegment {
    left_point: MapPoint,
    right_point: MapPoint,
    faces_outwards: bool,
    index: usize,
    order: usize,
    packet: RefCell<SegmentPacket>,
}

I do still need to call .borrow() or .borrow_mut() to get at the stuff in the “packet”, but that’s far less common, so there’s less noise overall. And I don’t need to wrap it in Rc because it’s part of a type that’s allocated in the arena and passed around only via references.


This still leaves me with the problem of how to actually perform the splits.

I’m not especially happy with what I came up with, I don’t know if I can defend it, and I suspect I could do much better. I changed possibleIntersection so that rather than performing splits, it returns the points at which each segment needs splitting, in the form (usize, Option<MapPoint>, Option<MapPoint>). (The usize is used as a flag for calling code and oughta be an enum, but, isn’t yet.)

Now the top-level function is responsible for all arena management, and all is well.

Except, er. possibleIntersection is called multiple times, and I don’t want to copy-paste a dozen lines of split code after each call. I tried putting just that code in its own function, which had the world’s most godawful signature, and that didn’t work because… uh… hm. I can’t remember why, exactly! Should’ve written that down.

I tried a local closure next, but closures capture their environment by reference, so now I had references to a bunch of locals for as long as the closure existed, which meant I couldn’t mutate those locals. Argh. (This seems a little silly to me, since the closure’s references cannot possibly be used for anything if the closure isn’t being called, but maybe I’m missing something. Or maybe this is just a limitation of lifetimes.)

Increasingly desperate, I tried using a macro. But… macros are hygienic, which means that any new name you use inside a macro is different from any name outside that macro. The macro thus could not see any of my locals. Usually that’s good, but here I explicitly wanted the macro to mess with my locals.

I was just about to give up and go live as a hermit in a cabin in the woods, when I discovered something quite incredible. You can define local macros! If you define a macro inside a function, then it can see any locals defined earlier in that function. Perfect!

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
macro_rules! _split_segment (
    ($seg:expr, $pt:expr) => (
        {
            let pt = $pt;
            let seg = $seg;
            // ... waaay too much code ...
        }
    );
);

loop {
    // ...
    // This is possibleIntersection, renamed because Rust rightfully complains about camelCase
    let cross = handle_intersections(Some(segment), maybe_above);
    if let Some(pt) = cross.1 {
        segment = _split_segment!(segment, pt);
    }
    if let Some(pt) = cross.2 {
        maybe_above = Some(_split_segment!(maybe_above.unwrap(), pt));
    }
    // ...
}

(This doesn’t actually quite match the original algorithm, which has one case where a segment can be split twice. I realized that I could just do the left-most split, and a later iteration would perform the other split. I sure hope that’s right, anyway.)

It’s a bit ugly, and I ran into a whole lot of implicit behavior from the C++ code that I had to fix — for example, the segment is sometimes mutated just before it’s split, purely as a shortcut for mutating the left part of the split. But it finally compiles! And runs! And kinda worked, a bit!

Aftermath

I still had a lot of work to do.

For one, this code was designed for intersecting two shapes, not mass-intersecting a big pile of shapes. The basic algorithm doesn’t care about how many polygons you start with — all it sees is segments — but the code for constructing the return value needed some heavy modification.

The biggest change by far? The original code traced each segment once, expecting the result to be only a single shape. I had to change that to trace each side of each segment once, since the vast bulk of the output consists of shapes which share a side. This violated a few assumptions, which I had to hack around.

I also ran into a couple very bad edge cases, spent ages debugging them, then found out that the original algorithm had a subtle workaround that I’d commented out because it was awkward to port but didn’t seem to do anything. Whoops!

The worst was a precision error, where a vertical line could be split on a point not quite actually on the line, which wreaked all kinds of havoc. I worked around that with some tasteful rounding, which is highly dubious but makes the output more appealing to my squishy human brain. (I might switch to the original workaround, but I really dislike that even simple cases can spit out points at 1500.0000000000003. The whole thing is parameterized over the coordinate type, so maybe I could throw a rational type in there and cross my fingers?)

All that done, I finally, finally, after a couple months of intermittent progress, got what I wanted!

This is Doom 2’s MAP01. The black area to the left of center is where the player starts. Gray areas indicate where the player can walk from there, with lighter shades indicating more distant areas, where “distance” is measured by the minimum number of line crossings. Red areas can’t be reached at all.

(Note: large playable chunks of the map, including the exit room, are red. That’s because those areas are behind doors, and this code doesn’t understand doors yet.)

(Also note: The big crescent in the lower-right is also black because I was lazy and looked for the player’s starting sector by checking the bbox, and that sector’s bbox happens to match.)

The code that generated this had to go out of its way to delete all the unreachable zones around solid walls. I think I could modify the algorithm to do that on the fly pretty easily, which would probably speed it up a bit too. Downside is that the algorithm would then be pretty specifically tied to this problem, and not usable for any other kind of polygon intersection, which I would think could come up elsewhere? The modifications would be pretty minor, though, so maybe I could confine them to a closure or something.

Some final observations

It runs surprisingly slowly. Like, multiple seconds. Unless I add --release, which speeds it up by a factor of… some number with multiple digits. Wahoo. Debug mode has a high price, especially with a lot of calls in play.

The current state of this code is on GitHub. Please don’t look at it. I’m very sorry.

Honestly, most of my anguish came not from Rust, but from the original code relying on lots of fairly subtle behavior without bothering to explain what it was doing or even hint that anything unusual was going on. God, I hate C++.

I don’t know if the Rust community can learn from this. I don’t know if I even learned from this. Let’s all just quietly forget about it.

Now I just need to figure this one out…

[$] Energy-aware scheduling on asymmetric systems

Post Syndicated from corbet original https://lwn.net/Articles/749900/rss

Energy-aware scheduling — running a system’s workload in a way that
minimizes the amount of energy consumed — has been a topic of active
discussion and development for some time; LWN first covered the issue at the beginning of 2012.
Many approaches have been tried during the intervening years, but little in
the way of generalized energy-aware scheduling work has made it into the
mainline. Recently, a new patch set was
posted by Dietmar Eggemann that
only tries to address one aspect of the problem; perhaps the problem domain
has now been simplified enough that this support can finally be merged.

What John Oliver gets wrong about Bitcoin

Post Syndicated from Robert Graham original http://blog.erratasec.com/2018/03/what-john-oliver-gets-wrong-about.html

John Oliver covered bitcoin/cryptocurrencies last night. I thought I’d describe a bunch of things he gets wrong.

How Bitcoin works

Nowhere in the show does it describe what Bitcoin is and how it works.
Discussions should always start with Satoshi Nakamoto’s original paper. The thing Satoshi points out is that there is an important cost to normal transactions, namely, the entire legal system designed to protect you against fraud, such as the way you can reverse the transactions on your credit card if it gets stolen. The point of Bitcoin is that there is no way to reverse a charge. A transaction is done via cryptography: to transfer money to me, you decrypt it with your secret key and encrypt it with mine, handing ownership over to me with no third party involved that can reverse the transaction, and essentially no overhead.
All the rest of the stuff, like the decentralized blockchain and mining, is all about making that work.
Bitcoin crazies forget about the original genesis of Bitcoin. For example, they talk about adding features to stop fraud, reversing transactions, and having a central authority that manages that. This misses the point, because the existing electronic banking system already does that, and does a better job at it than cryptocurrencies ever can. If you want to mock cryptocurrencies, talk about the “DAO”, which did exactly that — and collapsed in a big fraudulent scheme where insiders made money and outsiders didn’t.
Sticking to Satoshi’s original ideas are a lot better than trying to repeat how the crazy fringe activists define Bitcoin.

How does any money have value?

Oliver’s answer is currencies have value because people agree that they have value, like how they agree a Beanie Baby is worth $15,000.
This is wrong. A better way of asking the question why the value of money changes. The dollar has been losing roughly 2% of its value each year for decades. This is called “inflation”, as the dollar loses value, it takes more dollars to buy things, which means the price of things (in dollars) goes up, and employers have to pay us more dollars so that we can buy the same amount of things.
The reason the value of the dollar changes is largely because the Federal Reserve manages the supply of dollars, using the same law of Supply and Demand. As you know, if a supply decreases (like oil), then the price goes up, or if the supply of something increases, the price goes down. The Fed manages money the same way: when prices rise (the dollar is worth less), the Fed reduces the supply of dollars, causing it to be worth more. Conversely, if prices fall (or don’t rise fast enough), the Fed increases supply, so that the dollar is worth less.
The reason money follows the law of Supply and Demand is because people use money, they consume it like they do other goods and services, like gasoline, tax preparation, food, dance lessons, and so forth. It’s not like a fine art painting, a stamp collection or a Beanie Baby — money is a product. It’s just that people have a hard time thinking of it as a consumer product since, in their experience, money is what they use to buy consumer products. But it’s a symmetric operation: when you buy gasoline with dollars, you are actually selling dollars in exchange for gasoline. That you call one side in this transaction “money” and the other “goods” is purely arbitrary, you call gasoline money and dollars the good that is being bought and sold for gasoline.
The reason dollars is a product is because trying to use gasoline as money is a pain in the neck. Storing it and exchanging it is difficult. Goods like this do become money, such as famously how prisons often use cigarettes as a medium of exchange, even for non-smokers, but it has to be a good that is fungible, storable, and easily exchanged. Dollars are the most fungible, the most storable, and the easiest exchanged, so has the most value as “money”. Sure, the mechanic can fix the farmers car for three chickens instead, but most of the time, both parties in the transaction would rather exchange the same value using dollars than chickens.
So the value of dollars is not like the value of Beanie Babies, which people might buy for $15,000, which changes purely on the whims of investors. Instead, a dollar is like gasoline, which obey the law of Supply and Demand.
This brings us back to the question of where Bitcoin gets its value. While Bitcoin is indeed used like dollars to buy things, that’s only a tiny use of the currency, so therefore it’s value isn’t determined by Supply and Demand. Instead, the value of Bitcoin is a lot like Beanie Babies, obeying the laws of investments. So in this respect, Oliver is right about where the value of Bitcoin comes, but wrong about where the value of dollars comes from.

Why Bitcoin conference didn’t take Bitcoin

John Oliver points out the irony of a Bitcoin conference that stopped accepting payments in Bitcoin for tickets.
The biggest reason for this is because Bitcoin has become so popular that transaction fees have gone up. Instead of being proof of failure, it’s proof of popularity. What John Oliver is saying is the old joke that nobody goes to that popular restaurant anymore because it’s too crowded and you can’t get a reservation.
Moreover, the point of Bitcoin is not to replace everyday currencies for everyday transactions. If you read Satoshi Nakamoto’s whitepaper, it’s only goal is to replace certain types of transactions, like purely electronic transactions where electronic goods and services are being exchanged. Where real-life goods/services are being exchanged, existing currencies work just fine. It’s only the crazy activists who claim Bitcoin will eventually replace real world currencies — the saner people see it co-existing with real-world currencies, each with a different value to consumers.

Turning a McNugget back into a chicken

John Oliver uses the metaphor of turning a that while you can process a chicken into McNuggets, you can’t reverse the process. It’s a funny metaphor.
But it’s not clear what the heck this metaphor is trying explain. That’s not a metaphor for the blockchain, but a metaphor for a “cryptographic hash”, where each block is a chicken, and the McNugget is the signature for the block (well, the block plus the signature of the last block, forming a chain).
Even then that metaphor as problems. The McNugget produced from each chicken must be unique to that chicken, for the metaphor to accurately describe a cryptographic hash. You can therefore identify the original chicken simply by looking at the McNugget. A slight change in the original chicken, like losing a feather, results in a completely different McNugget. Thus, nuggets can be used to tell if the original chicken has changed.
This then leads to the key property of the blockchain, it is unalterable. You can’t go back and change any of the blocks of data, because the fingerprints, the nuggets, will also change, and break the nugget chain.
The point is that while John Oliver is laughing at a silly metaphor to explain the blockchain becuase he totally misses the point of the metaphor.
Oliver rightly says “don’t worry if you don’t understand it — most people don’t”, but that includes the big companies that John Oliver name. Some companies do get it, and are producing reasonable things (like JP Morgan, by all accounts), but some don’t. IBM and other big consultancies are charging companies millions of dollars to consult with them on block chain products where nobody involved, the customer or the consultancy, actually understand any of it. That doesn’t stop them from happily charging customers on one side and happily spending money on the other.
Thus, rather than Oliver explaining the problem, he’s just being part of the problem. His explanation of blockchain left you dumber than before.

ICO’s

John Oliver mocks the Brave ICO ($35 million in 30 seconds), claiming it’s all driven by YouTube personalities and people who aren’t looking at the fundamentals.
And while this is true, most ICOs are bunk, the  Brave ICO actually had a business model behind it. Brave is a Chrome-like web-browser whose distinguishing feature is that it protects your privacy from advertisers. If you don’t use Brave or a browser with an ad block extension, you have no idea how bad things are for you. However, this presents a problem for websites that fund themselves via advertisements, which is most of them, because visitors no longer see ads. Brave has a fix for this. Most people wouldn’t mind supporting the websites they visit often, like the New York Times. That’s where the Brave ICO “token” comes in: it’s not simply stock in Brave, but a token for micropayments to websites. Users buy tokens, then use them for micropayments to websites like New York Times. The New York Times then sells the tokens back to the market for dollars. The buying and selling of tokens happens without a centralized middleman.
This is still all speculative, of course, and it remains to be seen how successful Brave will be, but it’s a serious effort. It has well respected VC behind the company, a well-respected founder (despite the fact he invented JavaScript), and well-respected employees. It’s not a scam, it’s a legitimate venture.

How to you make money from Bitcoin?

The last part of the show is dedicated to describing all the scam out there, advising people to be careful, and to be “responsible”. This is garbage.
It’s like my simple two step process to making lots of money via Bitcoin: (1) buy when the price is low, and (2) sell when the price is high. My advice is correct, of course, but useless. Same as “be careful” and “invest responsibly”.
The truth about investing in cryptocurrencies is “don’t”. The only responsible way to invest is to buy low-overhead market index funds and hold for retirement. No, you won’t get super rich doing this, but anything other than this is irresponsible gambling.
It’s a hard lesson to learn, because everyone is telling you the opposite. The entire channel CNBC is devoted to day traders, who buy and sell stocks at a high rate based on the same principle as a ponzi scheme, basing their judgment not on the fundamentals (like long term dividends) but animal spirits of whatever stock is hot or cold at the moment. This is the same reason people buy or sell Bitcoin, not because they can describe the fundamental value, but because they believe in a bigger fool down the road who will buy it for even more.
For things like Bitcoin, the trick to making money is to have bought it over 7 years ago when it was essentially worthless, except to nerds who were into that sort of thing. It’s the same tick to making a lot of money in Magic: The Gathering trading cards, which nerds bought decades ago which are worth a ton of money now. Or, to have bought Apple stock back in 2009 when the iPhone was new, when nerds could understand the potential of real Internet access and apps that Wall Street could not.
That was my strategy: be a nerd, who gets into things. I’ve made a good amount of money on all these things because as a nerd, I was into Magic: The Gathering, Bitcoin, and the iPhone before anybody else was, and bought in at the point where these things were essentially valueless.
At this point with cryptocurrencies, with the non-nerds now flooding the market, there little chance of making it rich. The lottery is probably a better bet. Instead, if you want to make money, become a nerd, obsess about a thing, understand a thing when its new, and cash out once the rest of the market figures it out. That might be Brave, for example, but buy into it because you’ve spent the last year studying the browser advertisement ecosystem, the market’s willingness to pay for content, and how their Basic Attention Token delivers value to websites — not because you want in on the ICO craze.

Conclusion

John Oliver spends 25 minutes explaining Bitcoin, Cryptocurrencies, and the Blockchain to you. Sure, it’s funny, but it leaves you worse off than when it started. It admits they “simplify” the explanation, but they simplified it so much to the point where they removed all useful information.

Improve the Operational Efficiency of Amazon Elasticsearch Service Domains with Automated Alarms Using Amazon CloudWatch

Post Syndicated from Veronika Megler original https://aws.amazon.com/blogs/big-data/improve-the-operational-efficiency-of-amazon-elasticsearch-service-domains-with-automated-alarms-using-amazon-cloudwatch/

A customer has been successfully creating and running multiple Amazon Elasticsearch Service (Amazon ES) domains to support their business users’ search needs across products, orders, support documentation, and a growing suite of similar needs. The service has become heavily used across the organization.  This led to some domains running at 100% capacity during peak times, while others began to run low on storage space. Because of this increased usage, the technical teams were in danger of missing their service level agreements.  They contacted me for help.

This post shows how you can set up automated alarms to warn when domains need attention.

Solution overview

Amazon ES is a fully managed service that delivers Elasticsearch’s easy-to-use APIs and real-time analytics capabilities along with the availability, scalability, and security that production workloads require.  The service offers built-in integrations with a number of other components and AWS services, enabling customers to go from raw data to actionable insights quickly and securely.

One of these other integrated services is Amazon CloudWatch. CloudWatch is a monitoring service for AWS Cloud resources and the applications that you run on AWS. You can use CloudWatch to collect and track metrics, collect and monitor log files, set alarms, and automatically react to changes in your AWS resources.

CloudWatch collects metrics for Amazon ES. You can use these metrics to monitor the state of your Amazon ES domains, and set alarms to notify you about high utilization of system resources.  For more information, see Amazon Elasticsearch Service Metrics and Dimensions.

While the metrics are automatically collected, the missing piece is how to set alarms on these metrics at appropriate levels for each of your domains. This post includes sample Python code to evaluate the current state of your Amazon ES environment, and to set up alarms according to AWS recommendations and best practices.

There are two components to the sample solution:

  • es-check-cwalarms.py: This Python script checks the CloudWatch alarms that have been set, for all Amazon ES domains in a given account and region.
  • es-create-cwalarms.py: This Python script sets up a set of CloudWatch alarms for a single given domain.

The sample code can also be found in the amazon-es-check-cw-alarms GitHub repo. The scripts are easy to extend or combine, as described in the section “Extensions and Adaptations”.

Assessing the current state

The first script, es-check-cwalarms.py, is used to give an overview of the configurations and alarm settings for all the Amazon ES domains in the given region. The script takes the following parameters:

python es-checkcwalarms.py -h
usage: es-checkcwalarms.py [-h] [-e ESPREFIX] [-n NOTIFY] [-f FREE][-p PROFILE] [-r REGION]
Checks a set of recommended CloudWatch alarms for Amazon Elasticsearch Service domains (optionally, those beginning with a given prefix).
optional arguments:
  -h, --help   		show this help message and exit
  -e ESPREFIX, --esprefix ESPREFIX	Only check Amazon Elasticsearch Service domains that begin with this prefix.
  -n NOTIFY, --notify NOTIFY    List of CloudWatch alarm actions; e.g. ['arn:aws:sns:xxxx']
  -f FREE, --free FREE  Minimum free storage (MB) on which to alarm
  -p PROFILE, --profile PROFILE     IAM profile name to use
  -r REGION, --region REGION       AWS region for the domain. Default: us-east-1

The script first identifies all the domains in the given region (or, optionally, limits them to the subset that begins with a given prefix). It then starts running a set of checks against each one.

The script can be run from the command line or set up as a scheduled Lambda function. For example, for one customer, it was deemed appropriate to regularly run the script to check that alarms were correctly set for all domains. In addition, because configuration changes—cluster size increases to accommodate larger workloads being a common change—might require updates to alarms, this approach allowed the automatic identification of alarms no longer appropriately set as the domain configurations changed.

The output shown below is the output for one domain in my account.

Starting checks for Elasticsearch domain iotfleet , version is 53
Iotfleet Automated snapshot hour (UTC): 0
Iotfleet Instance configuration: 1 instances; type:m3.medium.elasticsearch
Iotfleet Instance storage definition is: 4 GB; free storage calced to: 819.2 MB
iotfleet Desired free storage set to (in MB): 819.2
iotfleet WARNING: Not using VPC Endpoint
iotfleet WARNING: Does not have Zone Awareness enabled
iotfleet WARNING: Instance count is ODD. Best practice is for an even number of data nodes and zone awareness.
iotfleet WARNING: Does not have Dedicated Masters.
iotfleet WARNING: Neither index nor search slow logs are enabled.
iotfleet WARNING: EBS not in use. Using instance storage only.
iotfleet Alarm ok; definition matches. Test-Elasticsearch-iotfleet-ClusterStatus.yellow-Alarm ClusterStatus.yellow
iotfleet Alarm ok; definition matches. Test-Elasticsearch-iotfleet-ClusterStatus.red-Alarm ClusterStatus.red
iotfleet Alarm ok; definition matches. Test-Elasticsearch-iotfleet-CPUUtilization-Alarm CPUUtilization
iotfleet Alarm ok; definition matches. Test-Elasticsearch-iotfleet-JVMMemoryPressure-Alarm JVMMemoryPressure
iotfleet WARNING: Missing alarm!! ('ClusterIndexWritesBlocked', 'Maximum', 60, 5, 'GreaterThanOrEqualToThreshold', 1.0)
iotfleet Alarm ok; definition matches. Test-Elasticsearch-iotfleet-AutomatedSnapshotFailure-Alarm AutomatedSnapshotFailure
iotfleet Alarm: Threshold does not match: Test-Elasticsearch-iotfleet-FreeStorageSpace-Alarm Should be:  819.2 ; is 3000.0

The output messages fall into the following categories:

  • System overview, Informational: The Amazon ES version and configuration, including instance type and number, storage, automated snapshot hour, etc.
  • Free storage: A calculation for the appropriate amount of free storage, based on the recommended 20% of total storage.
  • Warnings: best practices that are not being followed for this domain. (For more about this, read on.)
  • Alarms: An assessment of the CloudWatch alarms currently set for this domain, against a recommended set.

The script contains an array of recommended CloudWatch alarms, based on best practices for these metrics and statistics. Using the array allows alarm parameters (such as free space) to be updated within the code based on current domain statistics and configurations.

For a given domain, the script checks if each alarm has been set. If the alarm is set, it checks whether the values match those in the array esAlarms. In the output above, you can see three different situations being reported:

  • Alarm ok; definition matches. The alarm set for the domain matches the settings in the array.
  • Alarm: Threshold does not match. An alarm exists, but the threshold value at which the alarm is triggered does not match.
  • WARNING: Missing alarm!! The recommended alarm is missing.

All in all, the list above shows that this domain does not have a configuration that adheres to best practices, nor does it have all the recommended alarms.

Setting up alarms

Now that you know that the domains in their current state are missing critical alarms, you can correct the situation.

To demonstrate the script, set up a new domain named “ver”, in us-west-2. Specify 1 node, and a 10-GB EBS disk. Also, create an SNS topic in us-west-2 with a name of “sendnotification”, which sends you an email.

Run the second script, es-create-cwalarms.py, from the command line. This script creates (or updates) the desired CloudWatch alarms for the specified Amazon ES domain, “ver”.

python es-create-cwalarms.py -r us-west-2 -e test -c ver -n "['arn:aws:sns:us-west-2:xxxxxxxxxx:sendnotification']"
EBS enabled: True type: gp2 size (GB): 10 No Iops 10240  total storage (MB)
Desired free storage set to (in MB): 2048.0
Creating  Test-Elasticsearch-ver-ClusterStatus.yellow-Alarm
Creating  Test-Elasticsearch-ver-ClusterStatus.red-Alarm
Creating  Test-Elasticsearch-ver-CPUUtilization-Alarm
Creating  Test-Elasticsearch-ver-JVMMemoryPressure-Alarm
Creating  Test-Elasticsearch-ver-FreeStorageSpace-Alarm
Creating  Test-Elasticsearch-ver-ClusterIndexWritesBlocked-Alarm
Creating  Test-Elasticsearch-ver-AutomatedSnapshotFailure-Alarm
Successfully finished creating alarms!

As with the first script, this script contains an array of recommended CloudWatch alarms, based on best practices for these metrics and statistics. This approach allows you to add or modify alarms based on your use case (more on that below).

After running the script, navigate to Alarms on the CloudWatch console. You can see the set of alarms set up on your domain.

Because the “ver” domain has only a single node, cluster status is yellow, and that alarm is in an “ALARM” state. It’s already sent a notification that the alarm has been triggered.

What to do when an alarm triggers

After alarms are set up, you need to identify the correct action to take for each alarm, which depends on the alarm triggered. For ideas, guidance, and additional pointers to supporting documentation, see Get Started with Amazon Elasticsearch Service: Set CloudWatch Alarms on Key Metrics. For information about common errors and recovery actions to take, see Handling AWS Service Errors.

In most cases, the alarm triggers due to an increased workload. The likely action is to reconfigure the system to handle the increased workload, rather than reducing the incoming workload. Reconfiguring any backend store—a category of systems that includes Elasticsearch—is best performed when the system is quiescent or lightly loaded. Reconfigurations such as setting zone awareness or modifying the disk type cause Amazon ES to enter a “processing” state, potentially disrupting client access.

Other changes, such as increasing the number of data nodes, may cause Elasticsearch to begin moving shards, potentially impacting search performance on these shards while this is happening. These actions should be considered in the context of your production usage. For the same reason I also do not recommend running a script that resets all domains to match best practices.

Avoid the need to reconfigure during heavy workload by setting alarms at a level that allows a considered approach to making the needed changes. For example, if you identify that each weekly peak is increasing, you can reconfigure during a weekly quiet period.

While Elasticsearch can be reconfigured without being quiesced, it is not a best practice to automatically scale it up and down based on usage patterns. Unlike some other AWS services, I recommend against setting a CloudWatch action that automatically reconfigures the system when alarms are triggered.

There are other situations where the planned reconfiguration approach may not work, such as low or zero free disk space causing the domain to reject writes. If the business is dependent on the domain continuing to accept incoming writes and deleting data is not an option, the team may choose to reconfigure immediately.

Extensions and adaptations

You may wish to modify the best practices encoded in the scripts for your own environment or workloads. It’s always better to avoid situations where alerts are generated but routinely ignored. All alerts should trigger a review and one or more actions, either immediately or at a planned date. The following is a list of common situations where you may wish to set different alarms for different domains:

  • Dev/test vs. production
    You may have a different set of configuration rules and alarms for your dev environment configurations than for test. For example, you may require zone awareness and dedicated masters for your production environment, but not for your development domains. Or, you may not have any alarms set in dev. For test environments that mirror your potential peak load, test to ensure that the alarms are appropriately triggered.
  • Differing workloads or SLAs for different domains
    You may have one domain with a requirement for superfast search performance, and another domain with a heavy ingest load that tolerates slower search response. Your reaction to slow response for these two workloads is likely to be different, so perhaps the thresholds for these two domains should be set at a different level. In this case, you might add a “max CPU utilization” alarm at 100% for 1 minute for the fast search domain, while the other domain only triggers an alarm when the average has been higher than 60% for 5 minutes. You might also add a “free space” rule with a higher threshold to reflect the need for more space for the heavy ingest load if there is danger that it could fill the available disk quickly.
  • “Normal” alarms versus “emergency” alarms
    If, for example, free disk space drops to 25% of total capacity, an alarm is triggered that indicates action should be taken as soon as possible, such as cleaning up old indexes or reconfiguring at the next quiet period for this domain. However, if free space drops below a critical level (20% free space), action must be taken immediately in order to prevent Amazon ES from setting the domain to read-only. Similarly, if the “ClusterIndexWritesBlocked” alarm triggers, the domain has already stopped accepting writes, so immediate action is needed. In this case, you may wish to set “laddered” alarms, where one threshold causes an alarm to be triggered to review the current workload for a planned reconfiguration, but a different threshold raises a “DefCon 3” alarm that immediate action is required.

The sample scripts provided here are a starting point, intended for you to adapt to your own environment and needs.

Running the scripts one time can identify how far your current state is from your desired state, and create an initial set of alarms. Regularly re-running these scripts can capture changes in your environment over time and adjusting your alarms for changes in your environment and configurations. One customer has set them up to run nightly, and to automatically create and update alarms to match their preferred settings.

Removing unwanted alarms

Each CloudWatch alarm costs approximately $0.10 per month. You can remove unwanted alarms in the CloudWatch console, under Alarms. If you set up a “ver” domain above, remember to remove it to avoid continuing charges.

Conclusion

Setting CloudWatch alarms appropriately for your Amazon ES domains can help you avoid suboptimal performance and allow you to respond to workload growth or configuration issues well before they become urgent. This post gives you a starting point for doing so. The additional sleep you’ll get knowing you don’t need to be concerned about Elasticsearch domain performance will allow you to focus on building creative solutions for your business and solving problems for your customers.

Enjoy!


Additional Reading

If you found this post useful, be sure to check out Analyzing Amazon Elasticsearch Service Slow Logs Using Amazon CloudWatch Logs Streaming and Kibana and Get Started with Amazon Elasticsearch Service: How Many Shards Do I Need?

 


About the Author

Dr. Veronika Megler is a senior consultant at Amazon Web Services. She works with our customers to implement innovative big data, AI and ML projects, helping them accelerate their time-to-value when using AWS.

 

 

 

Auto Scaling is now available for Amazon SageMaker

Post Syndicated from Ana Visneski original https://aws.amazon.com/blogs/aws/auto-scaling-is-now-available-for-amazon-sagemaker/

Kumar Venkateswar, Product Manager on the AWS ML Platforms Team, shares details on the announcement of Auto Scaling with Amazon SageMaker.


With Amazon SageMaker, thousands of customers have been able to easily build, train and deploy their machine learning (ML) models. Today, we’re making it even easier to manage production ML models, with Auto Scaling for Amazon SageMaker. Instead of having to manually manage the number of instances to match the scale that you need for your inferences, you can now have SageMaker automatically scale the number of instances based on an AWS Auto Scaling Policy.

SageMaker has made managing the ML process easier for many customers. We’ve seen customers take advantage of managed Jupyter notebooks and managed distributed training. We’ve seen customers deploying their models to SageMaker hosting for inferences, as they integrate machine learning with their applications. SageMaker makes this easy –  you don’t have to think about patching the operating system (OS) or frameworks on your inference hosts, and you don’t have to configure inference hosts across Availability Zones. You just deploy your models to SageMaker, and it handles the rest.

Until now, you have needed to specify the number and type of instances per endpoint (or production variant) to provide the scale that you need for your inferences. If your inference volume changes, you can change the number and/or type of instances that back each endpoint to accommodate that change, without incurring any downtime. In addition to making it easy to change provisioning, customers have asked us how we can make managing capacity for SageMaker even easier.

With Auto Scaling for Amazon SageMaker, in the SageMaker console, the AWS Auto Scaling API, and the AWS SDK, this becomes much easier. Now, instead of having to closely monitor inference volume, and change the endpoint configuration in response, customers can configure a scaling policy to be used by AWS Auto Scaling. Auto Scaling adjusts the number of instances up or down in response to actual workloads, determined by using Amazon CloudWatch metrics and target values defined in the policy. In this way, customers can automatically adjust their inference capacity to maintain predictable performance at a low cost. You simply specify the target inference throughput per instance and provide upper and lower bounds for the number of instances for each production variant. SageMaker will then monitor throughput per instance using Amazon CloudWatch alarms, and then it will adjust provisioned capacity up or down as needed.

After you configure the endpoint with Auto Scaling, SageMaker will continue to monitor your deployed models to automatically adjust the instance count. SageMaker will keep throughput within desired levels, in response to changes in application traffic. This makes it easier to manage models in production, and it can help reduce the cost of deployed models, as you no longer have to provision sufficient capacity in order to manage your peak load. Instead, you configure the limits to accommodate your minimum expected traffic and the maximum peak, and Amazon SageMaker will work within those limits to minimize cost.

How do you get started? Open the SageMaker console. For existing endpoints, you first access the endpoint to modify the settings.


Then, scroll to the Endpoint runtime settings section, select the variant, and choose Configure auto scaling.


First, configure the minimum and maximum number of instances.

Next, choose the throughput per instance at which you want to add an additional instance, given previous load testing.

You can optionally set cool down periods for scaling in or out, to avoid oscillation during periods of wide fluctuation in workload. If not, SageMaker will assume default values.

And that’s it! You now have an endpoint that will automatically scale with increasing inferences.

You pay for the capacity used at regular SageMaker pay-as-you-go pricing, so you no longer have to pay for unused capacity during relative idle periods!

Auto Scaling in Amazon SageMaker is available today in the US East (N. Virginia & Ohio), EU (Ireland), and U.S. West (Oregon) AWS regions. To learn more, see the Amazon SageMaker Auto Scaling documentation.


Kumar Venkateswar is a Product Manager in the AWS ML Platforms team, which includes Amazon SageMaker, Amazon Machine Learning, and the AWS Deep Learning AMIs. When not working, Kumar plays the violin and Magic: The Gathering.