Tag Archives: eset

From Framework to Function: Deploying AWS Lambda Functions for Java 8 using Apache Maven Archetype

Post Syndicated from Ryosuke Iwanaga original https://aws.amazon.com/blogs/compute/from-framework-to-function-deploying-aws-lambda-functions-for-java-8-using-apache-maven-archetype/

As a serverless computing platform that supports Java 8 runtime, AWS Lambda makes it easy to run any type of Java function simply by uploading a JAR file. To help define not only a Lambda serverless application but also Amazon API Gateway, Amazon DynamoDB, and other related services, the AWS Serverless Application Model (SAM) allows developers to use a simple AWS CloudFormation template.

AWS provides the AWS Toolkit for Eclipse that supports both Lambda and SAM. AWS also gives customers an easy way to create Lambda functions and SAM applications in Java using the AWS Command Line Interface (AWS CLI). After you build a JAR file, all you have to do is type the following commands:

aws cloudformation package 
aws cloudformation deploy

To consolidate these steps, customers can use Archetype by Apache Maven. Archetype uses a predefined package template that makes getting started to develop a function exceptionally simple.

In this post, I introduce a Maven archetype that allows you to create a skeleton of AWS SAM for a Java function. Using this archetype, you can generate a sample Java code example and an accompanying SAM template to deploy it on AWS Lambda by a single Maven action.

Prerequisites

Make sure that the following software is installed on your workstation:

  • Java
  • Maven
  • AWS CLI
  • (Optional) AWS SAM CLI

Install Archetype

After you’ve set up those packages, install Archetype with the following commands:

git clone https://github.com/awslabs/aws-serverless-java-archetype
cd aws-serverless-java-archetype
mvn install

These are one-time operations, so you don’t run them for every new package. If you’d like, you can add Archetype to your company’s Maven repository so that other developers can use it later.

With those packages installed, you’re ready to develop your new Lambda Function.

Start a project

Now that you have the archetype, customize it and run the code:

cd /path/to/project_home
mvn archetype:generate \
  -DarchetypeGroupId=com.amazonaws.serverless.archetypes \
  -DarchetypeArtifactId=aws-serverless-java-archetype \
  -DarchetypeVersion=1.0.0 \
  -DarchetypeRepository=local \ # Forcing to use local maven repository
  -DinteractiveMode=false \ # For batch mode
  # You can also specify properties below interactively if you omit the line for batch mode
  -DgroupId=YOUR_GROUP_ID \
  -DartifactId=YOUR_ARTIFACT_ID \
  -Dversion=YOUR_VERSION \
  -DclassName=YOUR_CLASSNAME

You should have a directory called YOUR_ARTIFACT_ID that contains the files and folders shown below:

├── event.json
├── pom.xml
├── src
│   └── main
│       ├── java
│       │   └── Package
│       │       └── Example.java
│       └── resources
│           └── log4j2.xml
└── template.yaml

The sample code is a working example. If you install SAM CLI, you can invoke it just by the command below:

cd YOUR_ARTIFACT_ID
mvn -P invoke verify
[INFO] Scanning for projects...
[INFO]
[INFO] ---------------------------< com.riywo:foo >----------------------------
[INFO] Building foo 1.0
[INFO] --------------------------------[ jar ]---------------------------------
...
[INFO] --- maven-jar-plugin:3.0.2:jar (default-jar) @ foo ---
[INFO] Building jar: /private/tmp/foo/target/foo-1.0.jar
[INFO]
[INFO] --- maven-shade-plugin:3.1.0:shade (shade) @ foo ---
[INFO] Including com.amazonaws:aws-lambda-java-core:jar:1.2.0 in the shaded jar.
[INFO] Replacing /private/tmp/foo/target/lambda.jar with /private/tmp/foo/target/foo-1.0-shaded.jar
[INFO]
[INFO] --- exec-maven-plugin:1.6.0:exec (sam-local-invoke) @ foo ---
2018/04/06 16:34:35 Successfully parsed template.yaml
2018/04/06 16:34:35 Connected to Docker 1.37
2018/04/06 16:34:35 Fetching lambci/lambda:java8 image for java8 runtime...
java8: Pulling from lambci/lambda
Digest: sha256:14df0a5914d000e15753d739612a506ddb8fa89eaa28dcceff5497d9df2cf7aa
Status: Image is up to date for lambci/lambda:java8
2018/04/06 16:34:37 Invoking Package.Example::handleRequest (java8)
2018/04/06 16:34:37 Decompressing /tmp/foo/target/lambda.jar
2018/04/06 16:34:37 Mounting /private/var/folders/x5/ldp7c38545v9x5dg_zmkr5kxmpdprx/T/aws-sam-local-1523000077594231063 as /var/task:ro inside runtime container
START RequestId: a6ae19fe-b1b0-41e2-80bc-68a40d094d74 Version: $LATEST
Log output: Greeting is 'Hello Tim Wagner.'
END RequestId: a6ae19fe-b1b0-41e2-80bc-68a40d094d74
REPORT RequestId: a6ae19fe-b1b0-41e2-80bc-68a40d094d74	Duration: 96.60 ms	Billed Duration: 100 ms	Memory Size: 128 MB	Max Memory Used: 7 MB

{"greetings":"Hello Tim Wagner."}


[INFO] ------------------------------------------------------------------------
[INFO] BUILD SUCCESS
[INFO] ------------------------------------------------------------------------
[INFO] Total time: 10.452 s
[INFO] Finished at: 2018-04-06T16:34:40+09:00
[INFO] ------------------------------------------------------------------------

This maven goal invokes sam local invoke -e event.json, so you can see the sample output to greet Tim Wagner.

To deploy this application to AWS, you need an Amazon S3 bucket to upload your package. You can use the following command to create a bucket if you want:

aws s3 mb s3://YOUR_BUCKET --region YOUR_REGION

Now, you can deploy your application by just one command!

mvn deploy \
    -DawsRegion=YOUR_REGION \
    -Ds3Bucket=YOUR_BUCKET \
    -DstackName=YOUR_STACK
[INFO] Scanning for projects...
[INFO]
[INFO] ---------------------------< com.riywo:foo >----------------------------
[INFO] Building foo 1.0
[INFO] --------------------------------[ jar ]---------------------------------
...
[INFO] --- exec-maven-plugin:1.6.0:exec (sam-package) @ foo ---
Uploading to aws-serverless-java/com.riywo:foo:1.0/924732f1f8e4705c87e26ef77b080b47  11657 / 11657.0  (100.00%)
Successfully packaged artifacts and wrote output template to file target/sam.yaml.
Execute the following command to deploy the packaged template
aws cloudformation deploy --template-file /private/tmp/foo/target/sam.yaml --stack-name <YOUR STACK NAME>
[INFO]
[INFO] --- maven-deploy-plugin:2.8.2:deploy (default-deploy) @ foo ---
[INFO] Skipping artifact deployment
[INFO]
[INFO] --- exec-maven-plugin:1.6.0:exec (sam-deploy) @ foo ---

Waiting for changeset to be created..
Waiting for stack create/update to complete
Successfully created/updated stack - archetype
[INFO] ------------------------------------------------------------------------
[INFO] BUILD SUCCESS
[INFO] ------------------------------------------------------------------------
[INFO] Total time: 37.176 s
[INFO] Finished at: 2018-04-06T16:41:02+09:00
[INFO] ------------------------------------------------------------------------

Maven automatically creates a shaded JAR file, uploads it to your S3 bucket, replaces template.yaml, and creates and updates the CloudFormation stack.

To customize the process, modify the pom.xml file. For example, to avoid typing values for awsRegion, s3Bucket or stackName, write them inside pom.xml and check in your VCS. Afterward, you and the rest of your team can deploy the function by typing just the following command:

mvn deploy

Options

Lambda Java 8 runtime has some types of handlers: POJO, Simple type and Stream. The default option of this archetype is POJO style, which requires to create request and response classes, but they are baked by the archetype by default. If you want to use other type of handlers, you can use handlerType property like below:

## POJO type (default)
mvn archetype:generate \
 ...
 -DhandlerType=pojo

## Simple type - String
mvn archetype:generate \
 ...
 -DhandlerType=simple

### Stream type
mvn archetype:generate \
 ...
 -DhandlerType=stream

See documentation for more details about handlers.

Also, Lambda Java 8 runtime supports two types of Logging class: Log4j 2 and LambdaLogger. This archetype creates LambdaLogger implementation by default, but you can use Log4j 2 if you want:

## LambdaLogger (default)
mvn archetype:generate \
 ...
 -Dlogger=lambda

## Log4j 2
mvn archetype:generate \
 ...
 -Dlogger=log4j2

If you use LambdaLogger, you can delete ./src/main/resources/log4j2.xml. See documentation for more details.

Conclusion

So, what’s next? Develop your Lambda function locally and type the following command: mvn deploy !

With this Archetype code example, available on GitHub repo, you should be able to deploy Lambda functions for Java 8 in a snap. If you have any questions or comments, please submit them below or leave them on GitHub.

Съвместима ли е таксата за радио и телевизия с правото на ЕС

Post Syndicated from nellyo original https://nellyo.wordpress.com/2018/05/13/fee_psm/

През март  2018 г. Oberverwaltungsgericht Rheinland-Pfalz (Върховен административен съд на Рейнланд-Пфалц – OVG Rheinland-Pfalz) решава, че таксата за радио и телевизия в Германия е съвместима с правото на ЕС (дело № 7 A 11938/17) . Съдът отхвърля тезата, че таксата е несъвместима с правото на ЕС, тъй като предоставя на обществените радио- и телевизионни доставчици на медийни услуги несправедливо предимство пред техните частни конкуренти.

Съдът посочва, че през 2016 г. Bundesverwaltungsgericht (Федерален административен съд – BVerwG) вече е установил съответствието на таксата  – в новата й форма, въведена през 2013 г. – с правото на ЕС (решение от 18 март 2016 г., BVerwG 6 С 6.15). Съгласно това решение въвеждането на таксата  не изисква съгласието на Европейската комисия и е с съвместимо с Директивата за аудиовизуалните медийни услуги.  Обществените и частните радио- и телевизионни оператори  неизбежно ще бъдат финансирани по различни начини. Това обаче не означава непременно, че обществените радио- и телевизионни оператори са получили несправедливо предимство, тъй като за разлика от частните радио- и телевизионни оператори те са подложени на много по-ограничителни правила за рекламиране и следователно са финансово зависими от таксата.

Междувременно Landgericht Tübingen (Районен съд в Тюбинген, решение от 3 август 2017 г., дело № 5 T 246/17 и др.) е постановил, че таксата  нарушава правото на ЕС  – и в резултат има подадено преюдициално запитване до Съда на ЕС –  дело  С-492/17.

 

Преюдициални въпроси:

1)

Несъвместим ли е с правото на Съюза националният Gesetz vom 18.10.2011 zur Geltung des Rundfunkbeitragsstaatsvertrags (RdFunkBeitrStVtrBW) vom 17 Dezember 2010 (Закон от 18 октомври 2011 г. за прилагане на Държавния договор за вноската за радио- и телевизионно разпространение от 17 декември 2010 г., наричан по-нататък „RdFunkBeitrStVtrBW“) на провинция Баден-Вюртемберг, последно изменен с член 4 от Neunzehnter Rundfunkänderungsstaatsvertrag (Деветнадесети държавен договор за изменение на Държавните договори за радио- и телевизионно разпространение) от 3 декември 2015 г. (Закон от 23 февруари 2016 г., GBl. стр. 126, 129), поради това че вноската, събирана от 1 януари 2013 г. съгласно този закон безусловно по принцип от всяко живеещо в германската федерална провинция Баден-Вюртемберг пълнолетно лице в полза на радио- и телевизионните оператори SWR и ZDF, представлява помощ, която противоречи на правото на Съюза и предоставя по-благоприятно третиране само в полза на тези обществени радио- и телевизионни оператори спрямо частни радио- и телевизионни оператори? Трябва ли членове 107 и 108 ДФЕС да се тълкуват в смисъл, че за Закона за вноската за радио- и телевизионно разпространение е трябвало да се получи разрешението на Комисията и поради липсата на разрешение той е невалиден?

2)

Трябва ли член 107 ДФЕС, съответно член 108 ДФЕС да се тълкува в смисъл, че в обхвата му попада правна уредба, установена в националния закон „RdFunkBeitrStVtrBW“, която предвижда, че по принцип от всяко живеещо в Баден-Вюртемберг пълнолетно лице безусловно се събира вноска в полза само на държавни/обществени радио- и телевизионни оператори, поради това че тази вноска съдържа противоречаща на правото на Съюза и предоставяща по-благоприятно третиране помощ с цел изключването по технически причини на оператори от държави от Европейския съюз, доколкото вноските са предназначени да се използват за създаването на конкурентен начин на пренос (монопол върху DVB-T2), без да е предвидено той да се използва от чуждестранни оператори? Трябва ли член 107 ДФЕС, съответно член 108 ДФЕС да се тълкува в смисъл, че в обхвата му попадат не само преки субсидии, но и други релевантни от икономическа гледна точка привилегии (право на издаване на изпълнителен лист, правомощия за предприемане на действия както в качеството на стопанско предприятие, така и в качеството на орган, поставяне в по-благоприятно положение при изчисляването на дълговете)?

3)

Съвместимо ли е с принципа на равно третиране и със забраната за предоставящи привилегии помощи положение, при което на основание национален закон на провинция Баден-Вюртемберг германски телевизионен оператор, който се урежда от нормите на публичното право и има предоставени правомощия на орган, но същевременно се конкурира с частни радио- и телевизионни оператори на рекламния пазар, е привилегирован в сравнение с тези оператори поради това че не трябва като частните конкуренти да иска по общия съдебен ред да му бъде издаден изпълнителен лист за вземанията му срещу зрителите, преди да може да пристъпи към принудително изпълнение, а самият той има право, без участието на съд, да издаде титул, който същевременно му дава право на принудително изпълнение?

4)

Съвместимо ли е с член 10 от ЕКПЧ /член [11] от Хартата на основните права (свобода на информация) положение, при което държава членка предвижда в национален закон на провинция Баден-Вюртемберг, че телевизионен оператор, на който са предоставени правомощия на орган, има право да изисква плащането на вноска от всяко живеещо в зоната на радио- и телевизионното излъчване пълнолетно лице за целите на финансирането на точно този оператор, при неплащането на която е предвидена глоба, независимо дали това лице въобще разполага с приемник или само използва услугите на други, а именно чуждестранни или други, частни оператори?

5)

Съвместим ли е националният закон „RdFunkBeitrStVtrBW“, и по-специално членове 2 и 3, с установените в правото на Съюза принципи на равно третиране и на недопускане на дискриминация в положение, при което вноската, която следва да се плаща безусловно от всеки жител за целите на финансирането на обществен телевизионен оператор, налага на всяко лице, което само отглежда детето си, тежест в размер, многократно по-висок от сумата, дължима от лице, което живее в общо жилище с други хора? Следва ли Директива 2004/113/ЕО (1) да се тълкува в смисъл, че спорната вноска също попада в обхвата ѝ и че e достатъчно да е налице косвено поставяне в по-неблагоприятно положение, след като с оглед на реалните дадености 90 % от жените понасят по-голяма тежест?

6)

Съвместим ли националният закон „RdFunkBeitrStVtrBW“, и по-специално членове 2 и 3, с установените в правото на Съюза принципи на равно третиране и на недопускане на дискриминация в положение, при което вноската, която следва да се плаща безусловно от всеки жител за целите на финансирането на обществен телевизионен оператор, за нуждаещите се от второ жилище лица по свързана с работата причина е двойно по-голяма, отколкото за други работници?

7)

Съвместим ли е националният закон „RdFunkBeitrStVtrBW“, и по-специално членове 2 и 3, с установените в правото на Съюза принципи на равно третиране и на недопускане на дискриминация и със свободата на установяване, ако вноската, която следва да се плаща безусловно от всеки жител за целите на финансирането на обществен телевизионен оператор, е уредена по такъв начин, че при еднаква възможност за приемане на радио- и телевизионно разпространение непосредствено преди границата със съседна държава от ЕС германски гражданин дължи вноската само поради мястото си на пребиваване, докато германският гражданин, живущ непосредствено от другата страна на границата, не дължи вноската, също както гражданинът на друга държава — членка на ЕС, който по свързани с работата причини трябва да се установи непосредствено от другата страна на вътрешна граница на ЕС, понася тежестта на вноската, но не и гражданинът на ЕС, живущ непосредствено преди границата, дори и никой от двамата да не се интересува от приемането на излъчванията на германския оператор?

Коментар по въпрос №4:  допуснат е въпрос за съвместимост с чл.10 от Конвенцията за правата на човека. Съдът за правата на човека вече се е произнасял, има съображения за недопустимост по сходно дело отпреди десетина години –  ето тук съм писала – вж Faccio v Italy – но нека да се произнесе и Съдът на ЕС.

И – отново за характера на таксата: ако  плащат и хората без приемник, това очевидно не е такса в смисъл цена за услуга, а данъчно вземане, по мое мнение това е тенденцията.

Чакаме решението на Съда на ЕС. Нека да се развива и множи практиката.

A serverless solution for invoking AWS Lambda at a sub-minute frequency

Post Syndicated from Emanuele Menga original https://aws.amazon.com/blogs/architecture/a-serverless-solution-for-invoking-aws-lambda-at-a-sub-minute-frequency/

If you’ve used Amazon CloudWatch Events to schedule the invocation of a Lambda function at regular intervals, you may have noticed that the highest frequency possible is one invocation per minute. However, in some cases, you may need to invoke Lambda more often than that. In this blog post, I’ll cover invoking a Lambda function every 10 seconds, but with some simple math you can change to whatever interval you like.

To achieve this, I’ll show you how to leverage Step Functions and Amazon Kinesis Data Streams.

The Solution

For this example, I’ve created a Step Functions State Machine that invokes our Lambda function 6 times, 10 seconds apart. Such State Machine is then executed once per minute by a CloudWatch Events Rule. This state machine is then executed once per minute by an Amazon CloudWatch Events rule. Finally, the Kinesis Data Stream triggers our Lambda function for each record inserted. The result is our Lambda function being invoked every 10 seconds, indefinitely.

Below is a diagram illustrating how the various services work together.

Step 1: My sampleLambda function doesn’t actually do anything, it just simulates an execution for a few seconds. This is the (Python) code of my dummy function:

import time

import random


def lambda_handler(event, context):

rand = random.randint(1, 3)

print('Running for {} seconds'.format(rand))

time.sleep(rand)

return True

Step 2:

The next step is to create a second Lambda function, that I called Iterator, which has two duties:

  • It keeps track of the current number of iterations, since Step Function doesn’t natively have a state we can use for this purpose.
  • It asynchronously invokes our Lambda function at every loops.

This is the code of the Iterator, adapted from here.

 

import boto3

client = boto3.client('kinesis')

def lambda_handler(event, context):

index = event['iterator']['index'] + 1

response = client.put_record(

StreamName='LambdaSubMinute',

PartitionKey='1',

Data='',

)

return {

'index': index,

'continue': index < event['iterator']['count'],

'count': event['iterator']['count']

}

This function does three things:

  • Increments the counter.
  • Verifies if we reached a count of (in this example) 6.
  • Sends an empty record to the Kinesis Stream.

Now we can create the Step Functions State Machine; the definition is, again, adapted from here.

 

{

"Comment": "Invoke Lambda every 10 seconds",

"StartAt": "ConfigureCount",

"States": {

"ConfigureCount": {

"Type": "Pass",

"Result": {

"index": 0,

"count": 6

},

"ResultPath": "$.iterator",

"Next": "Iterator"

},

"Iterator": {

"Type": "Task",

"Resource": “arn:aws:lambda:REGION:ACCOUNT_ID:function:Iterator",

"ResultPath": "$.iterator",

"Next": "IsCountReached"

},

"IsCountReached": {

"Type": "Choice",

"Choices": [

{

"Variable": "$.iterator.continue",

"BooleanEquals": true,

"Next": "Wait"

}

],

"Default": "Done"

},

"Wait": {

"Type": "Wait",

"Seconds": 10,

"Next": "Iterator"

},

"Done": {

"Type": "Pass",

"End": true

}

}

}

This is how it works:

  1. The state machine starts and sets the index at 0 and the count at 6.
  2. Iterator function is invoked.
  3. If the iterator function reached the end of the loop, the IsCountReached state terminates the execution, otherwise the machine waits for 10 seconds.
  4. The machine loops back to the iterator.

Step 3: Create an Amazon CloudWatch Events rule scheduled to trigger every minute and add the state machine as its target. I’ve actually prepared an Amazon CloudFormation template that creates the whole stack and starts the Lambda invocations, you can find it here.

Performance

Let’s have a look at a sample series of invocations and analyse how precise the timing is. In the following chart I reported the delay (in excess of the expected 10-second-wait) of 30 consecutive invocations of my dummy function, when the Iterator is configured with a memory size of 1024MB.

Invocations Delay

Notice the delay increases by a few hundred milliseconds at every invocation. The good news is it accrues only within the same loop, 6 times; after that, a new CloudWatch Events kicks in and it resets.

This delay  is due to the work that AWS Step Function does outside of the Wait state, the main component of which is the Iterator function itself, that runs synchronously in the state machine and therefore adds up its duration to the 10-second-wait.

As we can easily imagine, the memory size of the Iterator Lambda function does make a difference. Here are the Average and Maximum duration of the function with 256MB, 512MB, 1GB and 2GB of memory.

Average Duration

Maximum Duration


Given those results, I’d say that a memory of 1024MB is a good compromise between costs and performance.

Caveats

As mentioned, in our Amazon CloudWatch Events documentation, in rare cases a rule can be triggered twice, causing two parallel executions of the state machine. If that is a concern, we can add a task state at the beginning of the state machine that checks if any other executions are currently running. If the outcome is positive, then a choice state can immediately terminate the flow. Since the state machine is invoked every 60 seconds and runs for about 50, it is safe to assume that executions should all be sequential and any parallel executions should be treated as duplicates. The task state that checks for current running executions can be a Lambda function similar to the following:

 

import boto3

client = boto3.client('stepfunctions')

def lambda_handler(event, context):

response = client.list_executions(

stateMachineArn='arn:aws:states:REGION:ACCOUNTID:stateMachine:LambdaSubMinute',

statusFilter='RUNNING'

)

return {

'alreadyRunning': len(response['executions']) > 0

}

About the Author

Emanuele Menga, Cloud Support Engineer

 

2018-05-03 python, multiprocessing, thread-ове и забивания

Post Syndicated from Vasil Kolev original https://vasil.ludost.net/blog/?p=3384

Всеки ден се убеждавам, че нищо не работи.

Открих забавен проблем с python и multiprocessing, който в момента още не мога да реша чий проблем е (в крайна сметка ще се окаже мой). Отне ми прилично количество време да го хвана и си струва да го разкажа.

Малко предистория: ползваме influxdb, в което тъпчем бая секундни данни, които после предъвкваме до минутни. InfluxDB има continuous queries, които вършат тази работа – на някакъв интервал от време хващат новите данни и ги сгъват. Тези заявки имаха няколко проблема:
– не се оправят с попълване на стари данни;
– изпълняват се рядко и минутните данни изостават;
– изпълняват се в общи линии в един thread, което кара минутните данни да изостават още повече (в нашия случай преди да ги сменим с около 12 часа).

Хванаха ме дяволите и си написах просто демонче на python, което да събира информация за различните бази какви данни могат да се сгънат, и паралелно да попълва данните. Работи в общи линии по следния начин:
– взима списък с базите данни
– пуска през multiprocessing-а да се събере за всяка база какви заявки трябва да се пуснат, на база на какви measurement-и има и докога са минутните и секундните данни в тях;
– пуска през multiprocessing-а събраните от предния pass заявки
– и така до края на света (или докато зависне).

След като навакса за няколко часа, успяваше да държи минутните данни в рамките на няколко минути от последните секундни данни, което си беше сериозно подобрение на ситуацията. Единственият проблем беше, че от време на време спираше да process-ва и увисваше.

Днес намерих време да го прегледам внимателно какво му се случва. Процесът изглежда като един parent и 5 fork()-нати child-а, като:
Parent-а спи във futex 0x22555a0;
Child 18455 във futex 0x7fdbfa366000;
Child 18546 read
Child 18457 във futex 0x7fdbfa366000
Child 18461 във futex 0x7fdbfa366000
Child 18462 във futex 0x7fdbfa366000
Child 18465 във futex 0x7fdbf908c2c0

Това не беше особено полезно, и се оказа, че стандартния python debugger (pdb) не може да се закача за съществуващи процеси, но за сметка на това gdb с подходящи debug символи може, и може да дава доста полезна информация. По този начин открих, че parent-а чака един child да приключи работата си:


#11 PyEval_EvalFrameEx (
[email protected]=Frame 0x235fb80, for file /usr/lib64/python2.7/multiprocessing/pool.py, line 543, in wait (self== 1525137960000000000 AND time < 1525138107000000000 GROUP BY time(1m), * fill(linear)\' in a read only context, please use a POST request instead', u'level': u'warning'}], u'statement_id': 0}]}, None], _callback=None, _chunksize=1, _number_left=1, _ready=False, _success=True, _cond=<_Condition(_Verbose__verbose=False, _Condition__lock=, acquire=, _Condition__waiters=[], release=) at remote 0x7fdbe0015310>, _job=45499, _cache={45499: < ...>}) a...(truncated), [email protected]=0) at /usr/src/debug/Python-2.7.5/Python/ceval.c:3040

Като в pool.py около ред 543 има следното:


class ApplyResult(object):

...

def wait(self, timeout=None):
self._cond.acquire()
try:
if not self._ready:
self._cond.wait(timeout)
finally:
self._cond.release()

Първоначално си мислех, че 18546 очаква да прочете нещо от грешното място, но излезе, че това е child-а, който е спечелил състезанието за изпълняване на следващата задача и чака да му я дадат (което изглежда се раздава през futex 0x7fdbfa366000). Един от child-овете обаче чака в друг lock:


(gdb) bt
#0 __lll_lock_wait () at ../nptl/sysdeps/unix/sysv/linux/x86_64/lowlevellock.S:135
#1 0x00007fdbf9b68dcb in _L_lock_812 () from /lib64/libpthread.so.0
#2 0x00007fdbf9b68c98 in __GI___pthread_mutex_lock ([email protected]=0x7fdbf908c2c0 ) at ../nptl/pthread_mutex_lock.c:79
#3 0x00007fdbf8e846ea in _nss_files_gethostbyname4_r ([email protected]=0x233fa44 "localhost", [email protected]=0x7fdbecfcb8e0, [email protected]=0x7fdbecfcb340 "hZ \372\333\177",
[email protected]=1064, [email protected]=0x7fdbecfcb8b0, [email protected]=0x7fdbecfcb910, [email protected]=0x0) at nss_files/files-hosts.c:381
#4 0x00007fdbf9170ed8 in gaih_inet (name=, [email protected]=0x233fa44 "localhost", service=, [email protected]=0x7fdbecfcbb90, [email protected]=0x7fdbecfcb9f0,
[email protected]=0x7fdbecfcb9e0) at ../sysdeps/posix/getaddrinfo.c:877
#5 0x00007fdbf91745cd in __GI_getaddrinfo ([email protected]=0x233fa44 "localhost", [email protected]=0x7fdbecfcbbc0 "8086", [email protected]=0x7fdbecfcbb90, [email protected]=0x7fdbecfcbb78)
at ../sysdeps/posix/getaddrinfo.c:2431
#6 0x00007fdbeed8760d in socket_getaddrinfo (self=
, args=) at /usr/src/debug/Python-2.7.5/Modules/socketmodule.c:4193
#7 0x00007fdbf9e5fbb0 in call_function (oparg=
, pp_stack=0x7fdbecfcbd10) at /usr/src/debug/Python-2.7.5/Python/ceval.c:4408
#8 PyEval_EvalFrameEx (
[email protected]=Frame 0x7fdbe8013350, for file /usr/lib/python2.7/site-packages/urllib3/util/connection.py, line 64, in create_connection (address=('localhost', 8086), timeout=3000, source_address=None, socket_options=[(6, 1, 1)], host='localhost', port=8086, err=None), [email protected]=0) at /usr/src/debug/Python-2.7.5/Python/ceval.c:3040

(gdb) frame 3
#3 0x00007fdbf8e846ea in _nss_files_gethostbyname4_r ([email protected]=0x233fa44 "localhost", [email protected]=0x7fdbecfcb8e0, [email protected]=0x7fdbecfcb340 "hZ \372\333\177",
[email protected]=1064, [email protected]=0x7fdbecfcb8b0, [email protected]=0x7fdbecfcb910, [email protected]=0x0) at nss_files/files-hosts.c:381
381 __libc_lock_lock (lock);
(gdb) list
376 enum nss_status
377 _nss_files_gethostbyname4_r (const char *name, struct gaih_addrtuple **pat,
378 char *buffer, size_t buflen, int *errnop,
379 int *herrnop, int32_t *ttlp)
380 {
381 __libc_lock_lock (lock);
382
383 /* Reset file pointer to beginning or open file. */
384 enum nss_status status = internal_setent (keep_stream);
385

Или в превод – опитваме се да вземем стандартния lock, който libc-то използва за да си пази reentrant функциите, и някой го държи. Кой ли?


(gdb) p lock
$3 = {__data = {__lock = 2, __count = 0, __owner = 16609, __nusers = 1, __kind = 0, __spins = 0, __elision = 0, __list = {__prev = 0x0, __next = 0x0}},
__size = "\002\000\000\000\000\000\000\000\[email protected]\000\000\001", '\000' , __align = 2}
(gdb) p &lock
$4 = (__libc_lock_t *) 0x7fdbf908c2c0

Тук се вижда как owner-а на lock-а всъщност е parent-а. Той обаче не смята, че го държи:


(gdb) p lock
$2 = 0
(gdb) p &lock
$3 = (__libc_lock_t *) 0x7fdbf9450df0
(gdb) x/20x 0x7fdbf9450df0
0x7fdbf9450df0
: 0x00000000 0x00000000 0x00000000 0x00000000
0x7fdbf9450e00 <__abort_msg>: 0x00000000 0x00000000 0x00000000 0x00000000
0x7fdbf9450e10 : 0x00000000 0x00000000 0x00000000 0x00000000
0x7fdbf9450e20 : 0x00000000 0x00000000 0x00000000 0x00000000
0x7fdbf9450e30 : 0x001762c9 0x00000000 0x00000000 0x00000000

… което е и съвсем очаквано, при условие, че са два процеса и тая памет не е обща.

Та, явно това, което се е случило е, че докато parent-а е правел fork(), тоя lock го е държал някой, и child-а реално не може да пипне каквото и да е, свързано с него (което значи никакви reentrant функции в glibc-то, каквито па всички ползват (и би трябвало да ползват)). Въпросът е, че по принцип това не би трябвало да е възможно, щото около fork() няма нищо, което да взима тоя lock, и би трябвало glibc да си освобождава lock-а като излиза от функциите си.

Първоначалното ми идиотско предположение беше, че в signal handler-а на SIGCHLD multiprocessing модула създава новите child-ове, и така докато нещо друго държи lock-а идва сигнал, прави се нов процес и той го “наследява” заключен. Това беше твърде глупаво, за да е истина, и се оказа, че не е…

Около въпросите с lock-а бях стигнал с търсене до две неща – issue 127 в gperftools и Debian bug 657835. Първото каза, че проблемът ми може да е от друг lock, който някой друг държи преди fork-а (което ме накара да се загледам по-внимателно какви lock-ове се държат), а второто, че като цяло ако fork-ваш thread-нато приложение, може после единствено да правиш execve(), защото всичко друго не е ясно колко ще работи.

И накрая се оказа, че ако се ползва multiprocessing модула, той пуска в главния процес няколко thread-а, които да се занимават със следенето и пускането на child-ове за обработка. Та ето какво реално се случва:

– някой child си изработва нужния брой операции и излиза
– parent-а получава SIGCHLD и си отбелязва, че трябва да види какво става
– главния thread на parent-а тръгва да събира списъка бази, и вика в някакъв момент _nss_files_gethostbyname4_r, който взима lock-а;
– по това време другия thread казва “а, нямам достатъчно child-ове, fork()”
– profit.

Текущото ми глупаво решение е да не правя нищо в главния thread, което може да взима тоя lock и да се надявам, че няма още някой такъв. Бъдещото ми решение е или да го пиша на python3 с някой друг модул по темата, или на go (което ще трябва да науча).

[$] The second half of the 4.17 merge window

Post Syndicated from corbet original https://lwn.net/Articles/751482/rss

By the time the 4.17 merge window was closed and 4.17-rc1 was
released, 11,769 non-merge changesets had been pulled into the
mainline repository. 4.17 thus looks to be a typically busy development
cycle, with a merge window only slightly more busy than 4.16 had.
Some 6,000 of those changes were pulled after last week’s summary was written. There was a
lot of the usual maintenance work in those patches (over 10% of those
changes were to device-tree files, for example), but also some more
significant changes.

COPPA Compliance

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

Interesting research: “‘Won’t Somebody Think of the Children?’ Examining COPPA Compliance at Scale“:

Abstract: We present a scalable dynamic analysis framework that allows for the automatic evaluation of the privacy behaviors of Android apps. We use our system to analyze mobile apps’ compliance with the Children’s Online Privacy Protection Act (COPPA), one of the few stringent privacy laws in the U.S. Based on our automated analysis of 5,855 of the most popular free children’s apps, we found that a majority are potentially in violation of COPPA, mainly due to their use of third-party SDKs. While many of these SDKs offer configuration options to respect COPPA by disabling tracking and behavioral advertising, our data suggest that a majority of apps either do not make use of these options or incorrectly propagate them across mediation SDKs. Worse, we observed that 19% of children’s apps collect identifiers or other personally identifiable information (PII) via SDKs whose terms of service outright prohibit their use in child-directed apps. Finally, we show that efforts by Google to limit tracking through the use of a resettable advertising ID have had little success: of the 3,454 apps that share the resettable ID with advertisers, 66% transmit other, non-resettable, persistent identifiers as well, negating any intended privacy-preserving properties of the advertising ID.

uTorrent Flagged as ‘Threat’ by Microsoft and Anti-Virus Vendors

Post Syndicated from Ernesto original https://torrentfreak.com/utorrent-flagged-as-threat-by-microsoft-and-anti-virus-vendors-180312/

Installed on dozens of millions of devices, uTorrent is the go-to torrent client for people all around the world.

While the software usually runs without hassle, many users started to experience problems recently. Several anti-virus tools, including Windows Defender, suddenly labeled the torrent client as dangerous.

Microsoft categorizes the affected clients as “Potentially Unwanted Software,” as can be seen below. The company has had a dedicated Utorrent page for a while, labeling it as a severe threat. This week, however, alarm bells started to go off on a broader scale.

uTorrent threat

It’s unclear what exactly triggered the recent warning. According to VirusTotal, a handful of anti-virus companies label uTorrent as problematic. ESET-NOD32 lists “Web Companion” as the trigger, which likely points to Lavasoft’s Ad-Aware software, which is sometimes bundled with uTorrent.

uTorrent parent company BitTorrent Inc. is aware of the problems but believes they’re false positives triggered by one of their recent releases.

“We believe that this passive flag changed to active just hours ago with the Windows patch Tuesday update, when a small percent of users started getting an explicit block,” the company told us.

“We had three uTorrent executables being served from our site. Two were going to 95% of our users and were not part of the Windows block. The third, which was going to 5% of users, was part of the Windows block. We stopped shipping that and confirmed we are no longer seeing any blocks.”

The issue doesn’t appear to be restricted to new installs only. Several users have reported that their uTorrent application was suddenly quarantined as unwanted software, possibly after an automatic update.

We rechecked the VirusTotal result with the most current uTorrent release, and this is still flagged by six anti-virus vendors.

VirusTotal results

But that’s not all. The uTorrent download page itself also triggers a warning from MalwareBytes’ real-time protection module, which brands the website itself as malicious.

Interestingly, when trying to install uTorrent, Windows lists Lavasoft Software Canada as the verified publisher. While Lavasoft’s “Ad-Aware WebCompanion” is regularly bundled with uTorrent as an ‘offer,’ we didn’t get that option when we last tried, nor was it installed.

After we installed it during an initial test yesterday, we did notice that WebCompanion was installed around the same time. However, we have been unable to replicate this result.

BitTorrent Inc. stresses that any of the offers users get during the install process are optional, Google-compliant, and in accordance with the Clean Software Alliance (CSA) standards.

Whatever is causing the red flags at Microsoft and the other companies remains a mystery for now, also for BitTorrent Inc.

“Based on our best assessment to date, we’ve found no reason why we would be blocked – especially on some builds and not others which are basically identical,” BitTorrent says.

“We are continuing to reach out, though, and hope to have more information,” the company adds.

Source: TF, for the latest info on copyright, file-sharing, torrent sites and more. We also have VPN reviews, discounts, offers and coupons.

User Authentication Best Practices Checklist

Post Syndicated from Bozho original https://techblog.bozho.net/user-authentication-best-practices-checklist/

User authentication is the functionality that every web application shared. We should have perfected that a long time ago, having implemented it so many times. And yet there are so many mistakes made all the time.

Part of the reason for that is that the list of things that can go wrong is long. You can store passwords incorrectly, you can have a vulnerably password reset functionality, you can expose your session to a CSRF attack, your session can be hijacked, etc. So I’ll try to compile a list of best practices regarding user authentication. OWASP top 10 is always something you should read, every year. But that might not be enough.

So, let’s start. I’ll try to be concise, but I’ll include as much of the related pitfalls as I can cover – e.g. what could go wrong with the user session after they login:

  • Store passwords with bcrypt/scrypt/PBKDF2. No MD5 or SHA, as they are not good for password storing. Long salt (per user) is mandatory (the aforementioned algorithms have it built in). If you don’t and someone gets hold of your database, they’ll be able to extract the passwords of all your users. And then try these passwords on other websites.
  • Use HTTPS. Period. (Otherwise user credentials can leak through unprotected networks). Force HTTPS if user opens a plain-text version.
  • Mark cookies as secure. Makes cookie theft harder.
  • Use CSRF protection (e.g. CSRF one-time tokens that are verified with each request). Frameworks have such functionality built-in.
  • Disallow framing (X-Frame-Options: DENY). Otherwise your website may be included in another website in a hidden iframe and “abused” through javascript.
  • Have a same-origin policy
  • Logout – let your users logout by deleting all cookies and invalidating the session. This makes usage of shared computers safer (yes, users should ideally use private browsing sessions, but not all of them are that savvy)
  • Session expiry – don’t have forever-lasting sessions. If the user closes your website, their session should expire after a while. “A while” may still be a big number depending on the service provided. For ajax-heavy website you can have regular ajax-polling that keeps the session alive while the page stays open.
  • Remember me – implementing “remember me” (on this machine) functionality is actually hard due to the risks of a stolen persistent cookie. Spring-security uses this approach, which I think should be followed if you wish to implement more persistent logins.
  • Forgotten password flow – the forgotten password flow should rely on sending a one-time (or expiring) link to the user and asking for a new password when it’s opened. 0Auth explain it in this post and Postmark gives some best pracitces. How the link is formed is a separate discussion and there are several approaches. Store a password-reset token in the user profile table and then send it as parameter in the link. Or do not store anything in the database, but send a few params: userId:expiresTimestamp:hmac(userId+expiresTimestamp). That way you have expiring links (rather than one-time links). The HMAC relies on a secret key, so the links can’t be spoofed. It seems there’s no consensus, as the OWASP guide has a bit different approach
  • One-time login links – this is an option used by Slack, which sends one-time login links instead of asking users for passwords. It relies on the fact that your email is well guarded and you have access to it all the time. If your service is not accessed to often, you can have that approach instead of (rather than in addition to) passwords.
  • Limit login attempts – brute-force through a web UI should not be possible; therefore you should block login attempts if they become too many. One approach is to just block them based on IP. The other one is to block them based on account attempted. (Spring example here). Which one is better – I don’t know. Both can actually be combined. Instead of fully blocking the attempts, you may add a captcha after, say, the 5th attempt. But don’t add the captcha for the first attempt – it is bad user experience.
  • Don’t leak information through error messages – you shouldn’t allow attackers to figure out if an email is registered or not. If an email is not found, upon login report just “Incorrect credentials”. On passwords reset, it may be something like “If your email is registered, you should have received a password reset email”. This is often at odds with usability – people don’t often remember the email they used to register, and the ability to check a number of them before getting in might be important. So this rule is not absolute, though it’s desirable, especially for more critical systems.
  • Make sure you use JWT only if it’s really necessary and be careful of the pitfalls.
  • Consider using a 3rd party authentication – OpenID Connect, OAuth by Google/Facebook/Twitter (but be careful with OAuth flaws as well). There’s an associated risk with relying on a 3rd party identity provider, and you still have to manage cookies, logout, etc., but some of the authentication aspects are simplified.
  • For high-risk or sensitive applications use 2-factor authentication. There’s a caveat with Google Authenticator though – if you lose your phone, you lose your accounts (unless there’s a manual process to restore it). That’s why Authy seems like a good solution for storing 2FA keys.

I’m sure I’m missing something. And you see it’s complicated. Sadly we’re still at the point where the most common functionality – authenticating users – is so tricky and cumbersome, that you almost always get at least some of it wrong.

The post User Authentication Best Practices Checklist appeared first on Bozho's tech blog.

Safety first: a Raspberry Pi safety helmet

Post Syndicated from Alex Bate original https://www.raspberrypi.org/blog/safety-helmet/

Jennifer Fox is back, this time with a Raspberry Pi Zero–controlled impact force monitor that will notify you if your collision is a worth a trip to the doctor.

Make an Impact Force Monitor!

Check out my latest Hacker in Residence project for SparkFun Electronics: the Helmet Guardian! It’s a Pi Zero powered impact force monitor that turns on an LED if your head/body experiences a potentially dangerous impact. Install in your sports helmets, bicycle, or car to keep track of impact and inform you when it’s time to visit the doctor.

Concussion

We’ve all knocked our heads at least once in our lives, maybe due to tripping over a loose paving slab, or to falling off a bike, or to walking into the corner of the overhead cupboard door for the third time this week — will I ever learn?! More often than not, even when we’re seeing stars, we brush off the accident and continue with our day, oblivious to the long-term damage we may be doing.

Force of impact

After some thorough research, Jennifer Fox, founder of FoxBot Industries, concluded that forces of 4 to 6 G sustained for more than a few seconds are dangerous to the human body. With this in mind, she decided to use a Raspberry Pi Zero W and an accelerometer to create helmet with an impact force monitor that notifies its wearer if this level of G-force has been met.

Jennifer Fox Raspberry Pi Impact Force Monitor

Obviously, if you do have a serious fall, you should always seek medical advice. This project is an example of how affordable technology can be used to create medical and citizen science builds, and not a replacement for professional medical services.

Setting up the impact monitor

Jennifer’s monitor requires only a few pieces of tech: a Zero W, an accelerometer and breakout board, a rechargeable USB battery, and an LED, plus the standard wires and resistors for these components.

After installing Raspbian, Jennifer enabled SSH and I2C on the Zero W to make it run headlessly, and then accessed it from a laptop. This allows her to control the Pi without physically connecting to it, and it makes for a wireless finished project.

Jen wired the Pi to the accelerometer breakout board and LED as shown in the schematic below.

Jennifer Fox Raspberry Pi Impact Force Monitor

The LED acts as a signal of significant impacts, turning on when the G-force threshold is reached, and not turning off again until the program is reset.

Jennifer Fox Raspberry Pi Impact Force Monitor

Make your own and more

Jennifer’s full code for the impact monitor is on GitHub, and she’s put together a complete tutorial on SparkFun’s website.

For more tutorials from Jennifer Fox, such as her ‘Bark Back’ IoT Pet Monitor, be sure to follow her on YouTube. And for similar projects, check out Matt’s smart bike light and Amelia Day’s physical therapy soccer ball.

The post Safety first: a Raspberry Pi safety helmet appeared first on Raspberry Pi.

[$] The first half of the 4.17 merge window

Post Syndicated from corbet original https://lwn.net/Articles/750928/rss

As of this writing, 5,392 non-merge changesets have been pulled into the
mainline repository for the 4.17 release. The 4.17 merge window is thus
off to a good start, but it is far from complete. The changes pulled thus
far cover a wide part of the core kernel as well as the networking, driver,
and filesystem subsystems.

Linux kernel lockdown and UEFI Secure Boot

Post Syndicated from Matthew Garrett original https://mjg59.dreamwidth.org/50577.html

David Howells recently published the latest version of his kernel lockdown patchset. This is intended to strengthen the boundary between root and the kernel by imposing additional restrictions that prevent root from modifying the kernel at runtime. It’s not the first feature of this sort – /dev/mem no longer allows you to overwrite arbitrary kernel memory, and you can configure the kernel so only signed modules can be loaded. But the present state of things is that these security features can be easily circumvented (by using kexec to modify the kernel security policy, for instance).

Why do you want lockdown? If you’ve got a setup where you know that your system is booting a trustworthy kernel (you’re running a system that does cryptographic verification of its boot chain, or you built and installed the kernel yourself, for instance) then you can trust the kernel to keep secrets safe from even root. But if root is able to modify the running kernel, that guarantee goes away. As a result, it makes sense to extend the security policy from the boot environment up to the running kernel – it’s really just an extension of configuring the kernel to require signed modules.

The patchset itself isn’t hugely conceptually controversial, although there’s disagreement over the precise form of certain restrictions. But one patch has, because it associates whether or not lockdown is enabled with whether or not UEFI Secure Boot is enabled. There’s some backstory that’s important here.

Most kernel features get turned on or off by either build-time configuration or by passing arguments to the kernel at boot time. There’s two ways that this patchset allows a bootloader to tell the kernel to enable lockdown mode – it can either pass the lockdown argument on the kernel command line, or it can set the secure_boot flag in the bootparams structure that’s passed to the kernel. If you’re running in an environment where you’re able to verify the kernel before booting it (either through cryptographic validation of the kernel, or knowing that there’s a secret tied to the TPM that will prevent the system booting if the kernel’s been tampered with), you can turn on lockdown.

There’s a catch on UEFI systems, though – you can build the kernel so that it looks like an EFI executable, and then run it directly from the firmware. The firmware doesn’t know about Linux, so can’t populate the bootparam structure, and there’s no mechanism to enforce command lines so we can’t rely on that either. The controversial patch simply adds a kernel configuration option that automatically enables lockdown when UEFI secure boot is enabled and otherwise leaves it up to the user to choose whether or not to turn it on.

Why do we want lockdown enabled when booting via UEFI secure boot? UEFI secure boot is designed to prevent the booting of any bootloaders that the owner of the system doesn’t consider trustworthy[1]. But a bootloader is only software – the only thing that distinguishes it from, say, Firefox is that Firefox is running in user mode and has no direct access to the hardware. The kernel does have direct access to the hardware, and so there’s no meaningful distinction between what grub can do and what the kernel can do. If you can run arbitrary code in the kernel then you can use the kernel to boot anything you want, which defeats the point of UEFI Secure Boot. Linux distributions don’t want their kernels to be used to be used as part of an attack chain against other distributions or operating systems, so they enable lockdown (or equivalent functionality) for kernels booted this way.

So why not enable it everywhere? There’s a couple of reasons. The first is that some of the features may break things people need – for instance, some strange embedded apps communicate with PCI devices by mmap()ing resources directly from sysfs[2]. This is blocked by lockdown, which would break them. Distributions would then have to ship an additional kernel that had lockdown disabled (it’s not possible to just have a command line argument that disables it, because an attacker could simply pass that), and users would have to disable secure boot to boot that anyway. It’s easier to just tie the two together.

The second is that it presents a promise of security that isn’t really there if your system didn’t verify the kernel. If an attacker can replace your bootloader or kernel then the ability to modify your kernel at runtime is less interesting – they can just wait for the next reboot. Appearing to give users safety assurances that are much less strong than they seem to be isn’t good for keeping users safe.

So, what about people whose work is impacted by lockdown? Right now there’s two ways to get stuff blocked by lockdown unblocked: either disable secure boot[3] (which will disable it until you enable secure boot again) or press alt-sysrq-x (which will disable it until the next boot). Discussion has suggested that having an additional secure variable that disables lockdown without disabling secure boot validation might be helpful, and it’s not difficult to implement that so it’ll probably happen.

Overall: the patchset isn’t controversial, just the way it’s integrated with UEFI secure boot. The reason it’s integrated with UEFI secure boot is because that’s the policy most distributions want, since the alternative is to enable it everywhere even when it doesn’t provide real benefits but does provide additional support overhead. You can use it even if you’re not using UEFI secure boot. We should have just called it securelevel.

[1] Of course, if the owner of a system isn’t allowed to make that determination themselves, the same technology is restricting the freedom of the user. This is abhorrent, and sadly it’s the default situation in many devices outside the PC ecosystem – most of them not using UEFI. But almost any security solution that aims to prevent malicious software from running can also be used to prevent any software from running, and the problem here is the people unwilling to provide that policy to users rather than the security features.
[2] This is how X.org used to work until the advent of kernel modesetting
[3] If your vendor doesn’t provide a firmware option for this, run sudo mokutil –disable-validation

comment count unavailable comments

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…

[$] An introduction to projectM

Post Syndicated from jake original https://lwn.net/Articles/750152/rss

Many people have seen music visualizations before, whether in a music
player on
their computer, at a live concert, or possibly on a home stereo
system. Those visualizations may have been generated using the open-source
music-visualization software library that is part of projectM.
Software-based
abstract visualizers first appeared along with early MP3 music players as a
sort of nifty thing to watch along with listening to your MP3s. One of
the most powerful and innovative of these was a plugin for Winamp known as
MilkDrop, which was
developed by a
Nullsoft (and later NVIDIA) employee named Ryan Geiss. The plugin was
extensible by using visualization
equation scripts
(also known as “presets”).

Subscribers can read on for a look at projectM by guest author (and
projectM maintainer) Mischa Spiegelmock.

New – Amazon DynamoDB Continuous Backups and Point-In-Time Recovery (PITR)

Post Syndicated from Randall Hunt original https://aws.amazon.com/blogs/aws/new-amazon-dynamodb-continuous-backups-and-point-in-time-recovery-pitr/

The Amazon DynamoDB team is back with another useful feature hot on the heels of encryption at rest. At AWS re:Invent 2017 we launched global tables and on-demand backup and restore of your DynamoDB tables and today we’re launching continuous backups with point-in-time recovery (PITR).

You can enable continuous backups with a single click in the AWS Management Console, a simple API call, or with the AWS Command Line Interface (CLI). DynamoDB can back up your data with per-second granularity and restore to any single second from the time PITR was enabled up to the prior 35 days. We built this feature to protect against accidental writes or deletes. If a developer runs a script against production instead of staging or if someone fat-fingers a DeleteItem call, PITR has you covered. We also built it for the scenarios you can’t normally predict. You can still keep your on-demand backups for as long as needed for archival purposes but PITR works as additional insurance against accidental loss of data. Let’s see how this works.

Continuous Backup

To enable this feature in the console we navigate to our table and select the Backups tab. From there simply click Enable to turn on the feature. I could also turn on continuous backups via the UpdateContinuousBackups API call.

After continuous backup is enabled we should be able to see an Earliest restore date and Latest restore date

Let’s imagine a scenario where I have a lot of old user profiles that I want to delete.

I really only want to send service updates to our active users based on their last_update date. I decided to write a quick Python script to delete all the users that haven’t used my service in a while.

import boto3
table = boto3.resource("dynamodb").Table("VerySuperImportantTable")
items = table.scan(
    FilterExpression="last_update >= :date",
    ExpressionAttributeValues={":date": "2014-01-01T00:00:00"},
    ProjectionExpression="ImportantId"
)['Items']
print("Deleting {} Items! Dangerous.".format(len(items)))
with table.batch_writer() as batch:
    for item in items:
        batch.delete_item(Key=item)

Great! This should delete all those pesky non-users of my service that haven’t logged in since 2013. So,— CTRL+C CTRL+C CTRL+C CTRL+C (interrupt the currently executing command).

Yikes! Do you see where I went wrong? I’ve just deleted my most important users! Oh, no! Where I had a greater-than sign, I meant to put a less-than! Quick, before Jeff Barr can see, I’m going to restore the table. (I probably could have prevented that typo with Boto 3’s handy DynamoDB conditions: Attr("last_update").lt("2014-01-01T00:00:00"))

Restoring

Luckily for me, restoring a table is easy. In the console I’ll navigate to the Backups tab for my table and click Restore to point-in-time.

I’ll specify the time (a few seconds before I started my deleting spree) and a name for the table I’m restoring to.

For a relatively small and evenly distributed table like mine, the restore is quite fast.

The time it takes to restore a table varies based on multiple factors and restore times are not neccesarily coordinated with the size of the table. If your dataset is evenly distributed across your primary keys you’ll be able to take advanatage of parallelization which will speed up your restores.

Learn More & Try It Yourself
There’s plenty more to learn about this new feature in the documentation here.

Pricing for continuous backups varies by region and is based on the current size of the table and all indexes.

A few things to note:

  • PITR works with encrypted tables.
  • If you disable PITR and later reenable it, you reset the start time from which you can recover.
  • Just like on-demand backups, there are no performance or availability impacts to enabling this feature.
  • Stream settings, Time To Live settings, PITR settings, tags, Amazon CloudWatch alarms, and auto scaling policies are not copied to the restored table.
  • Jeff, it turns out, knew I restored the table all along because every PITR API call is recorded in AWS CloudTrail.

Let us know how you’re going to use continuous backups and PITR on Twitter and in the comments.
Randall

Stone: A new era for Linux’s low-level graphics – Part 1

Post Syndicated from corbet original https://lwn.net/Articles/749790/rss

Daniel Stone begins
a series
on how the Linux graphic stack has improved in recent times.
This has made mainline Linux much more attractive: the exact same
generic codebases of GNOME and Weston that I’m using to write this blog
post on an Intel laptop run equally well on AMD workstations, low-power NXP
boards destined for in-flight entertainment, and high-end Renesas SoCs
which might well be in your car. Now that the drivers are easy to write,
and applications are portable, we’ve seen over ten new DRM drivers merged
to the upstream kernel since atomic modesetting was merged.

One LED Matrix Table to rule them all

Post Syndicated from Alex Bate original https://www.raspberrypi.org/blog/led-matrix-table/

Germany-based Andreas Rottach’s multi-purpose LED table is an impressive build within a gorgeous-looking body. Play games, view (heavily pixelated) images, and become hypnotised by flashy lights, once you’ve built your own using his newly released tutorial.

LED-Matrix Table – 300 LEDs – Raspberry Pi – C++ Engine – Custom Controllers

This is a short presentation of my LED-Matrix Table. The table is controlled by a raspberry pi computer that executes a control engine, written in c++. It supports input from keyboards or custom made game controllers. A full list of all features as well as the source code is available on GitHub (https://github.com/rottaca/LEDTableEngine).

Much excitement

Andreas uploaded a video of his LED Matrix Table to YouTube back in February, with the promise of publishing a complete write-up within the coming weeks. And so the members of Pi Towers sat, eagerly waiting and watching. Now the write-up has arrived, to our cheers of acclaim for this beautful, shiny, flashy, LED-based wonderment.

Build your own LED table

In his GitHub tutorial, Andreas goes through all the stages of building the table, from the necessary components to coding the Raspberry Pi 3 and 3D printing your own controllers.

Raspberry Pi LED Table

Find files for the controllers on Thingiverse

Andreas created the table’s impressive light matrix using a strip of 300 LEDs, chained together and connected to the Raspberry Pi via an LED controller.

Raspberry Pi LED Table

The LEDs are set out in zigzags

For the code, he used several open-source tools, such as SDL for image and audio support, and CMake for building the project software.

Anyone planning to recreate Andreas’ table can compile its engine by downloading the project repository from GitHub. Again, find full instructions for this on his GitHub.

Features

The table boasts multiple cool features, including games and visualisation tools. Using the controllers, you can play simplified versions of Flappy Bird and Minesweeper, or go on a nostalgia trip with Tetris, Pong, and Snake.

Raspberry Pi LED Table

There’s also a version of Conway’s Game of Life. Andreas explains: “The lifespan of each cell is color-coded. If the game field gets static, the animation is automatically reset to a new random cell population.”

Raspberry Pi LED Table

The table can also display downsampled Bitmap images, or show clear static images such as a chess board, atop of which you can place physical game pieces.

Raspberry Pi LED Table
Raspberry Pi LED Table
Raspberry Pi LED Table

Find all the 3D-printable aspects of the LED table on Thingiverse here and here, and the full GitHub tutorial and repository here. If you build your own, or have already dabbled in LED tables and displays, be sure to share your project with us, either in the comments below or via our social media accounts. What other functions would you integrate into this awesome build?

The post One LED Matrix Table to rule them all appeared first on Raspberry Pi.

Serverless Dynamic Web Pages in AWS: Provisioned with CloudFormation

Post Syndicated from AWS Admin original https://aws.amazon.com/blogs/architecture/serverless-dynamic-web-pages-in-aws-provisioned-with-cloudformation/

***This blog is authored by Mike Okner of Monsanto, an AWS customer. It originally appeared on the Monsanto company blog. Minor edits were made to the original post.***

Recently, I was looking to create a status page app to monitor a few important internal services. I wanted this app to be as lightweight, reliable, and hassle-free as possible, so using a “serverless” architecture that doesn’t require any patching or other maintenance was quite appealing.

I also don’t deploy anything in a production AWS environment outside of some sort of template (usually CloudFormation) as a rule. I don’t want to have to come back to something I created ad hoc in the console after 6 months and try to recall exactly how I architected all of the resources. I’ll inevitably forget something and create more problems before solving the original one. So building the status page in a template was a requirement.

The Design
I settled on a design using two Lambda functions, both written in Python 3.6.

The first Lambda function makes requests out to a list of important services and writes their current status to a DynamoDB table. This function is executed once per minute via CloudWatch Event Rule.

The second Lambda function reads each service’s status & uptime information from DynamoDB and renders a Jinja template. This function is behind an API Gateway that has been configured to return text/html instead of its default application/json Content-Type.

The CloudFormation Template
AWS provides a Serverless Application Model template transformer to streamline the templating of Lambda + API Gateway designs, but it assumes (like everything else about the API Gateway) that you’re actually serving an API that returns JSON content. So, unfortunately, it won’t work for this use-case because we want to return HTML content. Instead, we’ll have to enumerate every resource like usual.

The Skeleton
We’ll be using YAML for the template in this example. I find it easier to read than JSON, but you can easily convert between the two with a converter if you disagree.

---
AWSTemplateFormatVersion: '2010-09-09'
Description: Serverless status page app
Resources:
  # [...Resources]

The Status-Checker Lambda Resource
This one is triggered on a schedule by CloudWatch, and looks like:

# Status Checker Lambda
CheckerLambda:
  Type: AWS::Lambda::Function
  Properties:
    Code: ./lambda.zip
    Environment:
      Variables:
        TABLE_NAME: !Ref DynamoTable
    Handler: checker.handler
    Role:
      Fn::GetAtt:
      - CheckerLambdaRole
      - Arn
    Runtime: python3.6
    Timeout: 45
CheckerLambdaRole:
  Type: AWS::IAM::Role
  Properties:
    ManagedPolicyArns:
    - arn:aws:iam::aws:policy/AmazonDynamoDBFullAccess
    - arn:aws:iam::aws:policy/service-role/AWSLambdaBasicExecutionRole
    AssumeRolePolicyDocument:
      Version: '2012-10-17'
      Statement:
      - Action:
        - sts:AssumeRole
        Effect: Allow
        Principal:
          Service:
          - lambda.amazonaws.com
CheckerLambdaTimer:
  Type: AWS::Events::Rule
  Properties:
    ScheduleExpression: rate(1 minute)
    Targets:
    - Id: CheckerLambdaTimerLambdaTarget
      Arn:
        Fn::GetAtt:
        - CheckerLambda
        - Arn
CheckerLambdaTimerPermission:
  Type: AWS::Lambda::Permission
  Properties:
    Action: lambda:invokeFunction
    FunctionName: !Ref CheckerLambda
    SourceArn:
      Fn::GetAtt:
      - CheckerLambdaTimer
      - Arn
    Principal: events.amazonaws.com

Let’s break that down a bit.

The CheckerLambda is the actual Lambda function. The Code section is a local path to a ZIP file containing the code and its dependencies. I’m using CloudFormation’s packaging feature to automatically push the deployable to S3.

The CheckerLambdaRole is the IAM role the Lambda will assume which grants it access to DynamoDB in addition to the usual Lambda logging permissions.

The CheckerLambdaTimer is the CloudWatch Events Rule that triggers the checker to run once per minute.

The CheckerLambdaTimerPermission grants CloudWatch the ability to invoke the checker Lambda function on its interval.

The Web Page Gateway
The API Gateway handles incoming requests for the web page, invokes the Lambda, and then returns the Lambda’s results as HTML content. Its template looks like:

# API Gateway for Web Page Lambda
PageGateway:
  Type: AWS::ApiGateway::RestApi
  Properties:
    Name: Service Checker Gateway
PageResource:
  Type: AWS::ApiGateway::Resource
  Properties:
    RestApiId: !Ref PageGateway
    ParentId:
      Fn::GetAtt:
      - PageGateway
      - RootResourceId
    PathPart: page
PageGatewayMethod:
  Type: AWS::ApiGateway::Method
  Properties:
    AuthorizationType: NONE
    HttpMethod: GET
    Integration:
      Type: AWS
      IntegrationHttpMethod: POST
      Uri:
        Fn::Sub: arn:aws:apigateway:${AWS::Region}:lambda:path/2015-03-31/functions/${WebRenderLambda.Arn}/invocations
      RequestTemplates:
        application/json: |
          {
              "method": "$context.httpMethod",
              "body" : $input.json('$'),
              "headers": {
                  #foreach($param in $input.params().header.keySet())
                  "$param": "$util.escapeJavaScript($input.params().header.get($param))"
                  #if($foreach.hasNext),#end
                  #end
              }
          }
      IntegrationResponses:
      - StatusCode: 200
        ResponseParameters:
          method.response.header.Content-Type: "'text/html'"
        ResponseTemplates:
          text/html: "$input.path('$')"
    ResourceId: !Ref PageResource
    RestApiId: !Ref PageGateway
    MethodResponses:
    - StatusCode: 200
      ResponseParameters:
        method.response.header.Content-Type: true
PageGatewayProdStage:
  Type: AWS::ApiGateway::Stage
  Properties:
    DeploymentId: !Ref PageGatewayDeployment
    RestApiId: !Ref PageGateway
    StageName: Prod
PageGatewayDeployment:
  Type: AWS::ApiGateway::Deployment
  DependsOn: PageGatewayMethod
  Properties:
    RestApiId: !Ref PageGateway
    Description: PageGateway deployment
    StageName: Stage

There’s a lot going on here, but the real meat is in the PageGatewayMethod section. There are a couple properties that deviate from the default which is why we couldn’t use the SAM transformer.

First, we’re passing request headers through to the Lambda in theRequestTemplates section. I’m doing this so I can validate incoming auth headers. The API Gateway can do some types of auth, but I found it easier to check auth myself in the Lambda function since the Gateway is designed to handle API calls and not browser requests.

Next, note that in the IntegrationResponses section we’re defining the Content-Type header to be ‘text/html’ (with single-quotes) and defining the ResponseTemplate to be $input.path(‘$’). This is what makes the request render as a HTML page in your browser instead of just raw text.

Due to the StageName and PathPart values in the other sections, your actual page will be accessible at https://someId.execute-api.region.amazonaws.com/Prod/page. I have the page behind an existing reverse-proxy and give it a saner URL for end-users. The reverse proxy also attaches the auth header I mentioned above. If that header isn’t present, the Lambda will render an error page instead so the proxy can’t be bypassed.

The Web Page Rendering Lambda
This Lambda is invoked by calls to the API Gateway and looks like:

# Web Page Lambda
WebRenderLambda:
  Type: AWS::Lambda::Function
  Properties:
    Code: ./lambda.zip
    Environment:
      Variables:
        TABLE_NAME: !Ref DynamoTable
    Handler: web.handler
    Role:
      Fn::GetAtt:
      - WebRenderLambdaRole
      - Arn
    Runtime: python3.6
    Timeout: 30
WebRenderLambdaRole:
  Type: AWS::IAM::Role
  Properties:
    ManagedPolicyArns:
    - arn:aws:iam::aws:policy/AmazonDynamoDBReadOnlyAccess
    - arn:aws:iam::aws:policy/service-role/AWSLambdaBasicExecutionRole
    AssumeRolePolicyDocument:
      Version: '2012-10-17'
      Statement:
      - Action:
        - sts:AssumeRole
        Effect: Allow
        Principal:
          Service:
          - lambda.amazonaws.com
WebRenderLambdaGatewayPermission:
  Type: AWS::Lambda::Permission
  Properties:
    FunctionName: !Ref WebRenderLambda
    Action: lambda:invokeFunction
    Principal: apigateway.amazonaws.com
    SourceArn:
      Fn::Sub:
      - arn:aws:execute-api:${AWS::Region}:${AWS::AccountId}:${__ApiId__}/*/*/*
      - __ApiId__: !Ref PageGateway

The WebRenderLambda and WebRenderLambdaRole should look familiar.

The WebRenderLambdaGatewayPermission is similar to the Status Checker’s CloudWatch permission, only this time it allows the API Gateway to invoke this Lambda.

The DynamoDB Table
This one is straightforward.

# DynamoDB table
DynamoTable:
  Type: AWS::DynamoDB::Table
  Properties:
    AttributeDefinitions:
    - AttributeName: name
      AttributeType: S
    ProvisionedThroughput:
      WriteCapacityUnits: 1
      ReadCapacityUnits: 1
    TableName: status-page-checker-results
    KeySchema:
    - KeyType: HASH
      AttributeName: name

The Deployment
We’ve made it this far defining every resource in a template that we can check in to version control, so we might as well script the deployment as well rather than manually manage the CloudFormation Stack via the AWS web console.

Since I’m using the packaging feature, I first run:

$ aws cloudformation package \
    --template-file template.yaml \
    --s3-bucket <some-bucket-name> \
    --output-template-file template-packaged.yaml
Uploading to 34cd6e82c5e8205f9b35e71afd9e1548 1922559 / 1922559.0 (100.00%) Successfully packaged artifacts and wrote output template to file template-packaged.yaml.

Then to deploy the template (whether new or modified), I run:

$ aws cloudformation deploy \
    --region '<aws-region>' \
    --template-file template-packaged.yaml \
    --stack-name '<some-name>' \
    --capabilities CAPABILITY_IAM
Waiting for changeset to be created.. Waiting for stack create/update to complete Successfully created/updated stack - <some-name>

And that’s it! You’ve just created a dynamic web page that will never require you to SSH anywhere, patch a server, recover from a disaster after Amazon terminates your unhealthy EC2, or any other number of pitfalls that are now the problem of some ops person at AWS. And you can reproduce deployments and make changes with confidence because everything is defined in the template and can be tracked in version control.

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.