spark-operator VS Hey

I have spent days now trying to figure out a dependency issue I'm experiencing with (Py)Spark running on Kubernetes. I'm using the spark-on-k8s-operator and Spark's Google Cloud connector.

We're using https://github.com/bitnami/charts/tree/main/bitnami/spark, but I have heard good things about https://github.com/GoogleCloudPlatform/spark-on-k8s-operator as well. Hudi should not need any long running deployments as per the docs https://hudi.apache.org/docs/0.5.1/deployment/#deploying

The SparkOperator must be installed before we can use Spark on Kubernetes. Google created this operator, which is available on Github. In a nutshell, the operator is in charge of monitoring the cluster for specific events related to the spark job, as known as kind: SparkApplication

Spark-Operator (as I understand, Spark-Operator integrates well with K8s while vanilla Spark integration with K8s seems extremely complex to maintain)

Spark on K8s Operator is a project from Google that allows submitting spark applications on Kubernetes cluster using CustomResource Definition SparkApplication. It uses mutating admission webhook to modify the pod spec and add the features not officially supported by spark-submit.

In this section, you use Helm to deploy the Kubernetes Operator for Apache Spark from the incubator Chart repository. Helm is a package manager you can use to configure and deploy Kubernetes apps.

Additional details of how SparkApplications are run can be found in the design documentation.

GoogleCloudPlatform/spark-on-k8s-operator (Go): Kubernetes operator for managing the lifecycle of Apache Spark applications on Kubernetes.

The results of the experiment below were based on reproducing approximately 100 cold starts for the duration of our experiment which ran for approximately 1 hour. For it (and all experiments from my previous articles) I used the load test tool hey, but you can use whatever tool you want, like Serverless-artillery or Postman

The results of the experiment to retrieve the existing product from the database by its id see GetProductByIdViaAuroraServerlessV2DataApiHandler with Lambda function with 1024 MB memory setting were based on reproducing more than 100 cold and approximately 10.000 warm starts with experiment which ran for approximately 1 hour. For it (and experiments from my previous article) I used the load test tool hey, but you can use whatever tool you want, like Serverless-artillery or Postman. We won't enable SnapStart on the Lambda function first.

The results of the experiment below were based on reproducing more than 100 cold and approximately 100.000 warm starts with experiment which ran for approximately 1 hour. For it (and experiments from my previous article) I used the load test tool hey, but you can use whatever tool you want, like Serverless-artillery or Postman. I ran all these experiments for all 3 scenarios using 2 different compilation options in template.yaml each:

In our experiment we'll re-use the application introduced in part 9 for this. There are basically 2 Lambda functions which both respond to the API Gateway requests and retrieve product by id received from the API Gateway from DynamoDB. One Lambda function GetProductByIdWithPureJava21Lambda can be used with and without SnapStart and the second one GetProductByIdWithPureJava21LambdaAndPriming uses SnapStart and DynamoDB request invocation priming. We'll measure cold and warm starts using the following memory settings in MBs : 256, 512, 768, 1024, 1536 and 2048. I also put the cold starts measured in the part 12 into the tables to see both cold and warm starts in one place. The results of the experiment below were based on reproducing more than 100 cold and approximately 100.000 warm starts for the duration of our experiment which ran for approximately 1 hour. Here is the code for the sample application. For it (and experiments from my previous article) I used the load test tool hey, but you can use whatever tool you want, like Serverless-artillery or Postman. Abbreviation c is for the cold start and w is for the warm start.

Por último, podemos executar os testes de carga usando hey.

For running our experiments to provoke anomalies we'll use the stress test tool. You can use the tool of your choice (like Gatling, JMeter, Fiddler or Artillery), I personally prefer to use the tool hey as it is easy to use and similar to curl. On Linux this tool can be installed by executing

Hey: is a fast HTTP load testing tool used to test web applications and APIs. It provides a CLI (command-line interface) and supports concurrent requests.

The client and server nodes are CentOS7.9/X86_64. If the HTTP POST requests were sent directly to the server with hey -c 1, there are about 0.2% of cases that may timeout. If the HTTP POST requests were sent through an NGINX proxy on the client node, there are about 20% of cases will timeout. I've confirmed that only one backend node has this problem. All other nodes are 100% succeeded even with higher throughput.

64 concurrent requests isn't a lot. Modern web apps can typically handle much more than that (depending on what the request does, of course). Try it yourself with a load tester like https://github.com/rakyll/hey against a Go HTTP server, for example the one I've built in https://www.golang.dk/articles/go-and-sqlite-in-the-cloud