Autoscaler parameter autotuning

Amazon EMR 7.2.0 and higher measures the actual required restart time to apply autoscaling decisions. In releases 7.1.0 and lower, you had to use the configuration job.autoscaler.restart.time to manually configure estimated maximum restart time. By using the configuration job.autoscaler.restart.time-tracking.enabled, you only need to enter a restart time for the first scaling. Afterwards, the operator records the actual restart time and will use it for subsequent scalings.

To enable this tracking, use the following command:

job.autoscaler.restart.time-tracking.enabled: true

The following are the related configurations for rescale time estimation.

The following is an example deployment spec you can use to try out rescale time estimation:

apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
  name: autoscaling-example
spec:
  flinkVersion: v1_18
  flinkConfiguration:

    # Autoscaler parameters
    job.autoscaler.enabled: "true"
    job.autoscaler.scaling.enabled: "true"
    job.autoscaler.stabilization.interval: "5s"
    job.autoscaler.metrics.window: "1m"
    
    job.autoscaler.restart.time-tracking.enabled: "true"
    job.autoscaler.restart.time: "2m"
    job.autoscaler.restart.time-tracking.limit: "10m"
    
    jobmanager.scheduler: adaptive
    taskmanager.numberOfTaskSlots: "1"
    pipeline.max-parallelism: "12"

  executionRoleArn: <JOB ARN>
  emrReleaseLabel: emr-7.2.0-flink-latest
  jobManager:
    highAvailabilityEnabled: false
    storageDir: s3://<s3_bucket>/flink/autoscaling/ha/
    replicas: 1
    resource:
      memory: "1024m"
      cpu: 0.5
  taskManager:
    resource:
      memory: "1024m"
      cpu: 0.5
  job:
    jarURI: s3://<s3_bucket>/some-job-with-back-pressure
    parallelism: 1
    upgradeMode: stateless

To simulate backpressure, use the following deployment spec.

job:
    jarURI: s3://<s3_bucket>/pyflink-script.py
    entryClass: "org.apache.flink.client.python.PythonDriver"
    args: ["-py", "/opt/flink/usrlib/pyflink-script.py"]
    parallelism: 1
    upgradeMode: stateless

Upload the following Python script to your S3 bucket.

import logging
import sys
import time
import random

from pyflink.datastream import StreamExecutionEnvironment
from pyflink.table import StreamTableEnvironment

TABLE_NAME="orders"
QUERY=f"""
CREATE TABLE {TABLE_NAME} (
  id INT,
  order_time AS CURRENT_TIMESTAMP,
  WATERMARK FOR order_time AS order_time - INTERVAL '5' SECONDS
)
WITH (
  'connector' = 'datagen',
  'rows-per-second'='10',
  'fields.id.kind'='random',
  'fields.id.min'='1',
  'fields.id.max'='100'
);
"""

def create_backpressure(i):
    time.sleep(2)
    return i

def autoscaling_demo():
    env = StreamExecutionEnvironment.get_execution_environment()
    t_env = StreamTableEnvironment.create(env)
    t_env.execute_sql(QUERY)
    res_table = t_env.from_path(TABLE_NAME)

    stream =  t_env.to_data_stream(res_table) \
      .shuffle().map(lambda x: create_backpressure(x))\
      .print()
    env.execute("Autoscaling demo")

if __name__ == '__main__':
    logging.basicConfig(stream=sys.stdout, level=logging.INFO, format="%(message)s")
    autoscaling_demo()

To verify that rescale time estimation is working, make sure that DEBUG level logging of the Flink operator is enabled. The example below demonstrates how to update the helm chart file values.yaml. Then reinstall the updated helm chart and run your Flink job again.

log4j-operator.properties: |+
  # Flink Operator Logging Overrides
  rootLogger.level = DEBUG

Getthe name of your leader pod.

ip=$(kubectl get configmap -n $NAMESPACE <job-name>-cluster-config-map -o json | jq -r ".data[\"org.apache.flink.k8s.leader.restserver\"]" | awk -F: '{print $2}' | awk -F '/' '{print $3}')

kubectl get pods -n $NAMESPACE -o json | jq -r ".items[] | select(.status.podIP == \"$ip\") | .metadata.name"

Run the following command to get the actual restart time used in metrics evaluations.

kubectl logs <FLINK-OPERATOR-POD-NAME> -c flink-kubernetes-operator -n <OPERATOR-NAMESPACE> -f | grep "Restart time used in scaling summary computation"

You should see logs similar to the following. Note that only the first scaling uses job.autoscaler.restart.time. Subsequent scalings use the observed restart time.

2024-05-16 17:17:32,590 o.a.f.a.ScalingExecutor        [DEBUG][default/autoscaler-example] Restart time used in scaling summary computation: PT2M
2024-05-16 17:19:03,787 o.a.f.a.ScalingExecutor        [DEBUG][default/autoscaler-example] Restart time used in scaling summary computation: PT14S
2024-05-16 17:19:18,976 o.a.f.a.ScalingExecutor        [DEBUG][default/autoscaler-example] Restart time used in scaling summary computation: PT14S
2024-05-16 17:20:50,283 o.a.f.a.ScalingExecutor        [DEBUG][default/autoscaler-example] Restart time used in scaling summary computation: PT14S
2024-05-16 17:22:21,691 o.a.f.a.ScalingExecutor        [DEBUG][default/autoscaler-example] Restart time used in scaling summary computation: PT14S

The open source built-in Flink Autoscaler uses numerous metrics to make the best scaling decisions. However, the default values it uses for its calculations are meant to be applicable to most workloads and might not optimal for a given job. The autotuning feature added into the Amazon EMR on EKS version of the Flink Operator looks at historical trends observed over specific captured metrics and then accordingly tries to calculate the most optimal value tailored for the given job.

To enable autotuning, you must have completed the following:

The following is an example deployment spec you can use to try out autotuning.

apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
  name: autoscaling-example
spec:
  flinkVersion: v1_18
  flinkConfiguration:

    # Autotuning parameters
    kubernetes.operator.job.autoscaler.autotune.enable: "true"
    kubernetes.operator.job.autoscaler.autotune.metrics.history.max.count: "2"
    kubernetes.operator.job.autoscaler.autotune.metrics.restart.count: "1"
    metrics.job.status.enable: TOTAL_TIME

    # Autoscaler parameters
    kubernetes.operator.job.autoscaler.enabled: "true"
    kubernetes.operator.job.autoscaler.scaling.enabled: "true"
    kubernetes.operator.job.autoscaler.stabilization.interval: "5s"
    kubernetes.operator.job.autoscaler.metrics.window: "1m"

    jobmanager.scheduler: adaptive

    taskmanager.numberOfTaskSlots: "1"
    state.savepoints.dir: s3://<S3_bucket>/autoscaling/savepoint/
    state.checkpoints.dir: s3://<S3_bucket>/flink/autoscaling/checkpoint/
    pipeline.max-parallelism: "4"

  executionRoleArn: <JOB ARN>
  emrReleaseLabel: emr-6.14.0-flink-latest
  jobManager:
    highAvailabilityEnabled: true
    storageDir: s3://<S3_bucket>/flink/autoscaling/ha/
    replicas: 1
    resource:
      memory: "1024m"
      cpu: 0.5
  taskManager:
    resource:
      memory: "1024m"
      cpu: 0.5
  job:
    jarURI: s3://<S3_bucket>/some-job-with-back-pressure
    parallelism: 1
    upgradeMode: last-state

To simulate backpressure, use the following deployment spec.

  job:
    jarURI: s3://<S3_bucket>/pyflink-script.py
    entryClass: "org.apache.flink.client.python.PythonDriver"
    args: ["-py", "/opt/flink/usrlib/pyflink-script.py"]
    parallelism: 1
    upgradeMode: last-state

Upload the following Python script to your S3 bucket.

import logging
import sys
import time
import random

from pyflink.datastream import StreamExecutionEnvironment
from pyflink.table import StreamTableEnvironment

TABLE_NAME="orders"
QUERY=f"""
CREATE TABLE {TABLE_NAME} (
  id INT,
  order_time AS CURRENT_TIMESTAMP,
  WATERMARK FOR order_time AS order_time - INTERVAL '5' SECONDS
)
WITH (
  'connector' = 'datagen',
  'rows-per-second'='10',
  'fields.id.kind'='random',
  'fields.id.min'='1',
  'fields.id.max'='100'
);
"""

def create_backpressure(i):
    time.sleep(2)
    return i

def autoscaling_demo():
    env = StreamExecutionEnvironment.get_execution_environment()
    t_env = StreamTableEnvironment.create(env)
    t_env.execute_sql(QUERY)
    res_table = t_env.from_path(TABLE_NAME)

    stream =  t_env.to_data_stream(res_table) \
      .shuffle().map(lambda x: create_backpressure(x))\
      .print()
    env.execute("Autoscaling demo")

if __name__ == '__main__':
    logging.basicConfig(stream=sys.stdout, level=logging.INFO, format="%(message)s")
    autoscaling_demo()

To verify that your autotuner is working, use the following commands. Note that you must use your own leader pod information for the Flink Operator.

First get the name of your leader pod.

ip=$(kubectl get configmap -n $NAMESPACE <job-name>-cluster-config-map -o json | jq -r ".data[\"org.apache.flink.k8s.leader.restserver\"]" | awk -F: '{print $2}' | awk -F '/' '{print $3}')

kubectl get pods -n $NAMESPACE -o json | jq -r ".items[] | select(.status.podIP == \"$ip\") | .metadata.name"

Once you have the name of your leader pod, you can run the following command.

kubectl logs -n $NAMESPACE  -c flink-kubernetes-operator --follow <YOUR-FLINK-OPERATOR-POD-NAME>  | grep -E 'EmrEks|autotun|calculating|restart|autoscaler'

You should see logs similar to the following.

[m[33m2023-09-13 20:10:35,941[m [36mc.a.c.f.k.o.a.EmrEksMetricsAutotuner[m [36m[DEBUG][flink/autoscaling-example] Using the latest Emr Eks Metric for calculating restart.time for autotuning: EmrEksMetrics(restartMetric=RestartMetric(restartingTime=65, numRestarts=1))

[m[33m2023-09-13 20:10:35,941[m [36mc.a.c.f.k.o.a.EmrEksMetricsAutotuner[m [32m[INFO ][flink/autoscaling-example] Calculated average restart.time metric via autotuning to be: PT0.065S