** This FLIP proposal is a collaborative effort between  Rui Fan and Samrat Deb  , with valuable input and consulting provided by Gyula Fora and Maximilian Michels **

1. Motivation

The proposal to introduce autoscaling for Flink (FLIP-271) has garnered significant interest due to its potential to greatly enhance the usability of Flink. The primary objective is to enable users to effortlessly enable the autoscaler for their Flink jobs without the need for intricate parallelism configurations. However, the current implementation of the flink-autoscaler is tightly coupled with Kubernetes and resides within the flink-kubernetes-operator repository.

There are compelling reasons to extend the use of flink-autoscaler to more types of Flink jobs:

1. With the recent merge of the Externalized Declarative Resource Management (FLIP-291, FLINK-31316), in-place scaling is now supported across all types of Flink jobs. This development has made scaling Flink on YARN a straightforward process.

2. Several discussions within the Flink user community, as observed in the mail list , have emphasized the necessity of flink-autoscaler supporting Flink on YARN.

Hence, this FLIP is centered around the crucial task of decoupling the autoscaler functionality from Kubernetes.
By achieving this decoupling, we aim to empower Flink users to leverage the benefits of autoscaling irrespective of their chosen deployment platform, whether it be Kubernetes or YARN.

2. Core Idea

1. JobAutoScaler and JobAutoScalerImpl: These components will define the generic autoscaling strategy for Flink jobs and are essential for the implementation of the autoscaler module.
   
   
2. Interfaces: The FLIP outlines the necessity of defining a few interfaces - ScalingRealizer, AutoScalerEventHandler and AutoScalerStateStore.
   
   1. The ScalingRealizer interface handles scaling action.
   2. The AutoScalerEventHandler  interface handles loggable events.
   3. The AutoScalerStateStore  interface is responsible for accessing and persisting the autoscaler's state.
      
      
3. Dependencies: The autoscaler module should be agnostic to any specific deployment framework like Kubernetes, YARN etc. Currently, core autoscaling framework closely coupled with Kubernetes-related dependencies such as fabric8, flink-kubernetes-operator, or flink-kubernetes. 
   Ideally, it should rely on Apache Flink project dependencies to gather metrics and make scaling decisions based on JobVertexID , Configuration , MetricGroup , and other relevant classes.
   
   
4. Optional Goal: For flink-autoscaler module, we prefer stay it at flink-k8s-operator first(Advance discussion with Gyula Fora).

       a. Since the autoscaler is not only compatible with the latest Flink version, moving it to the Flink repository may not be the most suitable option.
           It would need to function across various Flink versions, potentially necessitating a separate repository/subproject. 

       b. Removing the autoscaler from the flink-kubernetes-operator repository would introduce significant operational overhead and complicate the release process,
           which would be viewed negatively from the operator's perspective.

   Therefore, considering above two points preference is to retain core autoscaler framework/logic within the flink-kubernetes-operator as a submodule in the short term. Should it prove to be stable in the future,
   we can revisit the idea of relocating it. As it stabilizes, the release frequency will decrease, and this transition will have a minor impact on the operator.

   
   

Note:
Independent flink-kubernetes-operator-autoscaler  module is not necessary. Moving classes to flink-kubernetes-operator  will reduce complexity.

Why flink-kubernetes-operator-autoscaler as seperate module isn't necessary?
     If flink-kubernetes-operator-autoscaler  as an independent module, it must depend on flink-kubernetes-operator  module. flink-kubernetes-operator  cannot depend on
     flink-kubernetes-operator-autoscaler , so it's more difficult to load these classes than remove the flink-kubernetes-operator-autoscaler  module. flink-kubernetes-operator-autoscaler 
     just defines KubernetesScalingRealizer, KubernetesAutoScalerEventHandler and KubernetesAutoScalerStateStore.
     There are not many of these classes, so moving them to flink-kubernetes-operator  will reduce complexity. 

3. Public Interfaces

3.1 JobAutoScaler

The proposed interface maintains a resemblance to the existing JobAutoScaler, albeit with adjustments to its method parameters. In lieu of employing Kubernetes-specific classes, the parameters will be substituted with JobAutoScalerContext<KEY> , rendering it framework-agnostic, suitable for various deployment frameworks beyond Kubernetes.

The 'KEY' in this context corresponds to the 'jobKey', and when two instances share the same 'KEY,' they are regarded as representing the same Flink job.

/** The generic Autoscaler instance. */
public interface JobAutoScaler<KEY> {

    boolean scale(JobAutoScalerContext<KEY> context);

    /** Called when the job is deleted. */
    void cleanup(JobAutoScalerContext<KEY> context);

}

3.2 JobAutoScalerContext

The JobAutoScalerContext  plays a pivotal role in consolidating crucial information necessary for scaling Flink jobs. It encompasses essential data such as the jobKey, jobId, configuration, MetricGroup, and more.

Currently as of now, the jobKey is defined using io.javaoperatorsdk.operator.processing.event.ResourceID , as seen in the existing code and for Kubernetes jobs. However, it's important to note that there may be potential future developments where the jobKey for Flink jobs operating on YARN could differ.

The JobAutoScalerContext , encompassing all pertinent details, will be furnished to these implementations. This comprehensive context will be provided when the autoscaler invokes their respective callbacks.

/**
 * The job autoscaler context.
 * Encompassing all pertinent details, It provides comprehensive context for autoscaler invokes their respective callbacks.
 * @param <KEY> The job key.
 */
 public interface JobAutoScalerContext<KEY> {
 
    // The identifier of each flink job.
    KEY getJobKey();
 
    JobID getJobID();
  
    Configuration getConfiguration();
 
    MetricGroup getMetricGroup();
 
    RestClusterClient<String> getRestClusterClient() throws Exception;

 }

3.3 ScalingRealizer

The ScalingRealizer interface is responsible for managing scaling actions(upscale or downscale), and specifically, the ScalingRealizer#realize  method will be invoked from JobAutoscaler#scale  function.

/**
* The Scaling Realizer. 
* Responsible for managing scaling actions.
* @param <KEY> the job key.
* @param Context Instance of JobAutoScalerContext.
*/
public interface ScalingRealizer<KEY,
       Context extends JobAutoScalerContext<KEY>> {

   void realize(Context context,
                Map<String, String> parallelismOverrides);
}

3.4 AutoScalerEventHandler

The AutoScalerEventHandler interface is responsible for managing loggable events, and specifically, the AutoScalerEventHandler#handleEvent  method will be invoked by the auto scaler when there's a need to handle such events. These events might include scaling errors, reporting scaling results, and more.

It's important to note that the AutoScalerEventHandler object is shared across all Flink jobs and doesn't possess specific job information. That's precisely why the JobAutoScalerContext is passed as a parameter to the handleEvent  method, allowing it to access the necessary job-related details when handling events.

/**
 * Handler all loggable events during scaling.
 *
 * @param <KEY>
 */
public interface AutoScalerEventHandler<KEY,
       Context extends JobAutoScalerContext<KEY>> {

   void handleEvent(Context context,
                    Type type,
                    String reason,
                    String message,
                    @Nullable String messageKey);

   enum Type {
       Normal,
       Warning
   }
}

3.5 AutoScalerStateStore

The AutoScalerStateStore serves the crucial role of persisting and providing access to state information during the scaling process.

In the existing code and for Kubernetes jobs, this state is stored in a ConfigMap . Consequently, the KubernetesAutoScalerStateStore  is responsible for retrieving the ConfigMap before the scaling operation and preserving it after the scaling event.

However, for other types of jobs, such as those running on YARN or in standalone mode, the default behavior involves persisting scaling information in memory itself via introducing new implementation HeapedAutoScalerStateStore . It's important to note that, in the future, there is the possibility to pesist in RDBMS or any persistent storage. It can be new implementation such as JdbcAutoScalerStateStore  etc  to ensure persistent storage of the state.

/** It will be used store state during scaling. */
public interface AutoScalerStateStore<KEY,
       Context extends JobAutoScalerContext<KEY>> {

   void storeScalingHistory(Context jobContext, String decision);

   String getScalingHistory(Context jobContext);

   void removeScalingHistory(Context jobContext);

   void storeEvaluatedMetrics(Context jobContext, String metrics);

   String getEvaluatedMetrics(Context jobContext);

   void removeEvaluatedMetrics(Context jobContext);

   void storeParallelismOverrides(Context jobContext, String parallelismOverrides);

   String getParallelismOverrides(Context jobContext);

   void removeParallelismOverrides(Context jobContext);


   void removeInfoFromCache(Context jobContext);

   // The flush is needed because we just save data in cache for all store methods, and flush these data to the physical storage after the flush is called to improve the performance 
   void flush(Context jobContext);
}

3.6 The generic autoscaler

As discussed with Gyula Fora  and Samrat Deb ,  yarn autoscaler implementation will be out of scope for this FLIP. We will develop a generic autoscaler. This generic autoscaler will not have knowledge of specific jobs, and users will have the flexibility to pass the JobAutoScalerContext when utilizing the generic autoscaler. Communication with Flink jobs can be achieved through the RestClusterClient. 

• The generic ScalingRealizer based on the rescale API (FLIP-291).
• The generic EventHandler based on the logger.
• The generic StateStore based on the Heap. This means that the state information is stored in memory and can be lost if the autoscaler restarts.

4. Proposed Changes

4.1 Ensure new autoscaler module keeps the generic auto scaling strategy.

It includes JobAutoScalerImpl , ScalingMetricCollector , AutoScalerInfo, ScalingExecutor  etc.
kubernetes related dependencies should be removed from these classes and use JobAutoScalerContext , ScalingRealizer , AutoScalerEventHandler  and AutoScalerStateStore  instead.

Using the RestClusterClient  instead of org.apache.flink.kubernetes.operator.service.FlinkService 

• The FlinkService is related to kubernetes, so we shouldn't use it.The RestClusterClient is generic flink client, it supports all flink types, including: kubernetes, yarn, standalone.
  The RestClusterClient<String> is included in JobAutoScalerContext.

4.2 KubernetesJobAutoScalerContext

Note: some code can be extracted into a AbstractJobAutoScalerContext, such as: jobKey, jobId, configuration, metric group and restClieentSupplier.

These logic should be generic for k8s, yarn and standalone.

public class KubernetesJobAutoScalerContext implements JobAutoScalerContext<ResourceID> {

   private final ResourceID resourceID;

   private final JobID jobID;

   private final Configuration configuration;

   private final MetricGroup metricGroup;

   private final SupplierWithException<RestClusterClient<String>, Exception> restClientSupplier;

   private final AbstractFlinkResource<?, ?> resource;

   public KubernetesJobAutoScalerContext(
           ResourceID resourceID,
           JobID jobID,
           Configuration configuration,
           MetricGroup metricGroup,
           SupplierWithException<RestClusterClient<String>, Exception> restClientSupplier,
           AbstractFlinkResource<?, ?> resource) {
       this.resourceID = resourceID;
       this.jobID = jobID;
       this.configuration = configuration;
       this.metricGroup = metricGroup;
       this.restClientSupplier = restClientSupplier;
       this.resource = resource;
   }

   @Override
   public ResourceID getJobKey() {
       return resourceID;
   }

   @Override
   public JobID getJobID() {
       return jobID;
   }

   @Override
   public Configuration getConfiguration() {
       return configuration;
   }

   @Override
   public MetricGroup getMetricGroup() {
       return metricGroup;
   }

   @Override
   public RestClusterClient<String> getRestClusterClient() throws Exception {
       return restClientSupplier.get();
   }

   public AbstractFlinkResource<?, ?> getResource() {
       return resource;
   }
}

4.3 KubernetesScalingRealizer

public class KubernetesScalingRealizer
       implements ScalingRealizer<ResourceID, KubernetesJobAutoScalerContext> {
   @Override
   public void realize(
           KubernetesJobAutoScalerContext context,
           Map<String, String> parallelismOverrides) {
       spec.getFlinkConfiguration().put(
       PipelineOptions.PARALLELISM_OVERRIDES.key(),
       ConfigurationUtils.convertValue(parallelismOverrides, String.class));
       context.getResource().getStatus().setImmediateReconciliationNeeded(true);
   }
}

4.4 KubernetesAutoScalerEventHandler

public class KubernetesAutoScalerEventHandler
       implements AutoScalerEventHandler<ResourceID, KubernetesJobAutoScalerContext> {

   private final EventRecorder eventRecorder;

   public KubernetesAutoScalerEventHandler(EventRecorder eventRecorder) {
       this.eventRecorder = eventRecorder;
   }

   @Override
   public void handleEvent(KubernetesJobAutoScalerContext context,
                           Type type,
                           String reason,
                           String message,
                           @Nullable String messageKey) {
       eventRecorder.triggerEvent(
               context.getResource(),
               EventRecorder.Type.valueOf(type.name()),
               reason,
               EventRecorder.Component.Operator,
               message,
               messageKey);
   }

}

4.5 KubernetesAutoScalerStateStore

For current code or kubernetes job, the state is persisted to ConfigMap. So the KubernetesAutoScalerStateStore needs to fetch ConfigMap before scaling, and persist the ConfigMap after scaling.

/** The kubernetes auto scaler state store, it's based on the config map. */ 
public class KubernetesAutoScalerStateStore
       implements AutoScalerStateStore<ResourceID, KubernetesJobAutoScalerContext> {

   private static final Logger LOG = LoggerFactory.getLogger(KubernetesAutoScalerStateStore.class);

   private static final String LABEL_COMPONENT_AUTOSCALER = "autoscaler";
   private static final String COLLECTED_METRICS_KEY = "collectedMetrics";
   private static final String SCALING_HISTORY_KEY = "scalingHistory";
   private static final String PARALLELISM_OVERRIDES_KEY = "parallelismOverrides";

   private final KubernetesClient kubernetesClient;

   private final ConcurrentHashMap<ResourceID, ConfigMap> cache = new ConcurrentHashMap<>();

   public KubernetesAutoScalerStateStore(KubernetesClient kubernetesClient) {
       this.kubernetesClient = kubernetesClient;
   }

   // COLLECTED_METRICS_KEY and PARALLELISM_OVERRIDES_KEY is similar to SCALING_HISTORY_KEY
   @Override
   public void storeScalingHistory(KubernetesJobAutoScalerContext jobContext, String decision) {
     getState(jobContext).put(SCALING_HISTORY_KEY, decision);
   }

   @Override
   public String getScalingHistory(KubernetesJobAutoScalerContext jobContext) {
    getState(jobContext).get(SCALING_HISTORY_KEY);
   }

   @Override
   public void removeScalingHistory(KubernetesJobAutoScalerContext jobContext) {
      getState(jobContext).remove(SCALING_HISTORY_KEY);
   }

   private Map<String, String> getState(KubernetesJobAutoScalerContext jobContext) {
       return cache.computeIfAbsent(jobContext.getJobKey(), jobKey -> getOrCreateConfigMap(jobContext)).getData();
   }

   @Override
   public void flush(KubernetesJobAutoScalerContext jobContext) {
       ConfigMap configMap = cache.get(jobContext.getJobKey());
       Preconditions.checkState(configMap != null, "The configMap shouldn't be null.");
       try {
           cache.put(jobContext.getJobKey(), kubernetesClient.resource(configMap).update());
       } catch (Exception e) {
           LOG.error(
                   "Error while updating autoscaler info configmap, invalidating to clear the cache",
                   e);
           removeInfoFromCache(jobContext);
           throw e;
       }
   }

   @Override
   public void removeInfoFromCache(KubernetesJobAutoScalerContext jobContext) {
       cache.remove(jobContext.getJobKey());
   }

   private ConfigMap getOrCreateConfigMap(KubernetesJobAutoScalerContext jobContext) {
       AbstractFlinkResource<?, ?> cr = jobContext.getResource();
       var meta = createCmObjectMeta(ResourceID.fromResource(cr));
       return getScalingInfoConfigMap(meta).orElseGet(() -> createConfigMap(cr, meta));
   }

   private ConfigMap createConfigMap(HasMetadata cr, ObjectMeta meta) {
       LOG.info("Creating scaling info config map");
       var cm = new ConfigMap();
       cm.setMetadata(meta);
       cm.addOwnerReference(cr);
       cm.setData(new HashMap<>());
       return kubernetesClient.resource(cm).create();
   }

   private ObjectMeta createCmObjectMeta(ResourceID uid) {
       var objectMeta = new ObjectMeta();
       objectMeta.setName("autoscaler-" + uid.getName());
       uid.getNamespace().ifPresent(objectMeta::setNamespace);
       objectMeta.setLabels(
               Map.of(
                       Constants.LABEL_COMPONENT_KEY,
                       LABEL_COMPONENT_AUTOSCALER,
                       Constants.LABEL_APP_KEY,
                       uid.getName()));
       return objectMeta;
   }

   private Optional<ConfigMap> getScalingInfoConfigMap(ObjectMeta objectMeta) {
       return Optional.ofNullable(
               kubernetesClient
                       .configMaps()
                       .inNamespace(objectMeta.getNamespace())
                       .withName(objectMeta.getName())
                       .get());
   }
}

5. Compatibility, Deprecation, and Migration Plan

It must be compatible with current kubernetes operator.

6. Test Plan

UT & IT & Manually verify that the autoscaler is working as expected.

7. Rejected Alternatives

Rejected a couple of interface designs, the the whole solution is fine.