Deep dive into the Amazon Managed Service for Apache Flink utility lifecycle – Half 2

In Half 1 of this sequence, we mentioned basic operations to manage the lifecycle of your Amazon Managed Service for Apache Flink utility. If you’re utilizing higher-level instruments comparable to AWS CloudFormation or Terraform, the instrument will execute these operations for you. Nevertheless, understanding the basic operations and what the service routinely does can present some stage of Mechanical Sympathy to confidently implement a extra strong automation.

Within the first a part of this sequence, we targeted on the pleased paths. In a great world, failures don’t occur, and each change you deploy works completely. Nevertheless, the true world is much less predictable. Quoting Werner Vogels, Amazon’s CTO, “The whole lot fails, on a regular basis.”

On this put up, we discover failure eventualities that may occur throughout regular operations or once you deploy a change or scale the applying, and monitor operations to detect and get better when one thing goes improper.

The much less pleased path

A strong automation should be designed to deal with failure eventualities, specifically throughout operations. To try this, we have to perceive how Apache Flink can deviate from the pleased path. Because of the nature of Flink as a stateful stream processing engine, detecting and resolving failure eventualities requires completely different methods in comparison with different long-running purposes, comparable to microservices or short-lived serverless capabilities (comparable to AWS Lambda).

Flink’s conduct on runtime errors: The fail-and-restart loop

When a Flink job encounters an surprising error at runtime (an unhandled exception), the conventional conduct is to fail, cease the processing, and restart from the newest checkpoint. Checkpoints permit Flink to help knowledge consistency and no knowledge loss in case of failure. Additionally, as a result of Flink is designed for stream processing purposes, which run repeatedly, if the error occurs once more, the default conduct is to maintain restarting, hoping the issue is transient and the applying will finally get better the conventional processing.In some circumstances, the issue isn’t transient, nevertheless. For instance, once you deploy a code change that comprises a bug, inflicting the job to fail as quickly because it begins processing knowledge, or if the anticipated schema doesn’t match the information within the supply, inflicting deserialization or processing errors. The identical state of affairs may additionally occur in the event you mistakenly modified a configuration that stops a connector to succeed in the exterior system. In these circumstances, the job is caught in a fail-and-restart loop, indefinitely, or till you actively force-stop it.

When this occurs, the Managed Service for Apache Flink utility standing could be RUNNING, however the underlying Flink job is definitely failing and restarting. The AWS Administration Console provides you a touch, pointing that the applying may want consideration (see the next screenshot).

Within the following sections, we learn to monitor the applying and job standing, to routinely react to this case.

When beginning or updating the applying goes improper

To know the failure mode, let’s evaluation what occurs routinely once you begin the applying, or when the applying restarts after you issued UpdateApplication command, as we explored in Half 1 of this sequence. The next diagram illustrates what occurs when an utility begins.

The workflow consists of the next steps:

1. Managed Service for Apache Flink provisions a cluster devoted to your utility.
2. The code and configuration are submitted to the Job Supervisor node.
3. The code within the predominant() methodology of your utility runs, defining the dataflow of your utility.
4. Flink deploys to the Activity Supervisor nodes the substasks that make up your job.
5. The job and utility standing change to RUNNING. Nevertheless, subtasks begin initializing now.
6. Subtasks restore their state, if relevant, and initialize any sources. For instance, a Kafka connector’s subtask initializes the Kafka consumer and subscribes the subject.
7. When all subtasks are efficiently initialized, they modify to RUNNING standing and the job begins processing knowledge.

To new Flink customers, it may be complicated {that a} RUNNING standing doesn’t essentially indicate the job is wholesome and processing knowledge.When one thing goes improper in the course of the strategy of beginning (or restarting) the applying, relying on the section when the issue arises, you may observe two several types of failure modes:

• (a) An issue prevents the applying code from being deployed – Your utility may encounter this failure state of affairs if the deployment fails as quickly because the code and configuration are handed to the Job Supervisor (step 2 of the method), for instance if the applying code package deal is malformed. A typical error is when the JAR is lacking a mainClass or if mainClass factors to a category that doesn’t exist. This failure mode may additionally occur if the code of your predominant() methodology throws an unhandled exception (step 3). In these circumstances, the applying fails to alter to RUNNING, and reverts to READY after the try.
• (b) The applying is began, the job is caught in a fail-and-restart loop – An issue may happen later within the course of, after the applying standing has modified RUNNING. For instance, after the Flink job has been deployed to the cluster (step 4 of the method), a part may fail to initialize (step 6). This may occur when a connector is misconfigured, or an issue prevents it from connecting to the exterior system. For instance, a Kafka connector may fail to hook up with the Kafka cluster due to the connector’s misconfiguration or networking points. One other attainable state of affairs is when the Flink job efficiently initializes, however it throws an exception as quickly because it begins processing knowledge (step 7). When this occurs, Flink reacts to a runtime error and may get caught in a fail-and-restart loop.

The next diagram illustrates the sequence of utility standing, together with the 2 failure eventualities simply described.

Troubleshooting

Now we have examined what can go improper throughout operations, specifically once you replace a RUNNING utility or restart an utility after altering its configuration. On this part, we discover how we will act on these failure eventualities.

Roll again a change

If you deploy a change and understand one thing isn’t fairly proper, you usually wish to roll again the change and put the applying again in working order, till you examine and repair the issue. Managed Service for Apache Flink supplies a swish strategy to revert (roll again) a change, additionally restarting the processing from the purpose it was stopped earlier than making use of the fault change, offering consistency and no knowledge loss.In Managed Service for Apache Flink, there are two varieties of rollbacks:

• Automated – Throughout an computerized rollback (additionally referred to as system rollback), if enabled, the service routinely detects when the applying fails to restart after a change, or when the job begins however instantly falls right into a fail-and-restart loop. In these conditions, the rollback course of routinely restores the applying configuration model earlier than the final change was utilized and restarts the applying from the snapshot taken when the change was deployed. See Enhance the resilience of Amazon Managed Service for Apache Flink utility with system-rollback characteristic for extra particulars. This characteristic is disabled by default. You may allow it as a part of the applying configuration.
• Guide – A handbook rollback API operation is sort of a system rollback, however it’s initiated by the person. If the applying is working however you observe one thing not behaving as anticipated after making use of a change, you may set off the rollback operation utilizing the RollbackApplication API motion or the console. Guide rollback is feasible when the applying is RUNNING or UPDATING.

Each rollbacks work equally, restoring the configuration model earlier than the change and restarting with the snapshot taken earlier than the change. This prevents knowledge loss and brings you again to a model of the applying that was working. Additionally, this makes use of the code package deal that was saved on the time you created the earlier configuration model (the one you’re rolling again to), so there is no such thing as a inconsistency between code, configuration, and snapshot, even when within the meantime you may have changed or deleted the code package deal from the Amazon Easy Storage Service (Amazon S3) bucket.

Implicit rollback: Replace with an older configuration

A 3rd strategy to roll again a change is to easily replace the configuration, bringing it again to what it was earlier than the final change. This creates a brand new configuration model, and requires the proper model of the code package deal to be out there within the S3 bucket once you situation the UpdateApplication command.

Why is there a 3rd possibility when the service supplies system rollback and the managed RollbackApplication motion? As a result of most high-level infrastructure-as-code (IaC) frameworks comparable to Terraform use this technique, explicitly overwriting the configuration. You will need to perceive this risk despite the fact that you’ll most likely use the managed rollback in the event you implement your automation based mostly on the low-level actions.

The next are two necessary caveats to think about for this implicit rollback:

• You’ll usually wish to restart the applying from the snapshot that was taken earlier than the defective change was deployed. If the applying is at present RUNNING and wholesome, this isn’t the newest snapshot (RESTORE_FROM_LATEST_SNAPSHOT), however moderately the earlier one. It’s essential to set the restart from RESTORE_FROM_CUSTOM_SNAPSHOT and choose the proper snapshot.
• UpdateApplication solely works if the applying is RUNNING and wholesome, and the job might be gracefully stopped with a snapshot. Conversely, if the applying is caught in a fail-and-restart loop, you could force-stop it first, change the configuration whereas the applying is READY, and later begin the applying from the snapshot that was taken earlier than the defective change was deployed.

Pressure-stop the applying

In regular eventualities, you cease the applying gracefully, with the automated snapshot creation. Nevertheless, this may not be attainable in some eventualities, comparable to if the Flink job is caught in a fail-and-restart loop. This may occur, for instance, if an exterior system the job makes use of stops working, or as a result of the AWS Id and Entry Administration (IAM) configuration was erroneously modified, eradicating permissions required by the job.

When the Flink job will get caught in a fail-and-restart loop after a defective change, your first possibility must be utilizing RollbackApplication, which routinely restores the earlier configuration and begins from the proper snapshot. Within the uncommon circumstances you may’t cease the applying gracefully or use RollbackApplication, the final resort is force-stopping the applying. Pressure-stop makes use of the StopApplication command with Pressure=true. You can too force-stop the applying from the console.

If you force-stop an utility, no snapshot is taken (if that have been attainable, you’ll have been in a position to gracefully cease). If you restart the applying, you may both skip restoring from a snapshot (SKIP_RESTORE_FROM_SNAPSHOT) or use a snapshot that was beforehand taken, scheduled utilizing Snapshot Supervisor, or manually, utilizing the console or CreateApplicationSnapshot API motion.

We strongly suggest organising scheduled snapshots for all manufacturing purposes that you would be able to’t afford restarting with no state.

Monitoring Apache Flink utility operations

Efficient monitoring of your Apache Flink purposes throughout and after operations is essential to confirm the end result of the operation and permit lifecycle automation to lift alarms or react, in case one thing goes improper.

The principle indicators you need to use throughout operations embrace the FullRestarts metric (out there in Amazon CloudWatch) and the applying, job, and process standing.

Monitoring the end result of an operation

The only strategy to detect the end result of an operation, comparable to StartApplication or UpdateApplication, is to make use of the ListApplicationOperations API command. This command returns a listing of the newest operations of a particular utility, together with upkeep occasions that drive an utility restart.

For instance, to retrieve the standing of the newest operation, you need to use the next command:

aws kinesisanalyticsv2 list-application-operations 
    --application-name MyApplication 
   | jq '.ApplicationOperationInfoList 
   | sort_by(.StartTime) | final'

The output will likely be just like the next code:

{
  "Operation": "UpdateApplication",
  "OperationId": "12abCDeGghIlM",
  "StartTime": "2025-08-06T09:24:22+01:00",
  "EndTime": "2025-08-06T09:26:56+01:00",
  "OperationStatus": "IN_PROGRESS"
}

OperationStatus will comply with the identical logic as the applying standing reported by the console and by DescribeApplication. This implies it may not detect a failure in the course of the operator initialization or whereas the job begins processing knowledge. As we’ve got realized, these failures may put the applying in a fail-and-restart loop. To detect these eventualities utilizing your automation, you could use different methods, which we cowl in the remainder of this part.

Detecting the fail-and-restart loop utilizing the FullRestarts metric

The only strategy to detect whether or not the applying is caught in a fail-and-restart loop is utilizing the fullRestarts metric, out there in CloudWatch Metrics. This metric counts the variety of restarts of the Flink job after you began the applying with a StartApplication command or restarted with UpdateApplication.

In a wholesome utility, the variety of full restarts ought to ideally be zero. A single full restart could be acceptable throughout deployment or deliberate upkeep; a number of restarts usually point out some situation. We suggest to not set off an alarm on a single restart, and even a few consecutive restarts.

The alarm ought to solely be triggered when the applying is caught in a fail-and-restart loop. This means checking whether or not a number of restarts have occurred over a comparatively brief time period. Deciding the interval isn’t trivial, as a result of the time the Flink job takes to restart from a checkpoint relies on the dimensions of the applying state. Nevertheless, if the state of your utility is decrease than a number of GB per KPU, you may safely assume the applying ought to begin in lower than a minute.

The purpose is making a CloudWatch alarm that triggers when fullRestarts retains rising over a time interval ample for a number of restarts. For instance, assuming your utility restarts in lower than 1 minute, you may create a CloudWatch alarm that depends on the DIFF math expression of the fullRestarts metric. The next screenshot reveals an instance of the alarm particulars.

This instance is a conservative alarm, solely triggering if the applying retains restarting for over 5 minutes. This implies you detect the issue after at the very least 5 minutes. You may think about decreasing the time to detect the failure earlier. Nevertheless, watch out to not set off an alarm after only one or two restarts. Occasional restarts may occur, for instance throughout regular upkeep (patching) that’s managed by the service, or for a transient error of an exterior system. Flink is designed to get better from these situations with minimal downtime and no knowledge loss.

Detecting whether or not the job is up and working: Monitoring utility, job, and process standing

Now we have mentioned how you may have completely different statuses: the standing of the applying, job, and subtask. In Managed Service for Apache Flink, the applying and job standing change to RUNNING when the subtasks are efficiently deployed on the cluster. Nevertheless, the job isn’t actually working and processing knowledge till all of the subtasks are RUNNING.

Observing the applying standing throughout operations

The applying standing is seen on the console, as proven within the following screenshot.

In your automation, you may ballot the DescribeApplication API motion to look at the applying standing. The next command reveals use the AWS Command Line Interface (AWS CLI) and jq command to extract the standing string of an utility:

aws kinesisanalyticsv2 describe-application  
    --application-name  
    | jq -r '.ApplicationDetail.ApplicationStatus'

Observing job and subtask standing

Managed Service for Apache Flink provides you entry to the Flink Dashboard, which supplies helpful info for troubleshooting, together with the standing of all subtasks. The next screenshot, for instance, reveals a wholesome job the place all subtasks are RUNNING.

Within the following screenshot, we will see a job the place subtasks are failing and restarting.

In your automation, once you begin the applying or deploy a change, you wish to make sure the job is finally up and working and processing knowledge. This occurs when all of the subtasks are RUNNING. Be aware that ready for the job standing to turn out to be RUNNING after an operation isn’t utterly secure. A subtask may nonetheless fail and trigger the job to restart after it was reported as RUNNING.

After you execute a lifecycle operation, your automation can ballot the substasks standing ready for considered one of two occasions:

• All subtasks report RUNNING – This means the operation was profitable and your Flink job is up and working.
• Any subtask stories FAILING or CANCELED – This means one thing went improper, and the applying is probably going caught in a fail-and-restart loop. You should intervene, for instance, force-stopping the applying after which rolling again the change.

If you’re restarting from a snapshot and the state of your utility is sort of huge, you may observe subtasks will report INITIALIZING standing for longer. In the course of the initialization, Flink restores the state of the operator earlier than altering to RUNNING.

The Flink REST API exposes the state of the subtasks, and can be utilized in your automation. In Managed Service for Apache Flink, this requires three steps:

1. Generate a pre-signed URL to entry the Flink REST API utilizing the CreateApplicationPresignedUrl API motion.
2. Make a GET request to the /jobs endpoint of the Flink REST API to retrieve the job ID.
3. Make a GET request to the /jobs/ endpoint to retrieve the standing of the subtasks.

The next GitHub repository supplies a shell script to retrieve the standing of the duties of a given Managed Service for Apache Flink utility.

Monitoring subtasks failure whereas the job is working

The strategy of polling the Flink REST API can be utilized in your automation, instantly after an operation, to look at whether or not the operation was finally profitable.

We strongly suggest to not repeatedly ballot the Flink REST API whereas the job is working to detect failures. This operation is useful resource consuming, and may degrade efficiency or trigger errors.

To watch for suspicious subtask standing adjustments throughout regular operations, we suggest utilizing CloudWatch Logs as an alternative. The next CloudWatch Logs Insights question extracts all subtask state transitions:

fields , message
| parse message /^(?.+) switched from (?[A-Z]+) to (?[A-Z]+)./
| filter ispresent(process) and ispresent(fromStatus) and ispresent(toStatus)
| show , process, fromStatus, toStatus
| restrict 10000

How Managed Service for Apache Flink minimizes processing downtime

Now we have seen how Flink is designed for robust consistency. To ensure exactly-once state consistency, Flink quickly stops the processing to deploy any adjustments, together with scaling. This downtime is required for Flink to take a constant copy of the applying state and reserve it in a savepoint. After the change is deployed, the job is restarted from the savepoint, and there’s no knowledge loss. In Managed Service for Apache Flink, updates are totally managed. When snapshots are enabled, UpdateApplication routinely stops the job and makes use of snapshots (based mostly on Flink’s savepoints) to retain the state.

Flink ensures no knowledge loss. Nevertheless, your corporation necessities or Service Degree Goals (SLOs) may additionally impose a most delay for the information obtained by downstream methods, or end-to-end latency. This delay is affected by the processing downtime, or the time the job doesn’t course of knowledge to permit Flink deploying the change.With Flink, some processing downtime is unavoidable. Nevertheless, Managed Service for Apache Flink is designed to attenuate the processing downtime once you deploy a change.

Now we have seen how the service runs your utility in a devoted cluster, for full isolation. If you situation UpdateApplication on a RUNNING utility, the service prepares a brand new cluster with the required quantity of sources. This operation may take a while. Nevertheless, this doesn’t have an effect on the processing downtime, as a result of the service retains the job working and processing knowledge on the unique cluster till the final attainable second, when the brand new cluster is prepared. At this level, the service stops your job with a savepoint and restarts it on the brand new cluster.

Throughout this operation, you’re solely charged for the variety of KPU of a single cluster.

The next diagram illustrates the distinction between the length of the replace operation, or the time the applying standing is UPDATING, and the processing downtime, observable from the job standing, seen within the Flink Dashboard.

You may observe this course of, maintaining each the applying console and Flink Dashboard open, once you replace the configuration of a working utility, even with no adjustments. The Flink Dashboard will turn out to be quickly unavailable when the service switches to the brand new cluster. Moreover, you may’t use the script we offered to examine the job standing for this scope. Though the cluster retains serving the Flink Dashboard till it’s tore down, the CreateApplicationPresignedUrl motion doesn’t work whereas the applying is UPDATING.

The processing time (the time the job isn’t working on both clusters) relies on the time the job takes to cease with a savepoint (snapshot) and restore the state within the new cluster. This time largely relies on the dimensions of the applying state. Knowledge skew may additionally have an effect on the savepoint time because of the barrier alignment mechanism. For a deep dive into the Flink’s barrier alignment mechanism, confer with Optimize checkpointing in your Amazon Managed Service for Apache Flink purposes with buffer debloating and unaligned checkpoints, maintaining in thoughts that savepoints are all the time aligned.

For the scope of your automation, you usually wish to wait till the job is again up and working and processing knowledge. You usually wish to set a timeout. If each the applying and job don’t return to RUNNING inside this timeout, one thing most likely went improper and also you may wish to increase an alarm or drive a rollback. This timeout ought to think about the whole replace operation length.

Conclusion

On this put up, we mentioned attainable failure eventualities once you deploy a change or scale your utility. We confirmed how Managed Service for Apache Flink rollback functionalities can seamlessly convey you again to a secure place after a change went improper. We additionally explored how one can automate monitoring operations to look at utility, job, and subtask standing, and use the fullRestarts metric to detect when the job is in a fail-and-restart loop.

For extra info, see Run a Managed Service for Apache Flink utility, Implement fault tolerance in Managed Service for Apache Flink, and Handle utility backups utilizing Snapshots.

Concerning the authors