[Feature]: Java Workflow Insight plugin follow-up work #679

Activity

1. wangyb-A

   self-assigned this

   on Sep 2, 2026
2. wangyb-A

   added

   enhancementNew feature or requestNew feature or request

   on Sep 2, 2026
3. github-actions

   added

   parityProvides parity with other language implementations of the SDKProvides parity with other language implementations of the SDK

   pkg:sdkModule: sdkModule: sdk

   on Sep 2, 2026
4. wangyb-A commented on Sep 8, 2026

   

   wangyb-A

   on Sep 8, 2026

   Last edited by wangyb-A

   ContributorAuthor

   More actions

   What would you like?

   Track work deferred from the initial Java Workflow Insight plugin pull request,
   #661.

   Follow-up items

   • Move ON_CHANGE exports off the checkpoint callback thread, add bounded
     coalescing, and prevent a delayed RUNNING export from overwriting the
     terminal object.
     (comment)
   • Preserve Java runtime types when passing values to content transforms.
     (comment)
   • Guarantee that records satisfy the configured size limit after all
     truncation phases. (comment)
   • Distinguish an included JSON null value from an omitted or disabled field.
     (comment)
   • Complete exporter isolation for enum values and mutable map keys.
     (comment)
   • Coordinate sampling normalization across JavaScript, Python, and Java;
     Java should continue matching JavaScript until then.
     (comment)
   • Cloud base integration test
   • Skip input snapshotting when input collection is disabled
     (content.input(false)); cache explicit absence without serializing the value.
     (comment)
   • Validate required exporter configuration, including a non-blank destination
     and positive record-size limit, at build time for S3, CloudWatch Logs, and
     Lambda log exporters.
     (comment)
   • Do not add endTime or durationMs to non-terminal RUNNING records; a
     PENDING or RETRYING invocation boundary is not execution completion.
     (comment)
   • Omit error.message when an exception has no message instead of emitting
     "message": null.
     (comment)
   • Avoid forcing users of the default LambdaLogExporter to include the S3 and
     CloudWatch Logs clients; use optional dependencies or separate exporter
     modules.
     (comment)
   • Avoid re-rendering and serializing the complete record for every truncation
     candidate.
     (comment)
   • Parse durable execution ARNs by structure rather than fixed positions, and
     normalize empty function, qualifier, region, account, and execution names
     consistently.
     (comment)
   • Add public accessors for workflow record input, output, error, end time,
     duration, region, account, and truncation metadata so custom exporters do not
     need to serialize the record to inspect it.
     (comment)
   • Update the plugin README to state that per-invocation state is cleared at every
     invocation end and rebuilt after resume.
     (comment)
   • Move REVIEW_FINDINGS.md out of the Maven-published plugin module; internal
     review-triage notes should remain in the PR or tracking issue.
     (comment)
   • Preserve the Workflow Insight record when an input or output supported by a custom SDK SerDes cannot be copied by the plugin’s Jackson mapper; omit only the unsupported field instead of dropping the entire record.
     (comment)

   Missing exporters

   Java already has Lambda logs, CloudWatch Logs, and S3. Add the remaining
   exporters available in the JavaScript plugin:

   • CloudWatchLogsExporter
   • DynamoDBExporter
   • FirehoseExporter
   • EventBridgeExporter
   • SQSExporter
   • OTelExporter
   • HttpExporter
   • FileExporter
   • OpenSearchExporter
   • RedshiftExporter
   • AuroraExporter

   Links

   • Java plugin PR:
     feat: add Workflow Insight plugin #661
   • Python tracking issue:
     [Feature]: Add workflow insight plugin aws-durable-execution-sdk-python#687
   • JavaScript reference:
     https://github.com/aws/aws-durable-execution-sdk-js/tree/main/packages/aws-durable-execution-sdk-js-insight
5. wangyb-A commented on Sep 9, 2026

   

   wangyb-A

   on Sep 9, 2026

   Last edited by wangyb-A

   ContributorAuthor

   More actions

   Draft implementation for the bounded FIFO ON_CHANGE scheduler: #699
6. wangyb-A commented on Sep 11, 2026

   

   wangyb-A

   on Sep 11, 2026

   Last edited by wangyb-A

   ContributorAuthor

   More actions

   • pending slot coalesces across executions, so one execution's final record can be dropped

   Finding 1 (should fix): the pending slot coalesces across executions, so one execution's final record can be dropped
   
      ExportScheduler has one pending field. schedule() overwrites it unconditionally. The coalescing rule is "a newer record supersedes an older one because each record is a complete snapshot of the execution." That rule holds only for records of
      the same execution. A record for execution B does not contain execution A's information. So when records for two ARNs interleave, pending discards a record that nothing else will deliver.
   
      The end-of-invocation path makes this a data-loss bug, not a lost intermediate:
   
      1. Execution A's onInvocationEnd calls closeAndSchedule, which puts A's SUCCEEDED record in pending.
      2. A's hook thread enters drain() and waits on the in-flight pump.
      3. Execution B's onOperationChange calls schedule(), which replaces pending with B's RUNNING record.
      4. The pump takes pending, exports B's record, sees pending == null, and exits.
      5. A's drain() returns. A's terminal record was never exported and its state is removed from byArn.
   
7. wangyb-A commented on Sep 15, 2026

   

   wangyb-A

   on Sep 15, 2026

   ContributorAuthor

   More actions

   Fix for the cross-execution pending-slot coalescing: #716