[ParidelPooya]

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

Summary

Every Durable Execution Telemetry record timestamp now has one fixed-width UTC format: YYYY-MM-DDTHH:MM:SS.sssZ. The fraction always has 3 digits. The value always ends with the literal Z. This is the format of JavaScript's Date.prototype.toISOString().

Why:

1. We deployed the JS, Python and Java exporters and ran the same workflow through each.
2. Java wrote startTime with 3 fraction digits. In the same record it wrote endTime and emittedAt with 9, for example 2026-10-23T18:42:44.078757096Z.
3. The cause is Instant.toString(). It writes as many fraction digits as the value needs. It writes no fraction when the fraction is zero, for example 2026-10-23T18:42:40Z.
4. JS always writes 3 digits. Python writes 6.
5. DynamoDB, CloudWatch Logs Insights, and string columns compare these values as text.
6. Text order equals time order only when every value has the same width.
7. We measured the effect. …40Z sorts after …40.500Z, because Z sorts after .. The filter endtime >= '2026-10-23T18:42:46Z' excluded all 7 records that ended in that second.
8. So every SDK will write one fixed-width format. JS already writes it. So that format is the one we use.

Changes:

• emittedAt, startTime and endTime on the record, and startTime and endTime on each operation, now go through one helper: Timestamps.format(Instant).
• The helper uses DateTimeFormatter.ofPattern("uuuu-MM-dd'T'HH:mm:ss.SSS'Z'").withZone(ZoneOffset.UTC). Digits below the millisecond are dropped, not rounded. So .078757096 becomes .078, and .999999999 becomes .999.
• TelemetryPlugin produces all record timestamps. No other main code in telemetry-exporter produces a record timestamp string.
• LambdaLogExporter had its own identical formatter for the envelope timestamp. It now calls the same helper. Its output is unchanged.
• The helper must be public, because LambdaLogExporter is in the exporters package. I put it in the existing telemetry.internal package, next to FatalErrors, instead of adding a method to the public Json class. So the change adds no public API outside internal.
• A missing timestamp stays absent, as before.
• The README has a new "Timestamps" section with the format contract. The LambdaLogExporter section now points to it.
• No CHANGELOG exists in the repository, so there is no entry. There is no version bump, because the exporter is unpublished.

Python gets the same change in a sibling PR. JS already writes this format.

Testing

• mvn -o -pl telemetry-exporter -am spotless:apply: no remaining changes.
• mvn -o -pl telemetry-exporter -am verify: BUILD SUCCESS. sdk: 1589 tests, 0 failures, 3 skipped. sdk-testing: 53 tests, 0 failures. telemetry-exporter: 243 tests, 0 failures.
• New RecordTimestampFormatTest:
  • Timestamps.format gives .000Z for zero nanos, .078Z for 078757096 ns, and .999Z for 999999999 ns.
  • A whole second sorts before the same second plus 500 ms when compared as text.
  • A record built by the plugin has emittedAt, startTime, endTime, and each operation's startTime and endTime matching ^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$, with the exact expected values for the three cases above.
• Two existing tests asserted START.toString() (2026-08-05T00:00:00Z) as the record startTime. They now assert 2026-08-05T00:00:00.000Z. Hand-built test records now use the same format as the plugin.

Follow-up: emittedAt is unique within one invocation

emittedAt has millisecond precision. With ON_CHANGE, one invocation can build two records in the same millisecond, for example the last RUNNING record and the terminal record. SQSExporter builds the FIFO deduplication id from executionArn and emittedAt. DynamoDBExporter uses emittedAt as its default sort key. So SQS dropped the second message, and DynamoDB overwrote the first item.

Commit 28fa366 applies the rule of aws/aws-durable-execution-sdk-js#1071. TelemetryPlugin keeps the last emittedAt it assigned, in epoch milliseconds. If the clock is past that value, the clock is used. Otherwise the last value plus 1 ms is used. So emittedAt strictly increases within one plugin instance, and one instance serves one invocation. Hook threads can build records at the same time, so the last value is an AtomicLong updated with updateAndGet(prev -> Math.max(prev + 1, now)).

New EmittedAtOrderingTest: with a frozen clock, the RUNNING, RUNNING and SUCCEEDED records get .123Z, .124Z and .125Z. When the clock moves past the last value, the clock is used. 8 threads × 2000 calls give 16000 distinct values. Without the rule, all three tests fail. verify: telemetry-exporter 246 tests, 0 failures.

[github-actions]

Codex AI review

No actionable findings. The prior same-millisecond emission collision is addressed. Residual test risk: tests were inspected but not executed under the review constraints.

Reviewed commit a5fb2095044a73b3aed106c4602371f04a806b9d. Workflow run