🔍 [REPRO – DO NOT MERGE] reproduce early-resource-timings flake on CI

[thomas-lebeau]

Motivation

The E2E test rum resources › retrieve early requests timings (async / npm / bundle variants at test/e2e/scenario/rum/resources.scenario.ts:61) is one of the top flaky tests this week on webkit-pinned — 28 / 24 / 21 occurrences over the last 7 days across many branches.

Working theory: WebKit clamps PerformanceResourceTiming timestamps and performance.now() to 1ms resolution. When the entire /empty.css request (a zero-byte response) fits inside one 1ms tick, every relevant timestamp clamps to the same value. The SDK's existing Safari fallback in computeResourceEntryDuration (packages/rum-core/src/domain/resource/resourceUtils.ts:76) is guarded by startTime < responseEnd, which fails when they're equal, so the SDK truthfully emits duration: 0 — and the test asserts strictly > 0.

Locally on macOS, 50/50 repeats passed but every captured entry confirmed 1ms clamping. The failure presumably needs CI's faster Linux loopback to make collapse probable. This branch verifies that.

Changes

⚠️ Temporary repro branch — must NOT be merged. Closing once we have data.

• Instrument the early-timings test (test/e2e/scenario/rum/resources.scenario.ts) to log:
  • The raw PerformanceResourceTiming for empty.css (all 12 timestamps + size fields).
  • Four consecutive performance.now() samples to expose clock granularity.
  • The SDK-emitted resource.duration, download, first_byte, size, transfer_size.
• Narrow .gitlab-ci.yml e2e job to:
  • only webkit-pinned (skip chromium/firefox/etc.)
  • only the retrieve early requests timings test via -g
  • --repeat-each=50 × 3 sub-tests = 150 attempts
  • --retries=0 so each repeat is an independent observation

Test instructions

1. Wait for the GitLab e2e: [webkit-pinned] job to finish.
2. In the job log, look for [EARLY-TIMINGS-DEBUG webkit] lines.
3. Expected confirmation of the hypothesis:
   • nowSamples shows 4 identical integer-ms values (clock clamped to 1ms).
   • On failing attempts: startTime === responseStart === responseEnd in raw= and the corresponding sdk= shows duration: 0.
   • On passing attempts: at least one timestamp differs by ≥1ms; SDK shows duration: ≥1ms.
4. Failure count (compared to the local 0/50) tells us whether CI's faster loopback is the differentiator.

Checklist

• Tested locally (50/50 passed — that's the point of the CI run)
• Tested on staging
• Added unit tests for this change.
• Added e2e/integration tests for this change.
• Updated documentation and/or relevant AGENTS.md file

[datadog-prod-us1-6]

🎉 All green!

❄️ No new flaky tests detected
🧪 All tests passed

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 76.96% (+0.00%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 65c7e29 | Docs | Datadog PR Page | Give us feedback!}

[cit-pr-commenter-54b7da]

Bundles Sizes Evolution

📦 Bundle Name	Base Size	Local Size	𝚫	𝚫%	Status
Rum	169.51 KiB	169.51 KiB	0 B	0.00%	✅
Rum Profiler	5.97 KiB	5.97 KiB	0 B	0.00%	✅
Rum Recorder	21.23 KiB	21.23 KiB	0 B	0.00%	✅
Logs	54.70 KiB	54.70 KiB	0 B	0.00%	✅
Rum Slim	127.85 KiB	127.85 KiB	0 B	0.00%	✅
Worker	22.99 KiB	22.99 KiB	0 B	0.00%	✅

🚀 CPU Performance

Action Name	Base CPU Time (ms)	Local CPU Time (ms)	𝚫%
RUM - add global context	0.0019	0.0018	-5.26%
RUM - add action	0.0104	0.0099	-4.81%
RUM - add error	0.0126	0.0092	-26.98%
RUM - add timing	0.0007	0.0004	-42.86%
RUM - start view	0.0109	0.0095	-12.84%
RUM - start/stop session replay recording	0.0008	0.0007	-12.50%
Logs - log message	0.0176	0.0139	-21.02%

🧠 Memory Performance

Action Name	Base Memory Consumption	Local Memory Consumption	𝚫
RUM - add global context	42.69 KiB	38.52 KiB	-4.17 KiB
RUM - add action	67.95 KiB	67.82 KiB	-138 B
RUM - add timing	37.49 KiB	36.91 KiB	-598 B
RUM - add error	73.15 KiB	72.20 KiB	-974 B
RUM - start/stop session replay recording	41.06 KiB	46.22 KiB	+5.16 KiB
RUM - start view	492.59 KiB	488.16 KiB	-4.43 KiB
Logs - log message	55.59 KiB	56.66 KiB	+1.07 KiB

🔗 RealWorld

[thomas-lebeau]

Repro confirmed — see analysis in PR description. Closing without merging; fix follows in a separate PR.

[thomas-lebeau]

Re-running the same repro setup (webkit-pinned + chromium, repeat-each=30, retries=0) with the 50ms /empty.css delay applied to verify the fix lands 0 failures.