[MeharPro]

Status: needs owner decision.

I hunted for races in the Prometheus 3.x web UI (web/ui/mantine-ui, prometheus/prometheus main at 4a933f8) and found none.

• Coverage: seven scenarios (query tree, table, graph, format, Targets and Service discovery pool switch, delete double-click). Each ran as an explore session at 300, 600 and 900 ms, plus 8 loads without GenClass where the superseded request was held 900 ms.
• Oracle: the page matched a fresh load of the final state in 224 of 224 page loads (126 of them seeded reorderings).
• Explore's verdicts: it printed RACE 7 times, and all 7 are false reports (explained below).

Separately, I found a real non-race bug: the query tree view sends each leaf node's /api/v1/query twice and aborts the first. There is a two-line fix with a Vitest test. GenClass did not find this as a race, so the owner decides whether it is worth sending as a plain bug fix.

Files

• docs/adoption/hunts/prometheus-mantine-ui.md: the report (setup, scenarios, results table, every explore line classified, why the UI does not race, the duplicate request, GenClass problems, time spent, how to reproduce).
• docs/adoption/hunts/prometheus-mantine-ui.patch: the fix and test against upstream 4a933f89fa96c56ad69e7c02dc5f40e06d93440e, 59 changed lines.
• docs/adoption/hunts/prometheus-mantine-ui-issue.md: the pull request draft in Prometheus' template, plus an optional bug_report.yml issue, the disclosure paragraph and the "before sending" checklist.
• docs/adoption/hunts/prometheus-mantine-ui/: the harness (hunt.mjs, run-all.sh, requests.mjs, summarise.mjs), the hunt's prometheus.yml and rules.yml, results/ (28 files, 384 KB in all, each under 30 KB) and summaries/.
• docs/adoption/hunts/README.md: one new row.

To send (only if the owner says yes)

1. From a personal account, fork prometheus/prometheus.
2. Branch from current main and apply the patch.
3. Reproduce by hand. Either open the tree view with the network panel open and see 7 requests with 2 cancelled, then 5 after the fix; or run requests.mjs against both builds.
4. Run the test both ways and the suite: pnpm --filter @prometheus-io/mantine-ui exec vitest run src/pages/query/TreeNode.test.tsx, then CI=true pnpm run test and pnpm run lint in web/ui.
5. Commit with git commit -s (DCO, the sender's own name).
6. Open the PR titled ui: send each query tree node's request once, using the body in prometheus-mantine-ui-issue.md. It addresses the web/ui owner (juliusv) as CONTRIBUTING.md asks for trivial fixes. No issue is needed first.
7. Before sending, re-check open PR ui: respect evaluation time in Explain and query tree prometheus/prometheus#20279. It changes the same key without fixing the duplicate. If it has merged, the test needs endTime={null}; the patch was checked on its head.
8. If asked how it was found: the browser's request log during the hunt. Explore only listed it as a request count that differs.

Measured

• Duplicate request (requests.mjs, plain Playwright, vite preview production builds, 8 loads each):

  build	requests per load (5 tree nodes)	aborted, all 8 loads
  main	7, in 8 of 8 loads	16
  with the fix	5, in 8 of 8 loads	0

• Test: on main it fails 3 of 3 runs; with the fix it passes 3 of 3. The full UI suite passes (18 files, 208 tests); eslint, prettier and tsc are clean.

• Hunt (seed 1, 6 seeded runs per session, slow_ms 300/600/900):

  scenario	explore verdict	oracle matched reference
  tree, table, graph, format	"no race" at every latency	8 of 8 per session
  delete	"no race" at 300 and 600 ms; RACE 3 of 6 at 900 ms	8 of 8 per session
  Targets, Service discovery	RACE 6 of 6 at every latency	8 of 8 per session
  each scenario without GenClass	(no explore)	8 of 8

• False-report rate on the correct app: all 7 RACE verdicts (42 of 126 seeded runs) are false.

  • Targets and Service discovery (36 runs): a form field named by an id Mantine generates at random on every page load.
  • Delete at 900 ms (3 runs, one verdict): one line of TSDB status page text comes from a tie the server breaks in its top-10 list.

Problems with GenClass itself (report section 9)

1. Form fields are named by their id. Ids generated at random per load (Mantine's useId) cause false RACE verdicts: 6 sessions, 36 of 36 seeded runs. The baselines' noise filter removes exact names, not the shape.
2. A streaming fetch blocks the schedule. The UI's notifications stream (fetch-event-source) holds up delivery for its first 10 s, so 3 to 5 holds run out per load, baselines included (624 in 168 loads). With the stream blocked: 0 in 4 of 4 loads, and each load took about 7 s instead of about 16 s. Suggested: treat text/event-stream responses as delivered once their headers arrive.
3. A server-side difference in one line of page text reads as a race (1 verdict).
4. Docs: the Gitea harness that swarm.md points to uses the older GenClass.init/data-manual injection, while Ghost's uses GENCLASS_CONFIG.

Time

About 2 h 15 min. The 50-minute matrix ran in the background while I worked on the fix, test and drafts.

The branch is claude/bold-faraday-i2fpxs, the session's assigned branch, instead of adopt/prometheus-mantine-ui. Nothing was sent to any other repository.

---

Generated by Claude Code