Skip to content
Source
Sign inSign up
TanStack/db

Browser SQLite multi-tab oracle: random campaign misses reset recovery reach #2329

Send to my agent
OpenKyleAMathews(GitHub profile, opens in a new tab) opened this issue 0 comments
  • I've validated the bug against the latest version of DB packages

This report concerns a test reliability failure on the exact CI revision below. Published packages and a repair have not been validated.

Describe the bug

The Browser SQLite multi-tab oracle can fail its campaign reach gate because random histories never exercise recovery from a reset-abandoned transaction whose failed commit is first delivered during final catch-up.

Observed on main commit 85392bd1d349aa3bb77c9c11ffba671f0b2d2a84, in Test (sqlite-adapters), run 41453553336.

The failing test is keeps reserved commits lawful across generated multi-tab histories (random or replayed) in packages/browser-db-sqlite-persistence/tests/multi-tab-electric-oracle.test.ts.

browser.multi-tab-reserved-history:
generated histories reached these states fewer than 1 times,
so their laws were not enforced there:
reset-abandoned source recovered during final catch-up (0)

The same campaign reached:

peer reset abandoned an open transaction: 14
reset-abandoned commit put the Collection in error: 5
reset-abandoned source recovered during final catch-up: 0

The property comparisons completed; the failure came from the subsequent generatedReach.assertReached(). The fixed-seed campaign and both fixed regressions for reset-abandoned recovery passed in this job. This failure does not establish a production row, status, or recovery defect.

How to investigate

At the recorded revision, run:

pnpm --filter @tanstack/browser-db-sqlite-persistence exec vitest run \
  tests/multi-tab-electric-oracle.test.ts \
  -t 'keeps reserved commits lawful across generated multi-tab histories'

The random campaign may pass. No failing CI seed/path has been recovered. A local replay attempt did not complete, so the CI log is the executed failure evidence.

Source inspection: the grammar constructs a stop/peer/recovery prefix, then appends an arbitrary tail. It does not construct this reset/final-catch-up interaction. Final catch-up counts recovery only when the abandoned commit first fails there; an earlier failure is restarted before that delivery.

Expected behavior

Normal-budget generated campaigns should reliably reach the declared recovery premise and compare the resulting public and durable observations.

A candidate repair is to construct this interaction with generated variations and nearby controls through the existing model, driver, and comparison. Preserve the 80-run budget, reach minimum, fixed/random parity, and existing fixed regressions. Measure actual reach and demonstrate sensitivity; increasing the budget, deleting the reach, or adding only another fixed example would leave the grammar gap unresolved. This candidate has not been implemented or validated.

Related: #2289 (chance reach misses), #2287 and #2290 (source-health reach/replay reliability). The receiving path is Browser SQLite/Electric multi-tab coordination; no common production cause with its open data-correctness reports is established.

Sign in to join this conversation.