[kriszyp]

⊙ Problem

Refs #1398.

Concurrent reads of one key on a sourcedFrom table coalesce only when the first read stores a record. When the source reports the record missing, or the read uses noCacheStore, 20 concurrent reads make 20 origin calls, one after another.

🤔 Is the requirement right?

Yes, with one exception. A response is not shared if its source set noCacheStore on its own context. That marks the response as per-request.

💡 Solution

• Waiters on the same thread now take the first read's result, "not found" included, when the store has nothing valid.
• Failed fills are not shared; each waiter retries as before.
• A shared record must be plain data. Records with a stream, iterator, Headers or blob refetch: a response consumes or changes those.
• An unstored noCacheStore record goes only to waiters that also asked for noCacheStore.
• Waiters on other worker threads still refetch in turn.

🤔 Should concurrent no-store requests share one source answer?

Yes, as the task asked. A caller-dependent source must set noCacheStore itself; the cache already shares its other answers. Reverting is one condition.

⚖️ Alternatives

• Share the leader's promise: resolves before the commit and leaks the leader's stale-on-error answers.
• Cache a short-lived "not found" entry: changes caching policy and does nothing for noCacheStore.
• Fresh decode per waiter: costs an encode on every shared fill.

🔧 Changes

• resources/Table.ts: a per-table sourceFills map of this thread's in-flight fills, keyed by writeKeyId so 0 and -0 stay apart. Each caller captures the in-flight fill before tryLock. The leader registers its fill in the same synchronous run as its successful tryLock. The waiter callback still reads the store first, then falls back to the fill's outcome. ⚠️ Look hardest: what is shared. The fill is kept unless the source opted out. The write counts as stored only once staged and kept by its commit callback. The first waiter that wants the outcome runs a bounded plain-data scan once.
• resources/DESIGN.md, DESIGN.md: the sharing rules and their index entry.

✅ Verification

• New unit suite, RocksDB and LMDB: 20 cases pass, including the shapes the scan refuses and a stale entry behind a no-store leader. On base the negative-miss, noCacheStore and 0/-0 cases fail (20 calls where 1 is expected).
• New one-worker REST test (source fixture, schema, config): 20 concurrent 404s and 20 no-store reads each make one origin request (base: 20).
• Rebased onto main at ec7b755: build, typecheck, lint, the new suites, caching and Resource-get-context pass.

Suite numbers

Before the rebase, test:unit:resources 4519 passing, 2 failing; test:unit:main 7425 passing, 2 failing (all four pass alone); test:integration:all 2558 passing, 12 cancelled (6 Ollama models missing locally, 6 in sql-not-predicate.test.ts, which passes alone). check:design-docs fails on main too: resources/DESIGN.md is over its 1000-line budget, and this PR adds 4 lines. Tests hold each burst's first source call until the other 19 reads park on the record lock.

Unit cases

absent record → 1 call; no negative caching across bursts; noCacheStore → 1 call, nothing stored; no-store leader with a caching waiter → the waiter caches; source-set noCacheStore → 20 calls, writable and read-only; cached record → 1 call; failed call → retried, failure not shared; failure then absence → 2 calls; failure while preparing the record → not shared; Headers, stream body, passthrough stream, iterator, uncopyable record, symbol/hidden/hole/getter/dead-proxy shapes, over-budget record → not shared; nested plain data with a text body → shared; stale entry behind a no-store leader → caching waiter refetches; 0/-0 kept apart.

Framing-Verdict: better-alternative-exists (23a399dda158) — adopted: the existing lock waiter callback carries the fill.

PR QA: skipped @ 88c71ad — dispatch unavailable (no worker token in this session).

Docs: no companion needed; no public API or configuration change.

_{🤖 Generated with Claude Code (Claude Opus 5.5; rebased and updated by Claude Sonnet 5.5); posted via @kriszyp.}

https://claude.ai/code/session_01T9ub6mcFSCB8x9He8FS35Z

Release target: main — a performance fix with no known backport need; it changes shared source-fill behavior

Related PRs: #3490 overlaps (same lock-release and waiter code in getFromSource), 9 others independent
Complexity: medium

Deferred findings

• a44a4a185298b74f nit correctness — resources/Table.ts — outcome served when stored record later invalidated, evicted or expired — NOT FILED
• ec56d2c461bfc09c nit documentation — resources/Table.ts — suggest comment that sourceFills is thread-local — NOT FILED
• c4640d13051102da nit test-coverage — unitTests/resources/sourceFillCoalescing.test.js — source-identity guard and mixed-policy burst untested — NOT FILED
• 996d31ce76294bfa nit security — resources/Table.ts — Concurrent no-store waiters receive the leader's unstored record without comparing requester identity — NOT FILED
• bb1bbbbb0ec64391 nit reliability — resources/Table.ts — Synchronous throw between sourceFills.set and when(commitPromise) attachment orphans the fill map entry — NOT FILED
• 1f2fab373441427f nit maintainability — integrationTests/resources/sourcedfrom-negative-miss-coalescing.test.ts — '; runtime resolves fine' in the ts-expect-error directive adds no constraint — NOT FILED
• 1b3008c1343391d1 nit performance — resources/Table.ts — unstoredRecord counts a tombstone existingEntry, forcing caching waiters behind a no-store leader into one extra origin call — NOT FILED
• 158e3ef374d7c494 nit test-coverage — unitTests/resources/sourceFillCoalescing.test.js — unit retry leaders rely on a fixed delay(10) to outlast re-parking waiters — NOT FILED
• becf8ece6a59110d nit correctness — resources/Table.ts — waiters share the leader's live source object; Date and ArrayBuffer-view leaves skip own-property inspection, admitting an own one-shot iter — NOT FILED
• 183bc5e8f8c1978e nit documentation — resources/DESIGN.md — note leaves resources/DESIGN.md at 998 of 1000 lines without pruning — NOT FILED

_{Dispatch: task harper-sourcedfrom-negative-miss-coalescing · queued by unknown · ran by claude/opus/high · worker kzyp-xps-1}

_{Review-Coverage: authored=claude; ran=gemini,codex,cursor-composer,cursor-muse; adjudicated=domain; blocked=muse(failed); declined=cursor-grok,cursor-kimi; rounds=17; full=2 @ 88c71ad}

_{Review-Attention: study ~20m (critical: Table.ts; decisions: do-less-alternative, same-thread-only, no-store-sharing-semantics, scan-allowlist-vs-clone) @ 88c71ad}

[gemini-code-assist]

Code Review

This pull request introduces a mechanism in getFromSource() where concurrent same-key waiters on the same thread share a resolved source fill (even if nothing was stored, such as for absent records or under noCacheStore) to prevent duplicate origin requests. This change is accompanied by comprehensive unit and integration tests, as well as updated design documentation. The review feedback highlights a potential deadlock risk in resources/Table.ts where a synchronous error thrown by transaction() could escape the Promise executor, preventing the record lock from being released. Wrapping the transaction call in a try...catch block is recommended to handle synchronous failures safely.