perf(channels): coalesce concurrent GetChannels/GetEncryptedChannels cache misses

[dborup]

Upstream source

Evaluate and port the behaviour from Kpa-clawbot/CoreScope#2059: coalesce concurrent cache misses for GetChannels and GetEncryptedChannels so one cold/expired key causes one SQLite query rather than one query per request.

Why this is still relevant here

Current fork master still has no channel-query singleflight. GetChannels uses a 60-second cache, but concurrent callers can all miss before the first result is stored. GetEncryptedChannels follows a separate query path.

This fork has diverged from upstream:

• PR perf(db): pin GetChannels and CountFloodAdvertsForNode to their indexes (issue #100) #107 pins the channel query to the intended index with a safe fallback.
• PR fix(server): avoid mutating the cached channels slice on includeEncrypted #98 protects the cached channels slice when encrypted results are appended.
• The channel cache currently has a single cached region key rather than upstream's map-shaped cache.

Do not cherry-pick blindly. Preserve those behaviours and adapt request coalescing to the current cache contract.

Acceptance criteria

• Concurrent misses for the same normalized region execute at most one real GetChannels query.
• Concurrent misses for the same normalized region execute at most one real GetEncryptedChannels query.
• Different region keys are never coalesced together and cannot receive each other's result.
• Errors are shared only within the in-flight call and are not cached as successful data.
• A second cache check occurs inside the singleflight winner.
• Existing index-pinning/fallback behaviour from perf(db): pin GetChannels and CountFloodAdvertsForNode to their indexes (issue #100) #107 remains intact.
• Existing cache-slice protection from fix(server): avoid mutating the cached channels slice on includeEncrypted #98 remains intact.
• Deterministic tests count actual query executions; timing alone is not sufficient.
• Run the relevant Go tests with -race and measure cold concurrent behaviour before and after.

No deploy is part of this issue.

[dborup]

Additional scope found in upstream Kpa-clawbot#1936

The selective cache work in Kpa-clawbot/CoreScope#1936 reinforces this issue but should not be ported wholesale.

Current fork findings:

• GetChannels still has a single-slot cache rather than a cache keyed by normalized region.
• GetEncryptedChannels has no equivalent cache.
• The fork's perf(db): pin GetChannels and CountFloodAdvertsForNode to their indexes (issue #100) #107 index-pinning/fallback and fix(server): avoid mutating the cached channels slice on includeEncrypted #98 cached-slice protection must remain intact.
• The upstream message-page cache is unsafe to copy directly here: it returns shared mutable maps, while fork handlers mutate them through annotateMessageAreas and annotateBotReplyTouchedAreas.
• Do not add upstream's composite index blindly; perf(db): pin GetChannels and CountFloodAdvertsForNode to their indexes (issue #100) #107 deliberately chose the existing pinned index.

Recommended extension to this issue: use keyed caches plus per-key singleflight for normalized-region channel and encrypted-channel queries. Benchmark message-page caching as a separate decision, and require an immutable/copy-on-read contract before adding it.

[dborup]

Draft PR #138 klar til review: CI grøn (Go Build & Test, Playwright E2E, Docker build). GetChannels og GetEncryptedChannels bruger nu en begrænset cache pr. normaliseret region med singleflight og et ekstra cache-tjek inde i flighten. #107 og #98 er bevaret. Der er ingen message-page-cache og intet nyt indeks. 16 samtidige kolde kald giver 2 queries (master: 16).

---

Generated by Claude Code

[dborup]

Fixed by #138 (merged as e829d1ea). The list cache is per normalized region, concurrent misses share one query, rows.Err() is checked, and keys over 256 bytes are stored as a digest while the query still gets the full region. Review note, not blocking: TTL expiry itself has no dedicated test.