mirror of
https://github.com/Kpa-clawbot/meshcore-analyzer.git
synced 2026-11-22 03:17:04 +00:00
## Summary Two pre-existing ways a running server **permanently** missed data. The database was correct; the server's in-memory store was not, and stayed wrong until the next restart. Found while comparing several server replicas side by side in #2396: two replicas that restarted mid-ingest each lacked one transmission and its 12 observations. Split out of #2396 at review's request, because it affects every SQLite install. ## The bugs 1. **Startup window.** The poller took its starting cursors from a fresh `MAX(id)` *after* the startup load finished. A row committed between the load's last chunk snapshot and that query ended up in neither the store nor any later poll. Its later observations were then dropped as orphans. 2. **Failed poll.** A failed observation read still advanced the cursor to its upper bound, skipping up to 500 observations for good. ## The fix - `NewPacketStore` records the transmission and observation high-water marks **before** the load, as the advert-evidence and rewrite feeds already did. - The loaded max transmission id stays as a floor. - The loaded max observation id no longer is one: a chunk-1 transmission can gain an observation whose id is *below* one that chunk 2 loaded. - Rows read twice are deduplicated and broadcast at most once. - A failed read keeps the cursor and retries on the next tick, logging once per failure streak. - Comments and a test pin why an observation of a transmission whose chunk has not merged yet cannot be skipped: the poller's evidence refresh waits on that chunk's lock. ## Performance - **Startup:** two extra indexed `MAX(id)` queries, which are primary-key lookups. - **Poll loop:** unchanged complexity. The only new per-tick work is a dedupe check on rows read twice, a map lookup by id (O(1)). Re-reads happen only for rows in the startup window, which is bounded by what was committed during the load, and on a retry after a failed read. ## Testing `cmd/server/startup_poll_gap_test.go` seeds the ingestor's own DDL plus `dbschema.Apply`: | Test | What it pins | |------|--------------| | `TestStartupLoadPollerGap_SQLite` | A transmission and observations committed between the load's snapshot and the poller's start reach the store | | `TestStartupLoadPollerGapMultiChunk_SQLite` | Same, across 3 chunks, including an observation id below the max loaded one | | `TestStartupObsDuringChunkMerge_SQLite` | An observation of a transmission whose chunk has not merged yet is not skipped | | `TestPollerRetriesFailedRead_SQLite` | A failed observation read does not advance the cursor | **Red/green:** against master's `store.go`, `websocket.go`, `chunked_load.go` and `db.go`, the first, second and fourth tests fail with the rows missing, e.g.: ``` tx 3, committed between the load's snapshot and the poller's cursor init, is missing from the store (lost forever) obs 3, committed while the observation read failed, never reached the store: the cursor passed it ``` The third depends on test hooks this change adds, so it cannot run against the old code. `go vet` and `go test` pass on all modules, with gofmt clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --- ## Merge order These four PRs were split out of #2396 and are **independent — they can merge in any order.** Verified: every pair is conflict-free, and all four merged together pass `go vet`, `go test ./...` (all 16 modules) and `sh test-all.sh`. Recommended **priority** (not a dependency): 1. **#2410** — musl 8 MiB thread stacks. A public, unauthenticated GET can SIGSEGV the whole server on the shipped image, so merge this first. 2. **#2408** — never lose rows during startup / a failed poll. 3. **#2407** — explicit `observers.idx` primary key. 4. **#2409** — content-hash migration moved into the ingestor. After these land, #2396 (Postgres) will be rebased onto master. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>