mirror of
https://github.com/Kpa-clawbot/meshcore-analyzer.git
synced 2026-05-11 07:16:55 +00:00
Kpa-clawbot
5a5df5d92b
## Why Diagnostic on #1129 shows PR #1117 (group commit M1 for #1115) is fundamentally broken: it starves the MQTT goroutine via `gcMu` lock contention, causing pingresp disconnects and lost packets at modest ingest rates. ## Three structural defects 1. **Lock held across `sql.Stmt.Exec`** — every concurrent `InsertTransmission` blocks for the full SQLite write latency, not just the brief queue mutation. 2. **Lock held across `tx.Commit`** — the WAL fsync runs *under* `gcMu`, so any backlog blocks all ingest writers AND the flusher ticker, snowballing under load. 3. **Single-conn DB** (`MaxOpenConns=1`) — the flusher and the ingest path serialise on one connection, turning the lock into a global ingest stall. Net effect: at modest packet rates the MQTT client loop misses its own pingresp deadline, the broker drops the connection, and packets received during the stall are lost. ## What this PR removes - `Store.SetGroupCommit`, `Store.FlushGroupTx`, `Store.flushLocked`, `Store.GroupCommitMs` - `gcMu`, `activeTx`, `pendingRows`, `groupCommitMs`, `groupCommitMaxRows` Store fields - `groupCommitMs` / `groupCommitMaxRows` config fields and `GroupCommitMsOrDefault` / `GroupCommitMaxRowsOrDefault` accessors - The flusher goroutine in `cmd/ingestor/main.go` - `cmd/ingestor/group_commit_test.go` - The `if s.activeTx != nil { … pendingRows … }` branch in `InsertTransmission` — reverts to plain prepared-stmt usage ## What this PR keeps (merged after #1117) - #1119 `BackfillPathJSON` `path_json='[]'` fix - #1120/#1123 perf metrics endpoints — `WALCommits` counter retained - `GroupCommitFlushes` JSON field on `/api/perf/write-sources` is kept as always-0 for API stability (server `perf_io.go` references it as a string field name; no client breakage) - `DBStats.GroupCommitFlushes` atomic field is removed from the Go struct ## Tests `cd cmd/ingestor && go test ./... -run "Test"` → `ok` (47.8s). `cd cmd/server && go build ./...` → clean. ## #1115 stays open The group-commit *idea* is sound — batching observation INSERTs would meaningfully reduce WAL fsync rate. But it needs a redesign that does **not** hold a mutex across blocking SQLite calls. Suggested directions for a future M1: - Channel-fed writer goroutine (single owner of the tx, ingest path is non-blocking enqueue) - Per-batch DB handle so the flusher doesn't serialise the ingest connection - Bounded queue with backpressure rather than a shared lock Refs #1117 #1129