codenotary / immudb / 25360351619

85%

master: 85%

Build Type

Pull #2094

gh-ci

Committed by vchaindz

Commit Message

perf: hot-path allocation and lock-contention fixes

Seven independent improvements driven from a manual review of
allocations and lock-holding patterns in the read/write hot paths.
Each fix is self-contained and on-disk format-neutral.

P1 — sync() vLog lock convoy
embedded/store/immustore.go (sync)

  The for-range over s.vLogs called defer s.releaseVLog(i+1) inside
  the loop. Go stacks defers until the enclosing function returns,
  so all per-vLog locks were held for the full duration of every
  vLog.Flush + vLog.Sync of every other vLog, blocking concurrent
  readers on vLogsCond for the entire sync window. Replaced the
  defer with a per-iteration IIFE so each vLog drops its lock as
  soon as its own flush+sync completes. No reordering of the actual
  flush/sync calls; only the lock-release timing changes.

P2 — zero-copy snapshot reads
embedded/tbtree/{tbtree,snapshot}.go

  TBtree.Get and Snapshot.Get always finish with `return cp(v), …`
  where cp performs make+copy of the value bytes — even on cache
  hits served entirely from already-immutable in-memory leafValue
  slices. The defensive copy exists only to protect against caller
  mutation. Added GetReadonly companions next to each (additive
  API, no behaviour change for existing callers): same body minus
  the cp, with a doc-comment that the returned slice is borrowed
  and must not be mutated. Migrating callers is intentionally a
  separate audit pass — every callsite needs a hand-check before
  switching.

P4 — per-tx scratch buffer in commit loops
embedded/store/immustore.go (commitDurable, sync)

  Both commit loops allocated `cb := make([]byte, s.cLogEntrySize)`
  inside the loop body, once per transaction in a batch. cLog.Append
  ultimately routes to AppendableFile.write, which copies the input
  into its own writeBuffer (singleapp/single_app.go:477), so the
  input slice is not retained — safe to hoist a single 12- or
  44-byte buffer outside the loop and reuse it across iteration... (continued)

Pull Request Pull Request #2094: perf(s3): wave 1-4 remote-storage performance + sql index-only COUNT + hot-path fixes

Coverage Stats

601 of 855 new or added lines in 16 files covered. (70.29%)

8 existing lines in 3 files now uncovered.

45131 of 53102 relevant lines covered (84.99%)

126741.92 hits per line