mirror of
https://github.com/ModernRelay/omnigraph.git
synced 2026-06-12 01:45:14 +02:00
1b0a2c9310
Closes the PR 1b regression (-17% disjoint, -30% same-key) by eliminating the fresh `db.snapshot_for_branch(branch).await` that PR 1b's commit_all issued per mutation. Single-table mutations (`staged.len() + inline_committed.len() == 1`): skip revalidation entirely. The per-(table, branch) Mutex queue holds exclusive while we commit; the publisher's CAS catches any drift that slipped between expected_versions capture and queue acquisition. Conflict cost: 1 orphan Lance HEAD advance, recovered via the existing sidecar protocol on the next ReadWrite open. This is the same trade-off the master plan §"Revalidation perf optimization" prescribes. Multi-table mutations: replace `db.snapshot_for_branch(branch)` (fresh manifest read) with `db.snapshot()` (in-memory). Correct under MR-686's single-process scope because all in-process tenants share one `Arc<Omnigraph>` -> one coordinator; publishes update the shared coordinator BEFORE releasing queue guards, so a contending tenant reads a fresh in-memory view by the time it acquires its queue keys. The within-mutation race (A captures expected_versions[T2]=V0, B publishes T2->V1 during A's stage I/O, A then acquires T2's queue) is caught via the in-memory check. Multi-coordinator deployments (§VI.27 aspirational) would need force-refresh under the queue — documented in §VI's "Explicit non-commitments". Adds a SAFETY comment naming the two load-bearing premises: (1) per-table queue uses exclusive Mutex (not RwLock), and (2) single-coordinator invariant (one Omnigraph engine per process). Migrating either breaks this skip. Regression sentinel `change_conflict_returns_manifest_conflict_409` passes. 102 lib + 24 runs + 16 staged_writes pass with the new path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| omnigraph | ||
| omnigraph-cli | ||
| omnigraph-compiler | ||
| omnigraph-server | ||