docs: close residual #3 (R2 merged); register row 47 (CTS-dispose bug, fixed in-flight); correct stale MaxBatchSize story in Component-StoreAndForward (R1 finding)

2026-08-15 04:14:49 -04:00
parent c75c79f8ab
commit 88425f662c
3 changed files with 15 additions and 4 deletions
@@ -80,7 +80,7 @@ There is **no maximum buffer size**. Messages accumulate in the buffer until del
 - The standby node applies the same operations to its own local SQLite database but is **passive**: it never runs the delivery sweep. The retry sweep is **gated to the active node** (the oldest Up member / singleton host, re-evaluated every sweep tick), so only one node delivers at a time. The standby applies replicated operations purely to keep its copy warm for a future failover.
 - On failover, the new active node has a near-complete copy of the buffer. In rare cases, the most recent operations may not have been replicated (e.g., a message added or removed just before failover). This can result in a few **duplicate deliveries** (message delivered but its `Remove` not yet replicated) or a few **missed retries** (message added but not replicated). Duplicate deliveries are therefore confined to the **failover window** — an in-flight delivery whose `Remove` had not yet replicated — and never occur in steady-state operation (the standby's gate keeps it from delivering the same rows). Both are acceptable trade-offs for the latency benefit.
 - On failover, the new active node's gate flips to active within one sweep interval and it resumes delivery from its local copy.
- **Peer-join anti-entropy resync (LocalDb CDC).** *(Rewritten for LocalDb Phase 2, 2026-07-20. The previous specification of a chunked, ack-confirmed `SfBufferSnapshotChunk` protocol described the bespoke `ReplicationService`, which Phase 2 deleted. The discussion is rewritten rather than removed, because the failure modes it reasoned about still exist — they are simply bounded differently now.)* The buffer lives in the consolidated LocalDb database as the replicated `sf_messages` table, and both nodes exchange changes over a gRPC sync stream rather than Akka remoting. A node that was down for an extended period no longer requests a full-buffer snapshot and **replaces** its local buffer; LocalDb's snapshot resync merges **per row under last-writer-wins and never deletes**. Several of the old hazards are therefore structurally gone rather than guarded against: **(a) the 128 000-byte Akka frame limit no longer applies** — the transport is gRPC, whose successor ceiling is the 4 MB default receive limit, managed by bounding `LocalDb:Replication:MaxBatchSize` (set to 16 on the rig, sized against a ~70 KB worst-case `config_json`; see the Phase 2 plan, D6). Chunking, `ResyncId` assembly, assembly timeouts, and the truncation flag are all retired with it. **(b) The N1 directional-authority hazard is gone.** That guard existed because the bespoke resync applied a destructive delete-all-then-insert-all, so running it in the wrong direction wiped a live buffer. With a non-destructive merge there is no wipe to gate, and replication is symmetric — either node may write. `ActiveNodeEvaluator` survives, but only for the **delivery** gate and the heartbeat, which still genuinely need a single active node. **(c) The N5 orphan-row race is gone.** A `Remove` ordered before a re-add can no longer resurrect a row: deletes are tombstoned with an HLC, and a tombstone beats any older write for the same key. **The duplicate-delivery bound, stated explicitly.** Delivery remains single-node: only the primary runs the sweep (`IClusterNodeProvider.SelfIsPrimary`). A message can therefore be delivered twice only when the OLD primary delivered it and the resulting status change had not yet replicated at the instant the gate flipped. The window is one replication flush interval plus the in-flight ack, not an unbounded divergence — and unlike the old model it does not grow with backlog depth or with how long a node was absent. **One new bound replaces the old ones:** a node offline longer than `LocalDb:Replication:TombstoneRetention` (default 7 days) may resurrect deleted rows on rejoin, because the tombstones that would have suppressed them have been pruned. Stop-and-start a site pair together, and do not leave one node of a pair offline across that horizon.
+- **Peer-join anti-entropy resync (LocalDb CDC).** *(Rewritten for LocalDb Phase 2, 2026-07-20. The previous specification of a chunked, ack-confirmed `SfBufferSnapshotChunk` protocol described the bespoke `ReplicationService`, which Phase 2 deleted. The discussion is rewritten rather than removed, because the failure modes it reasoned about still exist — they are simply bounded differently now.)* The buffer lives in the consolidated LocalDb database as the replicated `sf_messages` table, and both nodes exchange changes over a gRPC sync stream rather than Akka remoting. A node that was down for an extended period no longer requests a full-buffer snapshot and **replaces** its local buffer; LocalDb's snapshot resync merges **per row under last-writer-wins and never deletes**. Several of the old hazards are therefore structurally gone rather than guarded against: **(a) the 128 000-byte Akka frame limit no longer applies** — the transport is gRPC, whose successor ceiling is the 4 MB default receive limit, managed since LocalDb 0.2.0 by the **byte-budget** bound `LocalDb:Replication:MaxBatchBytes` (default 2 MB, enforced per wire message). `MaxBatchSize` is demoted to a secondary row cap whose remaining job is bounding the DB **read page** that `OplogStore`/`SnapshotStreamer` materialize in memory before the byte split runs (arch-review adversarial finding F2) — the rig pins it to 64 for that reason; the old rig pin of 16, hand-sized against a ~70 KB worst-case `config_json` as a byte-budget proxy (Phase 2 plan, D6), is retired. Chunking, `ResyncId` assembly, assembly timeouts, and the truncation flag are all retired with it. **(b) The N1 directional-authority hazard is gone.** That guard existed because the bespoke resync applied a destructive delete-all-then-insert-all, so running it in the wrong direction wiped a live buffer. With a non-destructive merge there is no wipe to gate, and replication is symmetric — either node may write. `ActiveNodeEvaluator` survives, but only for the **delivery** gate and the heartbeat, which still genuinely need a single active node. **(c) The N5 orphan-row race is gone.** A `Remove` ordered before a re-add can no longer resurrect a row: deletes are tombstoned with an HLC, and a tombstone beats any older write for the same key. **The duplicate-delivery bound, stated explicitly.** Delivery remains single-node: only the primary runs the sweep (`IClusterNodeProvider.SelfIsPrimary`). A message can therefore be delivered twice only when the OLD primary delivered it and the resulting status change had not yet replicated at the instant the gate flipped. The window is one replication flush interval plus the in-flight ack, not an unbounded divergence — and unlike the old model it does not grow with backlog depth or with how long a node was absent. **One new bound replaces the old ones:** a node offline longer than `LocalDb:Replication:TombstoneRetention` (default 7 days) may resurrect deleted rows on rejoin, because the tombstones that would have suppressed them have been pruned. Stop-and-start a site pair together, and do not leave one node of a pair offline across that horizon.

 ### Operation Tracking Table (lives in Site Runtime, not here)