From 72f22ec46e883c625a7d1670e66f786c231cd8a3 Mon Sep 17 00:00:00 2001 From: pingqiu Date: Sat, 2 May 2026 18:21:47 -0700 Subject: [PATCH] docs(p15): record G9A best-effort ACK oracle --- .../design/v3-phase-15-g9a-ack-reintegration-mini-plan.md | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/sw-block/design/v3-phase-15-g9a-ack-reintegration-mini-plan.md b/sw-block/design/v3-phase-15-g9a-ack-reintegration-mini-plan.md index 76125a6f0..f13d5cbec 100644 --- a/sw-block/design/v3-phase-15-g9a-ack-reintegration-mini-plan.md +++ b/sw-block/design/v3-phase-15-g9a-ack-reintegration-mini-plan.md @@ -56,15 +56,16 @@ Rule: recovery progress is not a substitute for synchronous ACK eligibility. | `5232e3a` | `cmd/blockvolume --replication-ack=best-effort|sync-quorum|sync-all` wires daemon mode to replication durability and durable write ACK policy. | | `520b25b` | `/status` append-only G9A vocabulary: returned old primary is `AuthorityRole=superseded`, `FrontendPrimaryReady=false`, `ReplicationRole=not_ready`; authority movement does not imply replica readiness. | | `8362d42` | subprocess L2 strict ACK oracle: real blockmaster + 2x blockvolume + iSCSI, `--replication-ack=sync-quorum`, secondary down => foreground WRITE returns non-GOOD. | +| `154bd96` | subprocess L2 best-effort oracle: same RF=2 daemon/iSCSI shape, secondary down => foreground WRITE still returns GOOD. | ### 3.2 Next red tests 1. `TestG9A_ReplicaCandidate_RequiresProgressFactBeforeReady` - Candidate readiness requires durable/progress fact, not assignment presence. -2. `TestG9A_BestEffortStillFeedsLaggingReplica` - - Best-effort write returns success, and progress policy still starts catch-up/rebuild for lagging peer. -3. `TestG9A_SyncQuorumWriteFailsWhenPeerInRecovery_Process` +2. `TestG9A_SyncQuorumWriteFailsWhenPeerInRecovery_Process` - Follow-up variant where peer is explicitly in recovery (not merely down) and the daemon path still fails foreground ACK. +3. `TestG9A_BestEffortStillFeedsLaggingReplica` + - Follow-up variant that proves best-effort success does not disable catch-up/rebuild for the lagging peer. ---