 pingqiuandClaude Opus 4.6
|
55013e103b
|
feat: Phase 20 Stage 0+1 closure — bootstrap + sustained workload on hardware
Stage 0 (bootstrap closure): PASS on m01/M02
- create RF=2 sync_all → 10s shipper wait → 4k fsync → publish_healthy
- Proves: BarrierAccepted observation, ShipperConnected, DurableLSN > 0
Stage 1 (sustained workload): 32/33 actions PASS
- bootstrap → fio 10s randwrite → dd_write 1M×2 fsync → data checksum
- Remaining: auto-failover promotion (separate issue)
Key fixes:
- BarrierAccepted callback: SyncCache success → core DurableLSN update
- BarrierRejected callback: barrier failures surface to core with reason
- Shipper state callback for new volumes (not just startup volumes)
- CatchUpTo ctrl conn reset: prevents stale control channel after recovery
- CP13-6 max-bytes budget suspended: uses replicaFlushedLSN which can't
advance without barrier; kills healthy shippers during async writes.
Will be replaced by v2 negotiated sync/recovery protocol.
- Barrier diagnostic logging: start/fail/success with reason and LSN
- Scenario restructured: Stage 0 (bootstrap-closure) + Stage 1 (failover)
- dd_write: sync_mode param + real stderr capture
- sw-test-runner suite command: deploy once, run N scenarios
- WAL size plumbing: proto + API + handler (forward-compatible)
Known: 6 blockvol/server test failures from Barrier() path change
(bounded catch-up in Barrier). Need test updates to match new semantics.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-06 19:55:12 -07:00 |
|
|
pingqiuandClaude Opus 4.6
|
3ec8fab2f1
|
feat: Phase 10 — control-plane closure (identity, convergence, idempotence)
Stable identity on wire:
- ServerID fields in proto (replica_server_id, server_id on ReplicaAddrMessage)
- volumeServerId wired through volume.go → BlockService.SetServerID
- Identity derived from canonical server ID, not transport addresses
Assignment convergence:
- V2 idempotence via lastAppliedAssignment.equals (full replica set comparison)
- setupPrimaryReplication/Multi idempotence guards
- ProcessAssignments with V2 + V1 dual-path assignment handling
Master-driven control loop:
- RecoveryManager: serialized cancel-and-drain via done channels
- Per-replica heartbeat state reporting (ReplicaShipperStatus)
- masterServerBackend: VolumeBackend calling real MasterServer in-process
- RestoreBlockSnapshot RPC (master + volume server proto)
QA tests (P10 P1-P4):
- Identity: ServerID on wire, fail-closed on missing
- Convergence: assignment delivery, epoch monotonicity, registry coherence
- Idempotence: repeated assignment, multi-replica set comparison
- Control loop: integrationMaster + real allocator + proto round-trip
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-02 16:25:43 -07:00 |
|