Files
seaweedfs/sw-block/.private/phase/phase-13-cp8-workload-validation.md
T
pingqiuandClaude Opus 4.6 d7cd415714 feat: CP13-8 — bounded real-workload validation scenario + envelope
One named workload validation package for RF=2 sync_all:
- Scenario: cp13-8-real-workload-validation.yaml (6 phases)
- ext4 proof: write 200 files → failover → fsck + file count + md5sum diff
- pgbench proof: TPC-B on promoted replica (database durability)
- Disturbance: one bounded failover (kill primary, promote replica)

Workload envelope doc: phase-13-cp8-workload-validation.md
- Named topology, transport, workloads, disturbance, exclusions
- Pass criteria: fsck passes, 200 files, checksums match, pgbench TPS > 0
- Maps each pass criterion to accepted CP13-1..7 semantics
- Explicit non-claims: not rollout approval, not NVMe, not soak, not CP13-9

Reuses existing infrastructure:
- cp85-db-ext4-fsck.yaml pattern (extended with checksums + pgbench)
- benchmark-pgbench.yaml actions (pgbench_init/pgbench_run)

Must run on real hardware (m01/M02). Cannot run in unit test harness.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-03 08:59:46 -07:00

3.3 KiB

CP13-8 Real-Workload Validation — Envelope + Contract

Date: 2026-04-03

Workload Envelope

Parameter Value
Topology RF=2, sync_all, cross-machine (m01 ↔ M02)
Transport iSCSI (primary frontend)
Filesystem workload ext4: 200 files, write + sync + failover + fsck + checksum verify
Application workload PostgreSQL pgbench TPC-B (scale=1, c=1, 10s) on promoted replica
Disturbance One bounded failover: kill primary, promote replica (epoch 1→2)
NOT included NVMe-TCP, RF>2, hours/days soak, degraded-mode perf, mode normalization

Scenario

weed/storage/blockvol/testrunner/scenarios/internal/cp13-8-real-workload-validation.yaml

Phase flow

  1. Setup: RF=2 sync_all pair (primary on M02, replica on m01), standalone iscsi-target binary
  2. ext4 write: iSCSI login → mkfs ext4 → write 200 files → md5sum → sync → umount → wait replication
  3. Failover: kill primary → promote replica to primary (epoch 2)
  4. ext4 verify: iSCSI login to promoted replica → fsck (filesystem integrity) → mount → file count == 200 → md5sum diff == MATCH
  5. pgbench: iSCSI login → pgbench_init (ext4, scale=1) → TPC-B run (c=1, 10s) → TPS reported
  6. Cleanup: always-run phase

Pass criteria

Proof Assertion What it validates
ext4 integrity fsck_ext4 passes Replicated writes are filesystem-consistent after failover
ext4 completeness file_count == 200 No files lost during replication + failover
ext4 correctness md5sum diff == MATCH File content identical to pre-failover (no corruption)
pgbench durability pgbench_run completes with TPS > 0 Database transactions are durable on sync_all promoted replica

Relation to CP13-1..7

Accepted checkpoint What this workload validates
CP13-2 (address truth) Cross-machine iSCSI replication uses canonical addresses
CP13-3 (durable progress) ext4 data survives failover because barrier guarantees flushed durability
CP13-4 (state eligibility) Only InSync replica was eligible for barrier during replication
CP13-5 (reconnect/catch-up) Replication completed before failover (all writes reached replica)
CP13-6 (retention) WAL retained long enough for replication to complete
CP13-7 (rebuild fallback) Not directly exercised — failover is clean (no WAL gap). Support-only.

Existing infrastructure reused

Existing scenario Relation
cp85-db-ext4-fsck.yaml CP13-8 extends this pattern with checksums, pgbench, and explicit envelope
benchmark-pgbench.yaml CP13-8 pgbench phase uses same pgbench_init + pgbench_run actions

Run instructions

# From m01 (client node):
sw-test-runner run cp13-8-real-workload-validation.yaml

# Or from Windows dev machine (testrunner SSH):
cd C:/work/seaweedfs
go run ./weed/storage/blockvol/testrunner/cmd/sw-test-runner run \
  weed/storage/blockvol/testrunner/scenarios/internal/cp13-8-real-workload-validation.yaml

What CP13-8 Does NOT Close

  • Mode normalization (CP13-9)
  • Broad launch approval
  • Performance floor (see Phase 12 P4)
  • Degraded-mode validation
  • NVMe-TCP transport validation
  • Hours/days soak under sustained load