v0.15.0 Milestone Soak
v0.15.0 milestone soak
Section titled “v0.15.0 milestone soak”A condensed (~2.5h) alternative to the 48–72h soak, run against a live 3-node Cassandra 5.0 cluster to answer one question before the milestone shipped: is 0.15 sound?
Verdict: yes — 7 / 8 claims VERIFIED. Every claim a condensed run can test passed. The milestone’s key open flag — the background snapshot grace-sweep (#2452) — is confirmed fixed in the field. Multi-day stability (claim 8) is out of scope for a short run by design; all its short-horizon proxies are green.
The full report is a self-contained dashboard with four linked views:
- Verdict — the 8-claim table with evidence.
- Stability — RSS, fd/thread, OOM/restart traces over the run.
- Phases — per-phase breakdown (cold start, load, burst, drain, failover).
- Comparison — 0.15 vs the R11b / R12 baselines and the Lane-B goals ladder.
What was verified
Section titled “What was verified”| # | Claim | Verdict |
|---|---|---|
| 1 | ~15× warm throughput (snapshot lifecycle, lazy index, point-read streaming) | VERIFIED — 90 min @ 32-thr: 211,373 queries, ~39 qps, p50 798ms, 0 client errors |
| 2 | Admission control under overload | VERIFIED — 80-thr burst: graceful queuing, not collapse; clean recovery |
| 3 | Saturation observability (blocking/egress/fd/threads/RSS) | VERIFIED — all gauges legible, returned to 0 after load |
| 4 | Snapshot retirement — background grace-sweep | VERIFIED — 660→6 snapshots at t=5min with zero queries (the R12 regression is fixed) |
| 5 | No silent row loss on large cells (≥1MB) | VERIFIED — integrity sweeps identical at 1,927,467 rows |
| 6 | Query-semantics correctness under concurrency | VERIFIED — 6 concurrent point-reads → 1 byte-identical result |
| 7 | Cold start / restart behavior | VERIFIED — cold first-query 2.2s, 0 index parses at boot, clean failover |
| 8 | Multi-day stability | COULD-NOT-OBSERVE — out of scope for ~2.5h; proxies all green (RSS flat 301–357Mi, 0 OOMKills) |
Three things to report (none a milestone blocker)
Section titled “Three things to report (none a milestone blocker)”- Fan-out skew caps throughput — a single pod is the ceiling (this is the one Lane-B goal, B2, not yet met).
- 0.89% server-side
do_geterror rate — client-invisible, and down from 1.2% in R12. - A harness metrics-pipeline bug inflated the dashboard’s
rate()panels ~1000× — found, fixed in the lab, and verified during this run; future runs render correctly.
Run metadata: flight
v0.15.0@sha256:30c2b10c(multi-arch INDEX) · connector 0.15.0 · Trino 481 · Cassandra 5.0 RF=3 · 3× i4i.xlarge · ~1.93M partitions/node · 2026-07-17.