Root cause of the ~4% error rate in audit-enabled runs (run-qc6newt3:
96.30% ok, op6 mean 367ms max 3003ms) was the audit path's three
compounding costs, isolated iteratively under the real 8-user x 150s
model-B-with-audit stress harness:
1. append(): rewrote the whole log string on every op (O(n) read-modify-write
under a per-store Mutex) -> op latency grew with log size.
2. dump(): walked 1..=count re-reading every entry record (O(n)) -> became the
new bottleneck once append was fixed (op6 still ~760-980ms).
3. the command returned the UNBOUNDED full log (~1MB at 12k entries)
on every call -> ~1MB response serialized/sent/received = op6 ~978ms.
Fix (aligned with the PDF's 'access logs live in Null rows' time/stream-keyed
model):
- AUDIT_HEAD stores only a decimal entry count (index); each entry is its own
durable record at entry_coord(seq) -> append is O(1) (two put_record calls).
- ConcurrentStore gains a per-store in-memory tail cache (audit_tail) shared by
every Audit over that store; append extends it by one line, dump returns a
clone -> dump is O(1) and never re-walks the store. Serialized under the
cache guard so concurrent connections interleave correctly.
- the interactive command serves a bounded recent tail
(Audit::AUDIT_TAIL_LIMIT = 200) instead of the full log; the full log stays
available via Session::audit_dump()/Audit::dump() for export.
Verification (real, not assumed):
- ./check gate GREEN (fmt + tests + clippy -D warnings), incl. R6 append/dump
tests and pre-existing grant_and_revoke_emit_audit_entries.
- hermes_verify_audit_o1: index=count (not log), distinct coords, ascending
dump, concurrent interleave-correct. PASS.
- model-B-with-audit re-run (8 users x 150s): 110,647 ops, 100.00% ok, 0
failures; op6 mean 11.9ms (p99 46.9ms, max 124.9ms) vs 367ms pre-fix. Final
run evidence: /root/cube-stress/run-kkaogy3b.
- Auth confirmed a non-factor (zero rejections) across all runs.
docs/stress-comparison-20260811.md: corrected the bogus '~0.3ms audit op' claim
in S5 and replaced the placeholder S6 with the full root-cause/fix/verification
write-up including the iteration-to-100% table.
Two harness drivers (model A per-command auth, model B persistent auth-once)
run across the real release cube-server to settle the "auth-each-time had ~0%
errors" memory. Conclusion: error rate is driven by the slow `audit` op /
3s socket cap, NOT the auth model — with audit removed, model A hits 94.81%
and model B 100%. Adds docs/stress-comparison-20260811.md §5 and the two
reusable harness scripts under tools/.
ConcurrentStore.inner is now Arc<RwLock<CubeStore>>: all read paths take the
read side, all mutations + checkpoint take the write side. Readers no longer
exclude each other and overlap an active writer (verified by
concurrent_reads_dont_block_on_writer + cube-bench Task 4 section). WAL,
checkpoint, and coordinate encoding are untouched, so durability/replay is
unchanged (.check green).
Honest finding recorded in docs/task4-reader-writer-sharding.md: on this 8-core
host std RwLock removes reader-vs-reader exclusion (correct) but shows no
wall-clock speedup for short reads (cache-line bounce on one shared lock). Real
read-throughput scaling would need sharded/lock-free storage, left as a
follow-up decision rather than invented.
- cube-bench crate: real-code-path throughput/latency over cubestore,
cubecrypt (aes/gcm/chacha/xts), cubecode VM, and cubesys Session.
Every section asserts correctness before timing. Wired into ./check
as an opt-in 'bench' stage.
- cubesys Session: per-command latency histogram + per-C-namespace record
counts, exposed via a new 'stats' command over the live socket.
- Deployed rebuilt cube-server to /home/luulu/.cubelinux/bin and
restarted the system cube.service; verified stats live.
CUBELinux-2