Skip to content

results(linux-x64): c7i.2xlarge 10× run — isolation ratios reproduce vs Mac baseline - #3

Open
kwit75 wants to merge 1 commit into
mainfrom
results/linux-x64-c7i-20260709
Open

results(linux-x64): c7i.2xlarge 10× run — isolation ratios reproduce vs Mac baseline#3
kwit75 wants to merge 1 commit into
mainfrom
results/linux-x64-c7i-20260709

Conversation

@kwit75

@kwit75 kwit75 commented Jul 9, 2026

Copy link
Copy Markdown

Linux-x64 (c7i.2xlarge, 8 fixed cores) 10× run of the isolation/concurrency harness — the diff vs main is the linux numbers side-by-side with the Mac baseline.

Runtime 3.2.1.30 (114509c6) x86_64 · langchain-core 0.3.86. All four groups 10/10 reps clean, and the isolation ratios reproduce cross-platform (absolute timings differ from Mac, as expected):

group stock LangChain stock RocketRide
concurrent-processing .batch (shared conn) crashes 0/64 ✅ safe, M=16 ~0.70s, 0 err
fault-isolation in-process .abatch loses all 0/4 ✅ survives (only crasher dies)
data-isolation shared dict, 32 workers silently loses 40–89% ✅ 0 lost / 0 leaked
authoring-effort 14–17 lines + up to 5 hidden decisions ✅ 0 concurrency lines

Provisioning: the --strip-components=1 bug (strips the top-level engine launcher) — your fix is already on main; this run used a correctly-extracted engine/engine + a 3.11 venv, equivalent to it. Ran natively via the real C++ ./engine ai/eaas.py (not the bundled python — that would void the isolation measurement).

Heads-up — a second harness bug (not blocking; not fixed in this PR): scripts/start_engine.sh readiness gate false-passes. On connection-refused, curl -w '%{http_code}' prints 000 and the || echo 000 appends another → "000000", which != "000" → it declares the engine healthy before uvicorn is listening, so make smoke races an ECONNREFUSED. Fix = gate on a real status, e.g. if [[ "$code" =~ ^[1-5][0-9][0-9]$ ]]. Worth landing so clean re-runs don't flake.

🤖 Generated with Claude Code

…vs Mac baseline

Runtime 3.2.1.30 (114509c6) x86_64, langchain-core 0.3.86. All 4 groups 10/10 clean:
concurrent-processing (LangChain .batch crashes 0/64), fault-isolation (LC loses all 0/4),
data-isolation (LC silently loses 40-89%), authoring-effort. RocketRide safe by default on every one.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant