Bench: Linux series follow-up at 1.0 CPU and 512M #793

Summary

The Linux series ran on 2026-11-14, with one REST replica at --cpus=1 --memory=512m. It answered the production question for most shapes (REPORT.md §1). This issue tracks what is still owed: shapes not run, single sessions, and sessions the OOM kill voided.

Measured

• Void and lost reruns: hang 20261114T043005 and closed loop 20261114T043953 (aborted: SIGTERM, the run agent ended its turn). Hang 20261114T045241 and closed loop 20261114T045740 (XProtect and logd CPU at an edge).
• Known limits: the VM does not charge socket and slab memory (nokmem,nosocket). The healthcheck was off. VM timer wake-ups add about 4 ms to the p99 at 200/s or less.

Cause

Owed work, not a defect.

What to change

1. Run the shapes that did not run on Linux: miss-real, huge, herd, redis-stall and limiter. miss-real comes first, because it is the production shape.
2. Repeat every scenario to k ≥ 2.
3. Run hang with steps at 24, 28, 32 and 36, to place the knee between hang@20 and hang@40.
4. Run hit above 12800/s.
5. Run an icon session on a quiet host, so the knee scan reaches 3200/s.
6. Get valid image, hostile and open-loop miss-fast sessions. Either run them after Link metadata: stream image relays or cap image bytes in flight #786 and Link metadata: shed cold work before the queue grows to an OOM kill #787, or stop each ramp before the kill step.
7. Repeat aa, because its p50 shift was significant.
8. Run the Docker-backend smokes owed on macOS: step0, aa and redis-stall.
9. Add a parse-attributed block timer for hostile pages (REPORT.md §3).
10. Keep the run agent alive until each session footer is written (REPORT.md §5.7 item 7).
11. Done: LINUX-SERIES.md now says the series ran.

Decisions

1. Change the send-lag limit inside the VM? Recommended: not yet. Count the void hit steps first, then rule.
2. Stop a ramp at the first OOM kill? Recommended: yes, for the after runs of Link metadata: stream image relays or cap image bytes in flight #786 and Link metadata: shed cold work before the queue grows to an OOM kill #787. A kill ends the session anyway.
3. Which Redis? Recommended: keep redis:alpine 8.8.0, recorded, as its own fact group.
4. Measure the core speed ratio to a production vCPU? Recommended: yes, with one short CPU-bound probe on the production host. The maintainer approves, because it touches production.

Benchmark

Before

The table in Measured. LINUX-RESULTS.md holds the full step tables. The hang and closed-loop reruns are in raw/linux/logs/{hang,miss-fast-conc}-rerun2.{replay,breaks}.out.

After

For each route, report:

• the rate where throttled_usec starts to rise;
• the memory.current max per step, and memory.events;
• each oom_kill, with its rate.

Condition

The Linux condition in Bench conditions below. It must say: a Linux VM on one laptop, not the production host. The core speed ratio to a production vCPU is not measured.

Content-safety checks

• None. The bench runs under --network none and touches no production data.

UX checks

• None. No reader-facing change.

Rollback

None needed. The work stays in local notes and the bench harness (1a0d21b7f).

Acceptance criteria

• miss-real has a valid Linux session.
• Every scenario has k ≥ 2 valid Linux sessions, or a note on why not.
• The hang knee is placed between hang@20 and hang@40.
• The Linux numbers stay in their own fact group, never mixed with macOS numbers.
• The status lines in LINUX-SERIES.md match the runs.

Related

• REPORT.md §1 (Linux series status) and §5.8.
• Link metadata: stream image relays or cap image bytes in flight #786 to Link metadata: join concurrent misses for one URL #791 need their Linux after from this runner.

Bench conditions

The bench is apps/hocuspocus.server/scripts/bench-link-tolerance/ (commit 1a0d21b7f). The headline numbers are in apps/hocuspocus.server/src/modules/link-metadata/BENCHMARK.md, section Load tolerance. REPORT.md, LINUX-RESULTS.md, research-*.md and raw/ are local bench notes (Notes/local-docs/2026-11-13-link-scraper-tolerance/), not in the repo. The Linux runner is local too.

Linux condition, used by every "Linux before": one REST container at --cpus=1 --memory=512m --memory-swap=512m, in the OrbStack VM on one laptop. Kernel 7.2.12-orbstack-00394-gbecdd6130068, booted with cgroup.memory=nokmem,nosocket, so socket and slab memory are not charged. --no-healthcheck. Bun 1.4.4+38afee6c8. Redis 8.8.0 (redis:alpine) in REST's network namespace. --network none. Pooled loopback to one local origin, no DNS, no TLS. benchTree 8774a6275a05005f, srcTree a9b6ca73ca68c2ec, runnerTree 516099a8c9839eac. Measured 2026-11-14. The core speed ratio to a production vCPU is not measured.

Laptop condition, used by every "Laptop before": local, one laptop (Apple M5 Max, 18 cores, 48 GB, macOS), Bun 1.4.4+38afee6c8, native Redis 8.10.3. REST runs under the sandbox-exec fence. Traffic is plain HTTP to one local origin, pooled loopback, no DNS, no TLS. No CPU quota and no cgroup. benchTree 8774a6275a05005f, srcTree a9b6ca73ca68c2ec. Measured 2026-11-14. Never compare it with a Linux number as one series.

"(void session)" marks a value from a valid step inside a void session. On Linux the OOM kill voided those sessions. The kill is the measured break.

After method, used by every "After":

• Base = the commit before the change. Head = the change.
• Run ABAB, at least 3 sessions per arm, each with its control step, at n ≥ 2,000 per step.
• The change moves srcTree, so the before must run again. Do not reuse the 2026-11-14 numbers as the before.
• The runner takes src/ from the working tree. So it first needs a source-folder option for base (REPORT.md §5.7 item 5).
• On Linux, start each session in a fresh REST container, and keep the run agent alive until the footer is written.
• Quote a change only when every head session sits outside the base range. Give σ and the MDE beside it.

Paths are under apps/hocuspocus.server/src/ unless they say otherwise.