Skip to content

Capacity: measured, not assumed

View as Markdown

A submit is acknowledged after the journal commit; the run then executes at machine speed. So the number that matters is the completion rate, not the accept rate — and when the offered rate outruns it long enough, the backlog hits the bound and admission refuses honestly rather than queueing silently:

  • --max-live-runs (default 100,000) — non-terminal runs admitted at once; past it, submits are 503 at_capacity.
  • --max-index-bytes (default ~512 MiB) — approximate bytes of retained runs; same refusal.

A 503 {"code":"at_capacity"} means the machine’s ceiling was crossed, not that work was lost — retry after the backlog drains, and the submit’s op_key dedupes whichever attempt actually landed.

Sustained runs, per-record fsync via group commit — development observations, not guarantees:

machine result
12-core laptop (M2 Pro) 100 submits/s sustained — 1,000,000/1,000,000 accepted and done, p99 41 ms
4-core hosted CI runner ~85 completions/s — a 100/s offer had 835,397 accepted and done, then refused 161,535 honestly as 503 at_capacity at the live-run bound; p99 hit the 10 s reply deadline. Zero 429s — the rate limiter wasn’t the bound; the machine was

The CI run is kept as evidence, not rerun until green: it shows exactly how an over-offered server behaves — it keeps completing at its own ~85/s and refuses the overflow by name. Ceilings move with disk latency (fsync) more than with cores, which is why a 2-vCPU cloud host lands below the laptop.

Do not plan from this table — plan from a probe on the machine that will actually serve. The product ships a load harness (perf/load.mjs, needs Node.js) that spawns its own server and handler on ports you name against a scratch journal — it does not touch a running unit’s state, though it shares the machine, so run it before real traffic:

Terminal window
cd /path/to/durable-runner # its target/release/durable-server is what gets measured
node perf/load.mjs sustained --rate=50 --total=30000 --port=58399 --hport=58398

Watch the done-rate and p99 in the result, then repeat a rung above or below to bracket the ceiling. If the refused share is nonzero, the offered rate was above the machine — not above the product’s bounds. Record the sustained completions/s; that number, with headroom, is your offered rate.

Restart cost is the other capacity question — how long an upgrade’s replay takes grows with queued runs, and deployment carries the measured table.

Was this page helpful?

Either answer opens a short issue on GitHub, so you can say what worked or what was missing.