4 of 4
the required failure cases, proven
Killed before acknowledging, killed after it, a handler reply that never arrives, a crash during compaction, a full disk, a second writer — each one mapped to the tests and live drills that cover it.
Measured. The kill -9 drill stops the server mid-work 36 times under load: 25,201 acknowledged submits and 9,119 cancels, zero lost, zero duplicated — a laptop drill, a development observation.
The failure-case map·The drill·The contract
1,000,000
of 1,000,000 submits accepted
Sent, accepted and drained to done — every single one. p50 12.5 ms, p99 41.3 ms; the journal grew to 496 MiB.
Measured. The sustained gate at 100 submits a second for about 2.8 hours on a busy laptop (Apple M2 Pro, median load 3.52) — a development observation, not a capacity figure. The same gate honestly misses on CI: a 4-core runner reaches the live-runs cap at about 85 completions a second, then refuses 503 at_capacity — a kept miss, never rerun until green.
The changelog entry·The run·The kept miss
0.4 s / 1.7 s
restart to ready
Process start to answering again, journal replay included: 0.4 s with 100,000 runs queued, 1.7 s with 800,000 — against a 5 s upgrade-unavailability target.
Measured. The restart-ready probe on the same laptop, after the journal became versioned; the unversioned format measured 1.1 s and 8.7 s the same night.
What a restart costs·The runs
64 · 100/s
per client address
Connections one untrusted address may hold, and requests it may start per second; over the line it gets 429 rate_limited with Retry-After. IPv6 counts per /64, with a /48 aggregate.
Default. Like every limit here, they are flags you set, not promises we make.
The change·Security
/inspector
the runner, inspected
Runs newest-first or oldest-first with cursor paging, job detail, submit, cancel, compact, backup, diagnostics — read and write keys kept separate, stored in sessionStorage.
Same-origin. Served by durable-server itself: no second deploy, no CORS.
The Inspector