Andrey Antukh 8328974f2e ✨ Complete jobs with the multipart result and skip status
`complete-job` accepts a second form: the worker sends the artifact
itself as a multipart body, the server stores it in the `job-resource`
bucket owned by the job profile, and the completion coalesces
`resource_id` and the result descriptor `{resource-uri, filename,
mtype, size}` in its own transaction (first terminal state wins; a
lost race releases the artifact it stored, so nothing lingers for the
storage GC). The checks run before any object is written: a lost race
against a cancel never stores bytes, the artifact type must be one the
export produces (new `cm/export-artifact-types` allowlist) and the size
must fit the new `:exporter-max-result-size` config (default 350 MiB)
below the HTTP multipart ceiling.

`complete-job` and `fail-job` take an optional `session-id`, deleted
best-effort after the commit, and every write method answers `:skip`
with the `:status` of the row it lost against, read fresh after the
failed write, so a worker tells a cancelled job apart from one it was
too late for.

AI-assisted-by: glm-5.3-flash
2026-10-07 14:03:09 +02:00
..
2026-09-22 10:22:31 +02:00
2026-09-29 09:07:34 +02:00
2026-10-06 15:12:15 +02:00
2026-10-06 15:12:15 +02:00
2026-03-11 15:24:40 +01:00