Leases: the worker renews its lease before the first side effect and
every JOB_TIMEOUT_SECONDS/3 while the variants are produced; a renewal
that finds the lease taken cancels the work (and its render process)
and settles the job as fenced without writing anything.
Retries: the delay before attempt n+1 is drawn from [d/2, d] with
d = min(cap, base * 2^(n-1)) (equal jitter), the next attempt time is
computed by PostgreSQL and the delay is kept in the job event.
Faults: the Faults port is async; relay reaches after_commit_before_publish
and after_publish_before_confirm, worker reaches after_claim_before_side_effect,
stall_after_claim (once, FAULT_STALL_SECONDS), after_s3_write_before_commit
and after_commit_before_ack. Crashes are os._exit(137).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The renderer adapter keeps one single-process pool per WORKER_CONCURRENCY
slot. A render that exceeds RENDER_TIMEOUT_SECONDS is killed and its slot
is respawned, the render process has RLIMIT_AS = RENDER_MEMORY_MB on
Linux, so both limits end in ImageTooDemanding, a permanent
processing_failed. A render process killed from outside is a transient
error. A cancelled render (lost lease, shutdown) kills its process too.
The worker container limit grows to 1g: two 448 MiB render processes
plus the worker itself.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The worker decodes the original once (JPEG draft down to the optimized
box, EXIF orientation applied in place, alpha composed on white) and
encodes three variants: the 256 px thumbnail, a centred 128x128 JPEG crop
that may upscale and a WebP fitted into 1280x1280 without upscaling,
quality 80, method 4. No EXIF, ICC or XMP is written. The byte limit is
checked before decoding. ready is set by the transaction that stores all
three rows, so a partial set is never published.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Upload now stages the checked original under staging/, and one
transaction moves the intent to queued, inserts the job (pipeline
version 1, attempt 1), its outbox row and an audit event; the 202 is
the in-memory snapshot, never a reread. Content answers 409 until the
attachment is ready, the thumbnail variant is served from the
published version and absent variants are 404.
The relay publishes outbox rows through FastStream with publisher
confirms, mandatory routing and persistent delivery, marks a row only
after the broker acked it, returns jobs whose lease expired and deletes
the staging object of ready attachments. The worker consumes the
quorum queue through a FastStream subscriber that declares nothing,
takes prefetch from WORKER_CONCURRENCY and acks manually after the
result commit, rejects messages off jobs.schema.json to the dead letter
exchange and requeues after an infrastructure error. A job is claimed
under a lease token; the original is copied to originals/ and a
256 px JPEG thumbnail is written under a content-addressed key; the
commit is fenced by the token, inserts variants with ON CONFLICT DO
NOTHING and publishes only a newer pipeline version. A transient error
reschedules the same job with an exponential delay and a new outbox
row, the last attempt fails with retry_exhausted, an undecodable image
fails at once with invalid_image, and a redelivered finished job is a
recorded no-op.
The worker and the relay are assembled in worker.py and relay.py with
the dishka FastStream integration and the platform providers of the
API; the shared core imports neither FastStream nor aio-pika. Tests
cover the thumbnail geometry and rendering, the job message against
the contract, the pipeline against PostgreSQL (duplicates, parallel
copies, retries, the attempt bound, leases and fencing) and the broker
adapters against a throwaway RabbitMQ with the stand's topology and
permissions.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>