Narrows two queue-overload scenarios to what the labels measure, adds the
S3 and broker images to the queue profile passport, and corrects report
figures. Thresholds and verdicts are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The host swapped out 26046 pages, all in one cell. A further label,
baseline4, follows by the user's decision; the main one is picked by the
number of affected cells, not by verdicts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The host swapped out 309432 pages during the run and five cells are
flagged host_swap, so this label stays as a record and baseline3 is
captured as the main one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reject uploads with 429 at [uploads] max_unfinished_jobs, counted by the existing partial indexes in the membership block. Add the bench_queue profile, hypotheses H1-H9 and the fifth failure-stand API.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Abort the transport when a bounded close ends, so a frozen server leaves
no sessions over the pool once the ping has no transaction. Under the
shared request deadline a block sets no timer of its own, so a single
cancel closes the connection and the log has one record. Add a test that
concurrent worker deliveries get separate units of work, and state the
cost of the pattern against both baselines in the report.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Services depend only on src/core/ports: async with self.uow is one
transaction, commit is explicit, leaving without it rolls back. A
connection is held per step, the JSON routers keep a shared request
deadline, and rollback runs under the block deadline. Adds
service-tests with fakes, a repo contract run on both the fake and SQL,
and make service_test / service_test_db. API, schema and settings are
unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Close the source image after converting to RGBA so PNGs at the pixel limit
fit the render memory limit. Strengthen the fencing test so that it fails
without the attempt check, and measure worker and relay stops under load.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Move publishing out of the upload path: a relay process publishes queued
jobs with confirms. Add lease and attempt fencing, backoff with jitter, a
per-job render process with CPU and memory limits, crop and optimized
variants, the failed registry with audited replay, and staging cleanup.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Evidence taken on ff1a83b. The report covers level 3; levels 4 and 5 are
closed by the following changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upload answers 202 after staging, a queued job in PostgreSQL and a broker
confirm; a separate worker publishes the original and a thumbnail with
bounded retries. Adds the job and variant tables, the broker service, the
worker S3 user, cleanup and reconcile support for staging and variants,
the failure stand with broker and worker, bench and driver updates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Source: mtusi2026lab4-safronov, solution/bvt2301-safronov, 2546deb.
Contracts, public tests and evidence of lab 4 are not carried over.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>