- bench stand is a separate compose project mtusi-bench (API, Postgres,
Valkey) with fixed limits, fresh volume per size and migrations
- k6 registers with a password, logs in with Basic and sends Bearer
- VACUUM (ANALYZE) and CHECKPOINT after prefill
- EXPLAIN (ANALYZE, BUFFERS) of the app's history query via auto_explain,
plus the same query without the index in a rolled back transaction
- Postgres CPU, table and index size in results; storages in passport
- bench_compare checks storage limits, images and settings
- remove inproc.py and make bench_inproc
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
make failure_checks brings up a throwaway compose project with its own
PostgreSQL and Valkey, a default and a tight API (2 s sessions, pool of
one), and runs the pytest tests marked fault through an explicit hook:
session expiry, database stop and return, pool exhaustion, chat creation
rollback, readiness, Valkey stop, hang, return and flush. The stand and
its volume are removed on any outcome; the working stand is untouched.
Regular runs deselect the fault tests, and 13 coverage exclusions go away.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Valkey 9.1 in compose caches the immutable chat card (title, members,
created_at) used by GET /chats/{id} and by the membership check of the
message endpoints. Cache-aside with a write after commit, keys per
database, TTL plus allkeys-lru, persistence off. redis-py runs without
retries; on failure the cache is skipped for a few seconds and requests
go to PostgreSQL. Readiness does not depend on Valkey.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
uv run forwards SIGINT to the child that already got it from the
process group, so the second KeyboardInterrupt aborted the cleanup and
left the test container and database behind. Run the script with
$(PYTHON) like the skeleton checks and ignore SIGINT during cleanup.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replace the in-memory store with PostgreSQL 18.6: SQLAlchemy 2.1 async
engine on asyncpg, alembic migrations with an async env.py, repos on
AsyncSession and services that own the transaction. Username
uniqueness is enforced by the database (ON CONFLICT), a chat and its
members are created in one transaction, member order is kept in
chat_members.position, and a composite foreign key makes the sender a
member of the chat. History and chat lists use keyset pagination on
(chat_id, created_at, id) and (user_id, created_at, chat_id).
The compose stand runs postgres, a one-shot migrate service on the
API image and the API. All Postgres addresses come from .env without
literals; only the 127.0.0.1 publish address stays literal. make
run_dev starts Postgres and applies migrations before the server, and
make env_setup generates a random Postgres password.
Unavailable PostgreSQL yields 503: get_db_session bounds the request
with asyncio and classifies driver errors. BoundedCloseConnection
limits asyncpg's graceful close, which otherwise waited without limit
on a frozen server and held the request until it was resumed.
Add scripts/restart.sh for make test-persistence and make
test_isolated, which runs the image on a throwaway test database.
Remove make run_no_compose. X-User-Id identification is unchanged
until the session-auth change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copy the committed lab 1 sources as-is: src/, student-tests/, bench/,
Dockerfile, docker-compose.yml, pyproject.toml, uv.lock,
.python-version, config_example.toml, AGENTS.md, CLAUDE.md, .claude/,
openspec/, .graphifyignore and KNOWN_ISSUES.md. Lab 1 evidence stays
in the lab 1 repository.
Merge the new skeleton files with own sections: Makefile (lab 2
targets plus own run, docker, bench and common targets), .gitignore,
.dockerignore and .env.example. Rename the image to
mtusi-messenger:lab2 and point AGENTS.md at lab 2.
The application is still the in-memory lab 1 implementation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>