유대선
← 프로젝트로
·트러블슈팅·3 분

PvP simulator surfaces 4 concurrency bugs (captured as xfail tests)

Driving a real Bo3 match through a PvP simulator against the live docker stack surfaced 4 concurrency/protocol bugs, which were then locked into a permanent pytest integration harness as xfail tests.

What happened

Jason wanted to see the PvP server work end-to-end, not just read code. We ran the full stack via docker compose (Postgres 16, Redis 7, FastAPI on port 8000) and drove a real match.

The first POST /api/v1/auth/guest returned 500 because the tables did not exist. Root cause: docker-compose.yml overrode the Dockerfile CMD and skipped alembic upgrade head. Fixed by running the 2 migrations manually (a67ac29dd885 + b8f3c72e1a44).

We wrote scripts/pvp_simulator.py (~210 lines): two virtual players (P1 aggro, P2 defensive), concurrent guest registration via httpx, concurrent WS via the websockets library, driving queue join -> match_found -> game WS -> submit_action loop -> match_result.

Running with --seed 42, a full Bo3 match completed in ~12s with a final score of 2-0 and ~400ms matchmaking pair latency — validating REST auth, matchmaking pairing, game WS, simultaneous reveal, perspective flipping, and round/match termination. Along the way the simulator exposed 4 bugs.

The 4 bugs

Bug 1 — spurious opponent_reconnected on first connect. When P2 first connects, the session already exists (P1 created it), so the code falls into the reconnect branch via handle_reconnect() at ws.py:158-162. There's no distinction between first connect and reconnect-after-disconnect.

Bug 2 — duplicate waiting_for_action per turn. session.start() is called from two places in ws.py (lines 151 and 171), both firing when both players connect, doubling _send_waiting_for_action. Lacks an idempotency check.

Bug 3 — message-order race. The server sends action_confirmed and turn_result via two separate await ws.send_json calls, and the _resolve_turn() path crosses two players' sockets, so turn_result can arrive before action_confirmed.

Bug 4 — action_confirmed lacks a turn sequence number. Originally described as "messages don't carry a turn sequence number." During test writing this was reassessed: reading app/schemas/ws.py showed that turn_result does carry turn_number, so the original description was inaccurate. The real gap is narrower — action_confirmed lacks turn_number. This was split into a passing positive regression test (test_turn_result_carries_turn_sequence_number) and an xfail (test_action_confirmed_carries_turn_number).

Note: the milestone title says the simulator "surfaces 4 concurrency bugs," but after the Bug 4 reassessment the 4th xfail reflects this revised, smaller scope. The code locations above are quoted from the engineering log, not independently re-verified against source.

Converting to a permanent test harness

Phase 2.1 turned the CLI simulator into a pytest integration harness under tests/integration/. conftest.py drives one full Bo3 match against the live docker stack and records every WebSocket event into a MatchRecording dataclass (auto-skips if the API is unreachable). test_pvp_flow.py asserts on the recording.

The test file splits into two classes:

Final run: 6 passed, 4 xfailed in 15.06s. The bugs are explicitly deferred to Phase 3 — neither commit is a feat fix.

Decisions

Commits