deer-flow/backend/tests/test_schedule_dispatch_race.py
rayhpeng a328af2917 refactor(schedule): wire the slice through adapters and delete the old path
The outer ring: SQL repositories, the run launcher, the thread lookup,
and the run-completion listener implementing the ports the domain
declared, plus the HTTP router and the poller driving them. All of it is
instantiated in one composition root, so no route or lifespan hook
builds an adapter of its own.

With the ports filled, the pre-hexagonal implementation is deleted
rather than left alongside: `app/scheduler/service.py` and its router
mixed policy, persistence, and HTTP into one class, which is why its
rules were only reachable through a live database.

Two bugs that shape hid are fixed here: a completion hook that replayed
a stale snapshot and rolled the launch write back, and a corrupt stored
row surfacing to the client as a 4xx. The concurrency invariants are
pinned by contract suites that run each port against both the fake and
real sqlite, plus dedicated dispatch-race tests.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-07-31 11:44:08 +08:00

224 lines
10 KiB
Python

"""Concurrency regression tests for the scheduled-task dispatch TOCTOU.
``ScheduleService.dispatch_task`` guards "at most one active run per task
when overlap_policy=skip" with a non-atomic ``has_active`` fast path followed
by a separate queued-record insert. Two concurrent dispatches (double-click,
client retry, or a manual trigger racing the poller) can both pass the check
and both launch. The database is the atomic arbiter via the partial unique
index ``uq_scheduled_task_run_active`` (``task_id WHERE status IN
('queued','running')``); the losing insert is translated to
``ActiveRunConflictError`` and collapsed to the same outcome as the fast
path.
These tests drive the REAL ``SqlScheduledRunRepository`` +
``SqlScheduledTaskRepository`` + ``ScheduleService`` against a real
file-backed sqlite database (so the index is actually enforced), with a fake
launcher that only records launches. The contract suite deliberately does not
own this: its doubles provide no atomicity, so a green contract run says
nothing about two dispatchers racing -- this file is where that is proven.
"""
from __future__ import annotations
import asyncio
from datetime import UTC, datetime
import pytest
from schedule_fakes import FakeThreadLookup
from app.adapters.schedule.scheduled_run_repository import SqlScheduledRunRepository
from app.adapters.schedule.scheduled_task_repository import SqlScheduledTaskRepository
from deerflow.config.database_config import DatabaseConfig
from deerflow.domain.schedule.exceptions import ActiveRunConflictError
from deerflow.domain.schedule.model import DispatchOutcome, RunStatus, ScheduledRun, ScheduledTask, SchedulePolicy, ScheduleSpec, TriggerKind
from deerflow.domain.schedule.ports import LaunchedRun
from deerflow.domain.schedule.service import ScheduleService
from deerflow.persistence.engine import close_engine, get_session_factory, init_engine_from_config
pytestmark = pytest.mark.asyncio
NOW = datetime(2026, 7, 27, 12, 0, tzinfo=UTC)
class _BarrierRunRepo(SqlScheduledRunRepository):
"""Real repository that only releases both dispatchers past ``has_active``
once both have read it, so their ``add()`` calls genuinely race for the
task's single active slot -- a deterministic reproduction of the
check-then-insert TOCTOU."""
def __init__(self, session_factory, barrier: asyncio.Barrier | None) -> None:
super().__init__(session_factory)
self._barrier = barrier
async def has_active(self, task_id: str) -> bool:
result = await super().has_active(task_id)
if self._barrier is not None:
await self._barrier.wait()
return result
class _RecordingLauncher:
"""Launch double that yields first, so a truly-concurrent sibling can
interleave before the launch is recorded."""
def __init__(self) -> None:
self.calls: list[dict] = []
async def launch(self, *, thread_id, assistant_id, prompt, owner_user_id, metadata) -> LaunchedRun:
await asyncio.sleep(0)
self.calls.append({"thread_id": thread_id, "metadata": metadata})
return LaunchedRun(run_id=f"run-{len(self.calls)}", thread_id=thread_id)
def _make_service(tasks, runs, launcher) -> ScheduleService:
return ScheduleService(
tasks=tasks,
runs=runs,
launcher=launcher,
threads=FakeThreadLookup(),
policy=SchedulePolicy(min_once_delay_seconds=60, max_concurrent_runs=10, lease_seconds=120),
)
async def _seed_task(tasks: SqlScheduledTaskRepository, title: str) -> ScheduledTask:
# fresh_thread_per_run: every dispatch gets a NEW thread_id, so #4003's
# per-thread uq_runs_thread_active can never fire for two dispatches of the
# same task -- this is precisely the gap the per-task index closes.
return await tasks.add(
ScheduledTask.create(
user_id="user-1",
title=title,
prompt="do the thing",
schedule=ScheduleSpec.cron_schedule("*/5 * * * *", "UTC"),
context_mode="fresh_thread_per_run",
thread_id=None,
now=NOW,
policy=SchedulePolicy(),
)
)
async def _active_run_count(runs: SqlScheduledRunRepository, task_id: str) -> int:
rows = await runs.list_by_task(task_id, limit=100)
return sum(1 for row in rows if row.is_active)
async def test_two_concurrent_manual_dispatches_launch_exactly_once(tmp_path):
await init_engine_from_config(DatabaseConfig(backend="sqlite", sqlite_dir=str(tmp_path)))
try:
sf = get_session_factory()
assert sf is not None
tasks = SqlScheduledTaskRepository(sf)
runs = _BarrierRunRepo(sf, asyncio.Barrier(2))
launcher = _RecordingLauncher()
service = _make_service(tasks, runs, launcher)
task = await _seed_task(tasks, "task-race-manual")
results = await asyncio.gather(
service.dispatch_task(task, now=NOW, trigger=TriggerKind.MANUAL),
service.dispatch_task(task, now=NOW, trigger=TriggerKind.MANUAL),
)
outcomes = sorted((result.outcome for result in results), key=str)
# Exactly one wins the active slot; the loser is a 409-style conflict.
assert outcomes == [DispatchOutcome.CONFLICT, DispatchOutcome.LAUNCHED], outcomes
assert len(launcher.calls) == 1, launcher.calls
assert await _active_run_count(runs, task.task_id) == 1
# The manual loser records no run-history row (nothing was scheduled).
conflict = next(r for r in results if r.outcome is DispatchOutcome.CONFLICT)
assert conflict.record_id is None
finally:
await close_engine()
async def test_scheduled_and_manual_dispatch_launch_exactly_once(tmp_path):
await init_engine_from_config(DatabaseConfig(backend="sqlite", sqlite_dir=str(tmp_path)))
try:
sf = get_session_factory()
assert sf is not None
tasks = SqlScheduledTaskRepository(sf)
runs = _BarrierRunRepo(sf, asyncio.Barrier(2))
launcher = _RecordingLauncher()
service = _make_service(tasks, runs, launcher)
task = await _seed_task(tasks, "task-race-mixed")
results = await asyncio.gather(
service.dispatch_task(task, now=NOW, trigger=TriggerKind.SCHEDULED),
service.dispatch_task(task, now=NOW, trigger=TriggerKind.MANUAL),
)
outcomes = [result.outcome for result in results]
# Whichever won launched; the loser is conflict (manual) or skipped
# (scheduled). Which one wins is timing-dependent, but exactly one runs.
assert outcomes.count(DispatchOutcome.LAUNCHED) == 1, outcomes
assert set(outcomes) <= {DispatchOutcome.LAUNCHED, DispatchOutcome.CONFLICT, DispatchOutcome.SKIPPED}, outcomes
assert len(launcher.calls) == 1, launcher.calls
assert await _active_run_count(runs, task.task_id) == 1
finally:
await close_engine()
async def test_natural_timing_concurrent_dispatch_launches_exactly_once(tmp_path):
# No barrier: exercise the fix under the same natural interleaving that
# reproduced the bug (5/5 both-launch before the index). The fix must hold
# whether the second dispatch is caught by the has_active fast path or by
# the index-violation path.
await init_engine_from_config(DatabaseConfig(backend="sqlite", sqlite_dir=str(tmp_path)))
try:
sf = get_session_factory()
assert sf is not None
tasks = SqlScheduledTaskRepository(sf)
runs = SqlScheduledRunRepository(sf)
for i in range(5):
launcher = _RecordingLauncher()
service = _make_service(tasks, runs, launcher)
task = await _seed_task(tasks, f"task-natural-{i}")
results = await asyncio.gather(
service.dispatch_task(task, now=NOW, trigger=TriggerKind.MANUAL),
service.dispatch_task(task, now=NOW, trigger=TriggerKind.MANUAL),
)
outcomes = [result.outcome for result in results]
assert outcomes.count(DispatchOutcome.LAUNCHED) == 1, (i, outcomes)
assert len(launcher.calls) == 1, (i, launcher.calls)
assert await _active_run_count(runs, task.task_id) == 1, i
finally:
await close_engine()
async def test_partial_unique_index_enforces_one_active_run_per_task(tmp_path):
# Focused repository-level test of the index semantics + the typed conflict.
await init_engine_from_config(DatabaseConfig(backend="sqlite", sqlite_dir=str(tmp_path)))
try:
sf = get_session_factory()
assert sf is not None
runs = SqlScheduledRunRepository(sf)
now = datetime(2026, 7, 2, 1, 0, tzinfo=UTC)
first = ScheduledRun.queued(task_id="t1", thread_id="th1", scheduled_for=now, trigger=TriggerKind.SCHEDULED)
await runs.add(first)
# queued -> running is a same-row UPDATE: keeps the one active slot, no
# violation (this is the normal launch transition).
await runs.update_status(first.record_id, status=RunStatus.RUNNING, run_id="run-1", started_at=now)
assert await runs.has_active("t1") is True
# A second active insert for the same task is a domain conflict.
with pytest.raises(ActiveRunConflictError):
await runs.add(ScheduledRun.queued(task_id="t1", thread_id="th2", scheduled_for=now, trigger=TriggerKind.MANUAL))
# Terminal-status rows for the same task are outside the index predicate.
await runs.add(ScheduledRun.skipped_tombstone(task_id="t1", thread_id="th3", scheduled_for=now, trigger=TriggerKind.SCHEDULED))
# A different task's active row is independent.
await runs.add(ScheduledRun.queued(task_id="t2", thread_id="th4", scheduled_for=now, trigger=TriggerKind.SCHEDULED))
# Finishing the active run frees the slot; a fresh active row is allowed.
await runs.update_status(first.record_id, status=RunStatus.SUCCESS, run_id="run-1", finished_at=now)
assert await runs.has_active("t1") is False
await runs.add(ScheduledRun.queued(task_id="t1", thread_id="th5", scheduled_for=now, trigger=TriggerKind.SCHEDULED))
assert await runs.has_active("t1") is True
finally:
await close_engine()