deer-flow/backend/app/composition.py
rayhpeng c38d291505 refactor(schedule): standardize the module on the hexagonal architecture
Replaces the pre-hexagonal scheduled-task implementation with a slice
built to the layering spec: a pure domain (two aggregates, two state
machines, the policy value object), output ports it declares itself,
SQL/launcher/thread adapters implementing them under `app/adapters/`,
and a composition root that is the one place any of them is
instantiated.

The old implementation mixed all of that into `app/scheduler/service.py`
and a router that reached straight into repositories, so the rules that
matter -- overlap policy, lease handling, which write owns which
timestamp -- were only reachable through a live database. They are now
unit-assertable on in-memory fakes, with the contract suite running each
port against both the fake and real sqlite, and the concurrency
invariants pinned by dedicated race tests.

Two bugs the old shape hid are fixed on the way: a completion hook that
replayed a stale snapshot and rolled back the launch write, and a
corrupt stored row surfacing to the client as a 4xx.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-07-31 11:16:46 +08:00

120 lines
5.1 KiB
Python

"""Composition root -- the one place adapters are instantiated and wired.
Every application service is assembled here and nowhere else. Ports are
declared by the domain, implemented under ``app/adapters/``, and the two are
introduced to each other in this file; no router, middleware, or lifespan hook
constructs an adapter of its own.
**Why this is a pure function rather than part of the lifespan.** Wiring used
to live inside ``deps.py::langgraph_runtime``, tangled with engine startup,
orphan recovery, and graceful shutdown -- so the one rule that actually
governs it ("no SQL backend means no service, and the routes answer 503")
could not be tested without driving a full application startup, and was held
up by a single comment. ``build_domain_services`` takes what it needs and
returns what it built, so that rule is an assertion in
``tests/test_composition.py`` instead.
**What "no SQL backend" means.** ``session_factory is None`` is how a
``database.backend: memory`` deployment presents itself. The context owns
tables, so it cannot run on it; the service is ``None`` and the dependency
providers translate that into 503. This is deliberately not a silent
degradation to an in-memory implementation -- scheduled work that vanishes on
restart is worse than scheduled work that is refused.
"""
from __future__ import annotations
from collections.abc import Awaitable, Callable, Mapping
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any
from deerflow.domain.schedule.model import SchedulePolicy
from deerflow.domain.schedule.service import ScheduleService
if TYPE_CHECKING:
from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
from deerflow.config.scheduler_config import SchedulerConfig
from deerflow.persistence.thread_meta.base import ThreadMetaStore
@dataclass(frozen=True)
class DomainServices:
"""Every application service the Gateway serves, or ``None`` where the
configured backend cannot support one."""
schedule: ScheduleService | None
def build_domain_services(
*,
session_factory: async_sessionmaker[AsyncSession] | None,
thread_store: ThreadMetaStore,
launch_run: Callable[..., Awaitable[Mapping[str, Any]]],
scheduler_config: SchedulerConfig,
) -> DomainServices:
"""Wire the domain services from already-built infrastructure.
Takes infrastructure rather than building it: engines, stores and the run
launcher have lifecycles (startup, recovery, shutdown) that belong to the
lifespan, while deciding *what is assembled from them* is this function's
only job. That split is what makes it callable from a test.
"""
if session_factory is None:
return DomainServices(schedule=None)
# Imported here, not at module scope: these pull in SQLAlchemy and the
# Gateway run path, and the composition root is imported by tests that
# only want the memory-backend branch.
from app.adapters.schedule.run_launcher import GatewayRunLauncher
from app.adapters.schedule.scheduled_run_repository import SqlScheduledRunRepository
from app.adapters.schedule.scheduled_task_repository import SqlScheduledTaskRepository
from app.adapters.schedule.thread_lookup import ThreadStoreThreadLookup
return DomainServices(
schedule=ScheduleService(
tasks=SqlScheduledTaskRepository(session_factory),
runs=SqlScheduledRunRepository(session_factory),
launcher=GatewayRunLauncher(launch_run),
threads=ThreadStoreThreadLookup(thread_store),
policy=build_schedule_policy(scheduler_config),
),
)
def build_run_completion_hook(
schedule_service: ScheduleService | None,
) -> Callable[[Any], Awaitable[None]] | None:
"""Install the schedule context on the run runtime's completion callback.
The inbound half of the wiring, and nothing more: which runs the context
cares about and what it does with them belongs to the adapter, not here.
This function's only decision is the same one the rest of this module
makes -- what to assemble, and what to leave unassembled.
Returns ``None`` when there is no service, so the runtime installs no hook
at all rather than one that always declines.
"""
if schedule_service is None:
return None
from app.adapters.schedule.run_completion import ScheduleRunCompletionListener
return ScheduleRunCompletionListener(schedule_service)
def build_schedule_policy(scheduler_config: SchedulerConfig) -> SchedulePolicy:
"""Project the operator's scheduler config onto the domain's policy.
Separate from the wiring above because it is the whole of the
configuration-to-domain translation: the domain declares which thresholds
it needs, and this names where each one comes from. `poll_interval_seconds`
is deliberately absent -- how often to look is the poller's business, not a
rule any task is subject to.
"""
return SchedulePolicy(
min_once_delay_seconds=scheduler_config.min_once_delay_seconds,
max_concurrent_runs=scheduler_config.max_concurrent_runs,
lease_seconds=scheduler_config.lease_seconds,
)