feat: allow unrestricted independent job concurrency

src/archive_clients/daemon.py

@@ -43,6 +43,19 @@ from archive_control.v1 import (
 logger = logging.getLogger(__name__)
 
 
+def _job_id_for_command(command: control_pb2.Command) -> str:
+    """Return the durable job key for commands whose execution is serialized."""
+
+    payload = command.WhichOneof("payload")
+    if payload == "assign_job":
+        return command.assign_job.job.job_id
+    if payload == "execute_step":
+        return command.execute_step.job_id
+    if payload == "cancel_job":
+        return command.cancel_job.job_id
+    raise ValueError(f"command {command.command_id} does not execute a job")
+
+
 class ArchiveClientDaemon:
     def __init__(
         self,
@@ -95,7 +108,10 @@ class ArchiveClientDaemon:
         self._lease = DatabaseLease(config.state_db)
         self._active_route_commands: set[str] = set()
         self._active_job_commands: set[str] = set()
-        self._job_execution_lock = asyncio.Lock()
+        # Commands for one job remain ordered locally, while unrelated jobs
+        # may use the node's available qB/Syncthing/filesystem capacity in
+        # parallel. The control daemon owns admission policy.
+        self._job_execution_locks: dict[str, asyncio.Lock] = {}
         self.jobs = (
             ClientJobExecutor(
                 client_id=config.client_id,
@@ -560,7 +576,9 @@ class ArchiveClientDaemon:
         correlation_id: str,
         outbound: asyncio.Queue[str],
     ) -> None:
-        async with self._job_execution_lock:
+        job_id = _job_id_for_command(command)
+        lock = self._job_execution_locks.setdefault(job_id, asyncio.Lock())
+        async with lock:
             await self._execute_job_command_locked(
                 command, correlation_id, outbound
             )

src/archive_clients/jobs.py

@@ -86,6 +86,8 @@ class ClientJobExecutor:
         self.verification_timeout = verification_timeout
         self.free_space_reserve_bytes = free_space_reserve_bytes
         self._cancel_events: dict[str, threading.Event] = {}
+        self._execution_locks: dict[str, threading.Lock] = {}
+        self._execution_locks_guard = threading.Lock()
 
     def request_cancel(self, job_id: str) -> None:
         self._cancel_events.setdefault(job_id, threading.Event()).set()
@@ -115,6 +117,22 @@ class ClientJobExecutor:
         self,
         command: control_pb2.ExecuteStepCommand,
         event_callback: Callable[[control_pb2.JobEvent], None] | None = None,
+    ) -> list[control_pb2.JobEvent]:
+        # asyncio cancellation of a connection-bound task cannot stop the
+        # synchronous filesystem/qB operation already running in its worker
+        # thread. A replay after reconnect therefore waits for that operation
+        # and then reads its durable event journal instead of executing twice.
+        with self._execution_lock(command.job_id):
+            return self._execute_locked(command, event_callback)
+
+    def _execution_lock(self, job_id: str) -> threading.Lock:
+        with self._execution_locks_guard:
+            return self._execution_locks.setdefault(job_id, threading.Lock())
+
+    def _execute_locked(
+        self,
+        command: control_pb2.ExecuteStepCommand,
+        event_callback: Callable[[control_pb2.JobEvent], None] | None = None,
     ) -> list[control_pb2.JobEvent]:
         definition = self._definition(command.job_id)
         replay = self._replay(