Preserve worker requests through storage admission pauses

src/durable_workflow/errors.py

@@ -171,6 +171,32 @@ def reason(self) -> str | None:
             return self.body.get("reason")
         return None
 
+    def is_storage_admission_failure(self, poll_request_id: str | None = None) -> bool:
+        """Whether the runtime explicitly refused admission and requested an identity-preserving retry."""
+        body = self.body
+        if (
+            self.status != 503
+            or not isinstance(body, dict)
+            or self.reason() not in ("storage_pressure", "storage_admission_unavailable")
+            or body.get("retryable") is not True
+            or type(body.get("retry_after_seconds")) is not int
+            or body["retry_after_seconds"] <= 0
+            or body.get("storage_state") not in ("draining", "fenced")
+            or (self.reason() == "storage_admission_unavailable" and body["storage_state"] != "fenced")
+            or ("request_admitted" in body and body["request_admitted"] is not False)
+        ):
+            return False
+        if poll_request_id is None:
+            return body.get("request_admitted") is False
+        return (
+            bool(poll_request_id)
+            and "task" in body and body["task"] is None
+            and body.get("poll_status") == self.reason()
+            and body.get("poll_request_id") == poll_request_id
+            and body.get("retry_same_poll_request_id") is True
+            and body.get("claim_admitted") is False
+        )
+
 
 class NexusOperationFailed(DurableWorkflowError):
     """A Nexus service operation completed with a typed service failure."""

src/durable_workflow/retry_policy.py

@@ -13,14 +13,50 @@
 from __future__ import annotations
 
 import asyncio
+import contextvars
+import json
+import logging
 import random
 from collections.abc import Awaitable, Callable
 from dataclasses import dataclass
 from typing import TypeVar
 
 import httpx
 
+from .errors import ServerError
+
 T = TypeVar("T")
+log = logging.getLogger("durable_workflow.worker")
+
+# Task-local so sharing a Client never changes unrelated client/control requests.
+_worker_storage_admission_stop: contextvars.ContextVar[Callable[[], bool] | None] = contextvars.ContextVar(
+    "worker_storage_admission_stop", default=None,
+)
+
+
+def _storage_refusal(exc: Exception) -> tuple[ServerError, str | None] | None:
+    if not isinstance(exc, httpx.HTTPStatusError):
+        return None
+    if "X-Durable-Workflow-Protocol-Version" not in exc.request.headers:
+        return None
+    try:
+        body = exc.response.json()
+    except ValueError:
+        return None
+    error = ServerError(exc.response.status_code, body)
+    if error.reason() not in ("storage_pressure", "storage_admission_unavailable"):
+        return None
+    poll_id = None
+    if exc.request.url.path.endswith("/poll"):
+        try:
+            request = json.loads(exc.request.content)
+            poll_id = request.get("poll_request_id") if isinstance(request, dict) else None
+        except ValueError:
+            pass
+        # An invalid submitted ID must not fall through to the non-poll contract.
+        if not isinstance(poll_id, str) or not poll_id:
+            poll_id = ""
+    return error, poll_id
 
 
 @dataclass
@@ -81,6 +117,7 @@ async def execute(self, fn: Callable[[], Awaitable[T]]) -> T:
         Raises the last exception if all retries are exhausted.
         """
         attempt = 0
+        storage_attempt = 0
         last_exc: Exception | None = None
 
         while attempt < self.max_attempts:
@@ -89,6 +126,32 @@ async def execute(self, fn: Callable[[], Awaitable[T]]) -> T:
                 return result
             except Exception as exc:
                 last_exc = exc
+                stop = _worker_storage_admission_stop.get()
+                refusal = _storage_refusal(exc) if stop is not None else None
+                if refusal is not None and stop is not None:
+                    error, poll_id = refusal
+                    if not error.is_storage_admission_failure(poll_id) or stop():
+                        raise
+                    storage_attempt += 1
+                    assert isinstance(error.body, dict)
+                    delay = min(
+                        5.0,
+                        max(
+                            self.backoff_seconds(min(storage_attempt - 1, 6)),
+                            error.body["retry_after_seconds"],
+                        ),
+                    )
+                    log.warning("storage admission paused; retrying the same worker request in %.2fs", delay)
+                    # Do not consume the finite transport budget or repeat serialization/uploads.
+                    while delay > 0:
+                        if stop():
+                            raise
+                        interval = min(0.1, delay)
+                        await asyncio.sleep(interval)
+                        delay -= interval
+                    if stop():
+                        raise
+                    continue
                 if not self.should_retry(exc, attempt):
                     raise
 

src/durable_workflow/worker.py

@@ -27,10 +27,11 @@
 import traceback
 import types
 import uuid
-from collections.abc import Awaitable, Callable, Iterable, Mapping
+from collections.abc import Awaitable, Callable, Coroutine, Iterable, Mapping
 from datetime import datetime, timezone
+from functools import wraps
 from types import FunctionType
-from typing import Annotated, Any, Literal, Union, get_args, get_origin, get_type_hints
+from typing import Annotated, Any, Concatenate, Literal, ParamSpec, TypeVar, Union, get_args, get_origin, get_type_hints
 
 from . import serializer
 from .activity import ActivityContext, ActivityInfo, _set_context
@@ -75,6 +76,7 @@
     WORKER_TASKS,
     MetricsRecorder,
 )
+from .retry_policy import _worker_storage_admission_stop
 from .workflow import (
     Command,
     NexusServiceCall,
@@ -137,6 +139,27 @@
 _WORKFLOW_TASK_COMPLETION_RETRY_DELAYS = (0.05, 0.2)
 _WORKFLOW_TASK_NEXUS_RESOLUTION_LIMIT = 100
 _WORKER_WORKFLOW_FINGERPRINTS: dict[tuple[str, str], str] = {}
+_P = ParamSpec("_P")
+_R = TypeVar("_R")
+
+
+def _with_storage_admission_retries(
+    fn: Callable[Concatenate[Worker, _P], Coroutine[Any, Any, _R]],
+) -> Callable[Concatenate[Worker, _P], Coroutine[Any, Any, _R]]:
+    @wraps(fn)
+    async def run(self: Worker, /, *args: _P.args, **kwargs: _P.kwargs) -> _R:
+        token = _worker_storage_admission_stop.set(self._stop.is_set)
+        try:
+            return await fn(self, *args, **kwargs)
+        finally:
+            _worker_storage_admission_stop.reset(token)
+    return run
+
+
+def _is_storage_admission_error(error: BaseException) -> bool:
+    return isinstance(error, ServerError) and error.reason() in (
+        "storage_pressure", "storage_admission_unavailable",
+    )
 
 
 def _command_payload_codec(codec: object) -> str:
@@ -208,6 +231,8 @@ def _should_fail_workflow_task_after_completion_error(error: BaseException) -> b
 
 
 def _should_retry_workflow_task_completion_error(error: BaseException) -> bool:
+    if _is_storage_admission_error(error):
+        return False
     if isinstance(error, ServerError):
         return error.status >= 500 or error.status == 429
 
@@ -1424,6 +1449,8 @@ async def _run_workflow_task_core(self, task: dict[str, Any]) -> list[dict[str,
                 )
             except Exception as e:
                 log.warning("failed to complete workflow update task %s: %s", task_id, e)
+                if _is_storage_admission_error(e):
+                    return None
                 if _should_fail_workflow_task_after_completion_error(e):
                     await self._report_workflow_task_after_completion_error(task_id, attempt, e)
                     return None
@@ -1550,6 +1577,8 @@ async def _run_workflow_task_core(self, task: dict[str, Any]) -> list[dict[str,
             )
         except Exception as e:
             log.warning("failed to complete workflow task %s: %s", task_id, e)
+            if _is_storage_admission_error(e):
+                return None
             if _should_fail_workflow_task_after_completion_error(e):
                 await self._report_workflow_task_after_completion_error(task_id, attempt, e)
                 return None
@@ -1736,6 +1765,8 @@ async def _run_activity_task(self, task: dict[str, Any]) -> str:
                 log.warning("failed to report activity failure: %s", fe)
             return "failed_non_retryable"
         except Exception as e:
+            if _is_storage_admission_error(e):
+                raise
             log.exception("activity failed")
             try:
                 await self.client.fail_activity_task(
@@ -1971,6 +2002,9 @@ async def _run_query_task_core(self, task: dict[str, Any], *, client: Client | N
                 **self._external_storage_completion_kwargs(),
             )
         except ServerError as e:
+            if _is_storage_admission_error(e):
+                log.warning("query task %s acknowledgement paused: %s", query_task_id, e)
+                return "complete_error"
             if _is_final_query_task_rejection(e):
                 log.info(
                     "query task %s completion was rejected after the task ended server-side: %s",
@@ -2147,6 +2181,8 @@ async def _poll_workflow_tasks(self) -> None:
                 self._release_workflow_capacity()
                 if self._stop.is_set():
                     return
+                if _is_storage_admission_error(e):
+                    raise
                 self._record_poll_metrics("workflow", "error", time.perf_counter() - poll_start)
                 log.warning("workflow poll error: %s", e)
                 await asyncio.sleep(1.0)
@@ -2236,10 +2272,15 @@ async def _poll_activity_tasks(self) -> None:
                     timeout=self._poll_http_timeout,
                     build_id=self.build_id,
                 )
+            except asyncio.CancelledError:
+                self._act_semaphore.release()
+                raise
             except Exception as e:
                 self._act_semaphore.release()
                 if self._stop.is_set():
                     return
+                if _is_storage_admission_error(e):
+                    raise
                 self._record_poll_metrics("activity", "error", time.perf_counter() - poll_start)
                 log.warning("activity poll error: %s", e)
                 await asyncio.sleep(1.0)
@@ -2289,6 +2330,8 @@ async def _poll_query_tasks(self, *, client: Client | None = None, track_tasks:
                     query_thread_stop is not None and query_thread_stop.is_set()
                 ):
                     return
+                if _is_storage_admission_error(e):
+                    raise
                 self._record_poll_metrics("query", "error", time.perf_counter() - poll_start)
                 log.warning("query poll error: %s", e)
                 await asyncio.sleep(1.0)
@@ -2525,6 +2568,7 @@ def _run_query_task_thread(self) -> None:
         except Exception:
             log.exception("query task poller thread stopped unexpectedly")
 
+    @_with_storage_admission_retries
     async def _query_task_thread_main(self) -> None:
         loop = asyncio.get_running_loop()
         task = asyncio.current_task()
@@ -2562,6 +2606,7 @@ async def _stop_query_task_thread(self, *, deadline: float) -> None:
                 "the worker registration remains active"
             )
 
+    @_with_storage_admission_retries
     async def run(self) -> None:
         """Register the worker and poll until `stop()` is called or the task is cancelled."""
         self._begin_run()
@@ -2622,6 +2667,8 @@ async def _heartbeat_loop(self) -> None:
                     process_metrics=self._current_process_metrics(),
                 )
             except Exception as e:
+                if _is_storage_admission_error(e) and not self._stop.is_set():
+                    raise
                 log.warning("worker heartbeat failed: %s", e)
                 continue
             if isinstance(ack, dict):
@@ -2703,6 +2750,7 @@ def _current_process_metrics(self) -> dict[str, Any]:
 
         return metrics
 
+    @_with_storage_admission_retries
     async def run_until(
         self,
         *,
@@ -2714,17 +2762,19 @@ async def run_until(
 
         This is intended for examples, smoke tests, and single-workflow scripts.
         Long-running workers should call :meth:`run` and coordinate shutdown from
-        their process supervisor.
+        their process supervisor. ``timeout`` includes registration and runtime
+        admission pauses; shutdown retains its separate drain timeout.
         """
         background_tasks: list[asyncio.Task[Any]] = []
+        deadline = asyncio.get_running_loop().time() + timeout
 
         self._begin_run()
         try:
-            await self._register()
-        finally:
-            self._registration_done.set()
+            try:
+                await asyncio.wait_for(self._register(), timeout=timeout)
+            finally:
+                self._registration_done.set()
 
-        try:
             if self._stop.is_set():
                 raise asyncio.CancelledError
             background_tasks.append(asyncio.create_task(self._heartbeat_loop()))
@@ -2742,7 +2792,11 @@ async def run_until(
                 )
             )
             self._poller_tasks.add(run_until_loop)
-            return await run_until_loop
+            return await asyncio.wait_for(
+                run_until_loop, timeout=max(0.0, deadline - asyncio.get_running_loop().time()),
+            )
+        except asyncio.TimeoutError as error:
+            raise TimeoutError(f"workflow {workflow_id} not terminal after {timeout}s") from error
         finally:
             primary_error = sys.exc_info()[1]
             primary_traceback = primary_error.__traceback__ if primary_error is not None else None

tests/fixtures/replay_regressions/storage-paused-cold-completion.json

@@ -0,0 +1,18 @@
+{
+  "$schema": "https://raw.githubusercontent.com/durable-workflow/.github/main/regression-corpus/evidence-schema.json",
+  "fixture_schema": "durable-workflow.replay-regression/v1",
+  "id": "storage-paused-cold-completion",
+  "protocol_version": "1.19",
+  "bindings": ["python"],
+  "workflow": {
+    "type": "golden.single-activity",
+    "input": ["Ada"],
+    "payload_codec": "avro"
+  },
+  "history": [
+    {"event_type": "WorkflowStarted", "payload": {"payload_codec": "avro"}},
+    {"event_type": "ActivityScheduled", "payload": {"sequence": 1, "activity_type": "golden.greet"}},
+    {"event_type": "ActivityCompleted", "payload": {"sequence": 1, "result": "wwHioz3/VYAiNwoSaGVsbG8gQWRh"}}
+  ],
+  "expected": {"command_type": "CompleteWorkflow", "result": "hello Ada"}
+}