Actors: a call cancelled while it runs frees the turn in daprd while the method is still executing, so the next call overlaps (port of dapr/dotnet-sdk#1937)
メンテナーはふだん 2 日以内に返信
まだ誰も着手していません。
評価
- 難易度
- 4/5
- 見積もり時間
- 3〜5日
- 初心者へのやさしさ
- 47/100
- issue の種類
- バグ
- 明瞭さ
- 明確に書かれている
- 活発さ
- 活発
調査の方向性
Start with ActorManager._dispatch_internal and trace how _on_pre_actor_method_internal, the actor method, and _on_post_actor_method_internal run for the same actor ID. Compare the proposed turn-lock behavior with the linked .NET ActorTurnLock and ActorManager; the issue describes the required cancellation and reentrancy cases. Done means actor turns stay serialized through state saving across HTTP and gRPC, including reminders, timers, and deactivation.
索引モデルが issue の本文から書いたものです。
説明
Summary
When a caller cancels an actor call while it's running, daprd releases that actor's turn and starts the next queued call, but this SDK keeps running the cancelled call's actor method. Two turns then run on the same actor ID at once, which breaks turn-based concurrency. It happens with reentrancy off as well as on.
The runtime alone can't prevent this. The .NET SDK fixed it on its side in dapr/dotnet-sdk#1937 (released in Dapr.Actors 1.19.0) by holding the turn in the app. This SDK has no equivalent, as far as we can see.
How it happens
- daprd takes the actor's lock, calls the app with the caller's context, and releases the lock when that call returns (
app.InvokeMethod). - If the caller cancels, the call to the app returns at once and the lock is released, even though the app is still running the method:
- Over HTTP, daprd aborts the request.
- Over the gRPC actor callback stream, daprd stops waiting for the response and sends the app nothing (
callbackstream.go).
- daprd then sends the next queued call for that actor ID. Neither an aborted HTTP request nor an abandoned stream call stops the actor method, so it overlaps with the new call.
dapr/dapr#10835 reported the overlap. dapr/dapr#10842, in runtime 1.19.0, fixed a related case: a call cancelled while still queued no longer evicts the running turn. A cancelled running call still releases the turn early, because daprd can't know when the app finishes once it stops waiting.
Evidence
From the standalone reproduction. Each trial sends 20 calls × 2s to one actor ID, cancels them all at a fixed time, and counts in-progress turns inside the actor. Kubernetes (kind), .NET actor host:
| Runtime | .NET SDK | Reentrancy off | Reentrancy on |
|---|---|---|---|
| 1.18.7 | 1.18.10 | 8 of 12 trials overlapped (up to 7 turns at once) | 7 of 12 (up to 9) |
| 1.19.0 | 1.18.10 (no app-side lock) | 3 of 12 (up to 2) | 4 of 12 (up to 4) |
| 1.19.0 | 1.19.0 (with #1937) | 0 of 12 | 0 of 12 |
We measured this with the .NET SDK only and haven't run the reproduction against this SDK. The code below has no lock that would prevent it, though.
Where (python-sdk)
ActorManager._dispatch_internal holds _active_actors_lock only to look up the actor. _on_pre_actor_method_internal, the method and _on_post_actor_method_internal (the state save) run outside any per-actor lock. Two coroutines for the same actor ID can therefore interleave at any await, over both HTTP (FastAPI/Flask extensions) and the gRPC callback host (grpc_host.py).
Suggested fix
Port the .NET approach: a per-actor turn lock in the SDK (ActorTurnLock, acquired in ActorManager).
- Acquire it per actor ID before activation and the method call.
- Hold it until the actor code has finished, including the post-method state save, whether or not the request was cancelled.
- When reentrancy is enabled, let a call with the same non-empty reentrancy ID as the running turn enter.
- Apply it to reminders, timers and deactivation as well as method calls.
Its own comment states the rule: "The sidecar can stop waiting before actor code finishes. Keep local ownership until every invocation in the turn has completed, not until its request is cancelled."
- 主要言語
- Python
- スター
- 274
- フォーク
- 162
- 平均マージ
- 2日 2時間
- マージ済み PR（30日）
- 71
環境構築
このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。
- Dockerfile・Docker Compose ファイルなし
- プルリクエストのテンプレートあり
- コントリビューションガイドを読む
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
dapr/python-sdk のほかの issue
-
Add the OpenSSF Scorecard GitHub Action and publish the results対応中かも @AdepuSriCharan が 7 日前に担当しました。 オープンgood first issue help wanted
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
dapr/python-sdk#1332 ·
メンテナーはふだん 2 日以内に返信
-
SDK Alignment of Workflow Examples再び着手できるかも このイシューのプルリクエストはマージされずにクローズされました。 オープンdapr-ext-workflow good first issue kind/enhancement P2
難易度 2/5 1〜2日 初心者へのやさしさ 72/100
dapr/python-sdk#853 · コメント 4 件 ·
メンテナーはふだん 2 日以内に返信
-
Add PyPI keywords and project links to the dapr package対応中かも @SpinachDigital が 8 日前に担当しました。 オープンgood first issue help wanted kind/enhancement
難易度 2/5 1〜3時間 初心者へのやさしさ 20/100
dapr/python-sdk#1335 · コメント 1 件 ·
メンテナーはふだん 2 日以内に返信
-
ci/release-failure release
難易度 2/5 1〜3時間 初心者へのやさしさ 25/100
dapr/python-sdk#1320 ·
メンテナーはふだん 2 日以内に返信
-
ci/release-failure release
難易度 2/5 1〜3時間 初心者へのやさしさ 49/100
dapr/python-sdk#1311 ·
メンテナーはふだん 2 日以内に返信
dapr/python-sdk の issue をすべて見る
似ている issue
-
enhancement help wanted
難易度 2/5 1〜3時間 初心者へのやさしさ 76/100
Git-Hospital/githealthcheck#61 ·
メンテナーはふだん 1 日以内に返信
-
changelog investigate
難易度 2/5 1〜3時間 初心者へのやさしさ 67/100
ramnes/notion-sdk-py#433 ·
-
Partial method descriptors raise on class access対応中かも @Gambit-Checkmate が今日担当しました。 オープン
難易度 2/5 1〜3時間 初心者へのやさしさ 76/100
メンテナーはふだん 19 日以内に返信
-
docs pkg:core
難易度 1/5 1時間未満 初心者へのやさしさ 84/100
pydantic/pydantic-ai#11929 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
intake/intake-esm#789 ·