Cancel resolver tasks if execution of an operation is terminated

まず、非同期 GraphQL オペレーションの実行を追跡し、nullable でない resolver が失敗したときに resolver タスクがどのようにスケジュールされるかを確認します。issue #3414 に記載された Strawberry のシナリオを再現します。実行が終了する前に不要になった resolver タスクがキャンセルされ、レスポンス後またはライフサイクルのクリーンアップ後も実行を継続しないことを確認できれば完了です。

TL;DR

Would it be possible to cancel outstanding resolver tasks as soon as before execution of a GraphQL operation is terminated? Outstanding / unused tasks can happen as a result of an incoercible field resolver error.

These outstanding tasks can run well after lifecycle hooks responsible for cleaning up their context have completed. As a result, they often fail and cause a lot of noisy errors for us.

More details

Hello!

I'm coming from the strawberry-graphql project. We're running into a problem with abandoned resolver tasks not being cancelled, and thus failing in unpredictable ways.

Strawberry has an extensions framework which allows you to wrap the execution phase of your operation processing in hooks – running code before and after execution of resolvers. We use these hooks to manage (set up and tear down) some state needed by the resolvers.

However, when a resolver of a non-nullable field fails, it may lead to the results of some other resolvers being no longer necessary. It appears that graphql-core will then short-circuit returning a result, while the no-longer-necessary resolvers are still running. We see that they can continue running well after the HTTP response was returned.

The problem is that Strawberry's lifecycle hooks (and I also imagine - the webserver's) would have completed by that time. As a result, we do things like terminate the DB session – making it very likely that these outstanding tasks will fail with various esoteric errors. This causes a lot of noise in our observability suite – which we'd love to avoid.

See https://github.com/strawberry-graphql/strawberry/issues/3414 for more details, and https://github.com/strawberry-graphql/strawberry/issues/3414#issuecomment-2012060461 for the comment which recommends I report the issue here.