Co-operative way to cancel request execution

[rstoyanchev]

If the underlying transport times out, or the client goes away or network is lost, it makes sense to cancel execution and prevent further DataFetcher invocations.

There is graphql.execution.AbortExecutionException that can be raised to that effect, but requires wrapping DataFetcher's to protect their invocation. It would be useful to have a more built-in option that provides a way to cancel and stop the engine from further processing.

For some further context, see spring-graphql#450 and spring-graphql#1153.

[bbakerman]

See #3880

[bclozel]

Additional context here as well: spring-projects/spring-graphql#1094 (comment)

As of Spring for GraphQL 1.3.5-SNAPSHOT we're supporting the cancellation of subscription publishers as well. But developers are still seeing exceptions at runtime. It looks like the main publisher is cancelled but in-flight data fetchers fail when they complete. I'm not sure this is something we can fix at all, or refine the exception handling.

I see that this PR attempts to cancel things before the Publisher is returned, but we should ensure that it is itself cancelled if publishing is already happening.

[bbakerman]

I see that this PR attempts to cancel things before the Publisher is returned, but we should ensure that it is itself cancelled if publishing is already happening.

I have added another test to the PR to show that the Publisher subscription is cancelled and the Subscriber goes into onError state.

Reactive semantics say that IF an exception happens on a publisher,
then onError is called and no more messages flow. But since the exception here happens
during the mapping, the upstream publisher does not know about this.

So we cancel the subscription to bring the semantics back together, that is as soon as an exception
has happened, no more messages flow.

This should cover most cases. In practice some Publisher may ignore the Subscription.cancel but there is not a lot we can do there

[bclozel]

Thanks @bbakerman !

[bbakerman]

The cancel operation support is not working as expected. At this late stage in the release process we think we should remove it and then re-look at it once we have released.

See #3889 and #3888

I am reopening this issue as its not done