Deadlock when using subscriptions with two Ktor worker threads

[micke-humla]

Library Version
7.0.2

Describe the bug
A deadlock occurs when calling subscriptions in graphql-kotlin-ktor-server if a suspendable function is called (that actually suspends execution) when there are only two workers (workerGroupSize is set to 2 in application.conf or using default settings on a machine with 2 CPUs). Tested with Netty as engine.

Explicitly setting workerGroupSize to 3 or higher seems to fix the problem but might not be optimal on machines with only 2 CPUs. Problem surfaced on AWS Elastic Container Service with ECS instance reporting to have 2 CPUs (even with different vCPU values).

To Reproduce
I created a minimal reproducible example here and this is how it can be reproduced:

1. Start the application with ./gradlew run
2. Connect to the websocket at ws://localhost:8080/subscriptions
3. Send init message: {"type":"connection_init"}
4. Send subscribe message: {"type":"subscribe", "id":"12345", "payload": {"query": "subscription { random }" } }
5. No messages are returned and subscription "hangs" and the worker thread (eventLoopGroupProxy) has state WAITING

Expected behavior
Subscriptions work even with "only" two worker threads

[dariuszkuc]

Hello 👋
I'm guessing the underlying issue is probably the same as with the #1898. If you change GraphQLServer to switch threads for processing requests then it should work even if your Ktor engine starts with just 2 worker threads.

class MyKtorGraphQLServer(
    requestParser: KtorGraphQLRequestParser,
    contextFactory: KtorGraphQLContextFactory,
    requestHandler: GraphQLRequestHandler
) : KtorGraphQLServer(requestParser, contextFactory, requestHandler) {
    override suspend fun execute(request: ServerRequest): GraphQLServerResponse? = withContext(Dispatchers.Default) {
        super.execute(request)
    }
}

[micke-humla]

Thanks for the reply! Looks promising but unfortunately I find no way to override KtorGraphQLServer (without also creating my custom com.expediagroup.graphql.server.ktor.GraphQL class with all it's logic) so I can't verify if it solves the issue.

Also, would that really solve the problem for subscriptions which is using KtorGraphQLWebSocketServer (which is also not possible to override)?

[dariuszkuc]

Thanks for the reply! Looks promising but unfortunately I find no way to override KtorGraphQLServer (without also creating my custom com.expediagroup.graphql.server.ktor.GraphQL class with all it's logic) so I can't verify if it solves the issue.

Good callout. I'll open up a separate issue to allow folks specifying their own GraphQL server impls.

[dariuszkuc]

Opened up #1902