fix(rt): coroutine leak in ktor engine #628

aajtodd · 2022-04-18T16:31:15Z

Issue #

Description of changes

Fixes a coroutine leak in the ktor engine due to invalid parent-child
relationship. The ktor engine was launching a coroutine as a child of
the request job (calLContext()). The issue was that this child
coroutine was relying on the parent job completion notification
to exit. Parent jobs can't complete though until all their children
do so the child coroutine never completed.

The ktor engine now uses the engine scope to launch this child
coroutine and establishes no parent-child relationship to the
request job explicitly.

Testing

fun main(): Unit = runBlocking {
    val executor = newFixedThreadPoolContext(1, "poll-thread")
    SqsClient { region = "us-east-1" }.use { sqs ->
        withContext(executor) {
            while(true) {
                pollQueue(sqs)
                delay(1.seconds)
            }
        }
    }

}

suspend fun pollQueue(sqs: SqsClient) {
    println("polling queue")
    val resp = sqs.receiveMessage {
        queueUrl = QUEUE_URL
        // waitTimeSeconds = 10
    }

    println("response:")
    println(resp)

    resp.messages?.forEach { m ->
        sqs.deleteMessage {
            queueUrl = QUEUE_URL
            receiptHandle = m.receiptHandle
        }
    }
}

This code polls SQS on a 1-sec interval. Before this fix you could see DefaultSqsClient slowly become the largest object (30-45+min of running):

After this fix (and after ~45 min of running) the SQS client is the same size it was when the program started:

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

Fixes a coroutine leak in the ktor engine due to invalid parent-child relationship. The ktor engine was launching a coroutine as a child of the request job (`calLContext()`). The issue was that this child coroutine was relying on the parent job completion notification to exit. Parent jobs can't complete though until all their children do so the child coroutine never completed. The ktor engine now uses the engine scope to launch this child coroutine and establishes no parent-child relationship to the request job explicitly. fixes: awslabs/aws-sdk-kotlin#587