hyperledger-labs / fabric-smart-client / 37775521433

91%

push

github

fix(comm)!: stop one non-draining session from stalling message dispatch (#2049)

The single dispatcher goroutine delivered to a session by blocking for up to a minute on its full queue, then closed the session inline and waited for its drain, so one session whose consumer stopped reading delayed delivery for every session on the node.

The dispatcher now never waits for a session's consumer. A message for a full session is dropped and counted, and the session is marked closing before the dispatcher moves on and finishes closing in the background, so no later message is delivered after the dropped one. The master session is never closed this way, since closing it would stop the node accepting new sessions.

A closing session takes no new stream lease, from the dispatcher, sendWithCachedStreams or sendTo, since it releases its streams only once. A Close that finds the session already closing waits for that close to finish and releases the session's streams, so no stream stays leased after Close returns.

DeleteSession closes the session after releasing the registry lock. Closing waits for the session's consumer, and the dispatcher takes that lock for every incoming message.

The dispatcher tests run through P2PNode.Start with the production worker count.

Docs: describe what fsc.p2p.incomingMessagesBufferSize sizes and its default, note that a closed session stays registered until its view context is disposed, and update the metrics file list.

Fixes #1987

BREAKING CHANGE: DefaultEnqueueTimeout, DefaultDispatcherTimeout and NetworkStreamSession.SetEnqueueTimeout are removed. Session delivery no longer waits on a full queue.

Signed-off-by: Said Altury <Said.Altury@ibm.com>

37 of 38 new or added lines in 3 files covered. (97.37%)

29101 of 32078 relevant lines covered (90.72%)

2604.29 hits per line