can‘t connect to ws server in v0.7.0

[huahuayu]

Env

cronos: 0.6.5 and 0.7.0

Issue

Can't subscribe newHead and newTxs:

SubscribeNewHeads / newTxs from golang

dial tcp <my_ip>:8546: connect: connection reset by peer ws://<my_ip>:8546

Subscribe from wscat

wscat -c ws://<my_ip>:8546
error: connect ECONNRESET <my_ip>:8546

Behavior in v0.6.5

You can connect to ws after restarting, it works only 5-10 mins, then you get error and need another restart.

Behavior in v0.7.0

Today I upgrade to v0.7.0, the longest record I have is about 1 hour, within that 1 hour, I can subscribe newHead and newTx, I test it many times, so I thought it is been fixed in v0.7.0.

But when I tried again just now, I can't connect ws server anymore. Even restarting cronosd is useless.

The issue is still there.

@yihuang Please help to check, thx.

[huahuayu]

What's the problem do you think, I am willing to dig into the issue, please share your findings.

[huahuayu]

My findings

In tendermint/state/txindex/indexer_service.go use unbuffered channel for blockHead and tx subscribe, which may block the channel

	blockHeadersSub, err := is.eventBus.SubscribeUnbuffered(
		context.Background(),
		subscriber,
		types.EventQueryNewBlockHeader)
	if err != nil {
		return err
	}

	txsSub, err := is.eventBus.SubscribeUnbuffered(context.Background(), subscriber, types.EventQueryTx)
	if err != nil {
		return err
	}

In tendermint/libs/pubsub/pubsub.go I do observed send event msg get blocked, the logic goes to --> mark.

func (state *state) send(msg interface{}, events map[string][]string) error {
	for qStr, clientSubscriptions := range state.subscriptions {
		q := state.queries[qStr].q

		match, err := q.Matches(events)
		if err != nil {
			return fmt.Errorf("failed to match against query %s: %w", q.String(), err)
		}

		if match {
			for clientID, subscription := range clientSubscriptions {
				if cap(subscription.out) == 0 {
					// block on unbuffered channel
-->					subscription.out <- NewMessage(msg, events)
				} else {
					// don't block on buffered channels
					select {
					case subscription.out <- NewMessage(msg, events):
					default:
						state.remove(clientID, qStr, ErrOutOfCapacity)
					}
				}
			}
		}
	}

	return nil
}

Solution

I changed blockHeadersSub, err := is.eventBus.SubscribeUnbuffered and txsSub, err := is.eventBus.SubscribeUnbuffered to buffered channel, so far so good, let me keep observing for a while.

[huahuayu]

Three days passed, still works. @yihuang

[yihuang]

Three days passed, still works. @yihuang

awsome, so the issue is dead lock on unbuffered channel? Can you open a PR to tendermint directly?

[huahuayu]

I don't know how to effectively reproduce the issue and am not sure if there are other side effects.

[JayT106]

Hi @huahuayu, I think the block on unbuffered channel was designed for the indexer services in Tendermint, it guaranteed that every event will be processed to the indexer. If the indexer has a heavy I/O loading, it will blocks the pubsub module temporarily for sure.

What's your experimental_websocket_write_buffer_size and experimental_subscription_buffer_size in config.toml?
it shouldn't be 0 then you will get a buffered channels subscription.

Do you need to use indexer service from the node? maybe you can set it to null, and to see if this issue still happens.

[yihuang]

I think ws server issue will eventually be fixed by this solution:#665