Deals must be manually restarted due to data transfer failures

[arajasek]

Edit: The original issue was about protecting connections lest we close them when cleaning up and interrupt a transfer. They should be unprotected when the transfer completes.
Edit: It seems that connections are already protected so the data transfer failures are being caused by a different issue. -@dirkmc

[arajasek]

Based on some discussion with @Stebalien and @hannahhoward, this is already happening, yet @whyrusleeping reports that the majority of deals he tried needed data transfer to be restarted because the connection was closed from the remote (miner) side.

[Stebalien]

Regardless of what's happening here, we should automatically reconnect.

[dirkmc]

I spent some time today digging into this issue with @aarshkshah1992. It seems that:

• We already are protecting the connection in go-data-transfer when opening push / pull channels
• We already are attempting to reconnect in go-graphsync when a message send fails

In order to resolve the issue I think we will need more information about why exactly the connection is being closed.

[Stebalien]

We believe the issue was yamux hitting a keepalive timeout. I believe this situation has gotten better since we fixed that bug.

[hannahhoward]

also: side note: the timeouts on reconnect in graphsync may not be long enough to overcome libp2p's dialing backoff and we may need to increase them.

[hannahhoward]

also I just realized the note about the disconnection is not resolved when we reconnect -- even if transfer resumes.

[dkkapur]

@dirkmc - fair to assume this can be closed with the fixes in markets v1.3?

[dirkmc]

Yes I'm closing, thanks Deep