Now that host is shared having separate config for listen and announce
addresses along with graphsync limit makes no sense since all of that is
shared with the markets host and datatransfer manager.
Remove the bespoke instantiation of libp2p host and datatransfer manager
for index provider and reuse the existing instances used by markets.
The rationale for reuse is the following:
1. Separation of host introduces a discovery problem, where without
gossipsub the index provider endpoint will not be discoverable.
Using the same host as markets would mean the chain can be used to
discover addresses, putting less empassis on criticality of
gossipsub considering its set-up cost and lack of message delivery
guarantees.
2. Only a single instance of graphsync/datatransfer can be instantiated
per libp2p host; therefore, if the host is shared, so should
datatransfer manager.
3. it is not clear if the assumptions under which separation was
decided still hold.
Add the peer ID of index provider host to the list of protected peers
before connecting to full node. Otherwise, it is possible for the
connection to be reset by full node before we reach the line that adds
the ID to list of protected peers via JsonRPC API.
Relates to:
- https://github.com/filecoin-project/index-provider/issues/177
Using the same pattern described in my previous commit.
I also added the CircuitBreaker to the itests kit as it may be useful
for other integration tests when debugging flakyness caused by timeouts.
The flake was caused by improper waiting for certain chain operations
to finish:
- We didn't wait for messages to be registered as pushed
- We improperly waited for a fixed time (10 seconds) for messages to be
mined, which in the best case would wait longer than necessary and in the
worst case would cause the test to break.
What I did:
- fixed by waiting in a loop for "just enough time". This fixed the flake
and made the test run faster, on average, because we don't have unnecessary
waiting.
- I added a "circuit-breaker" where the wait loop will timeout after 10 seconds.
deals_retry_deal_no_funds_test.go:TestDealsRetryLackOfFunds seems to be
flaky, sometimes it passes and sometimes it doesn't.
The first suspect for me was time.Sleep(), which I replaced with waiting
until a deal state where you can restart is reached.
