supabase / supabase-flutter / 39417868719

92%

push

github

fix(realtime): do not flush a join that its channel has since resent (#1986)

Resolves SDK-2280. This checks the two Realtime races fixed in
supabase/supabase-swift#1578 and supabase/supabase-swift#1585 against
the Dart client. The tests reproduced one bug, which this PR fixes.

## Fix

**A timed out join is no longer sent after the rejoin.** Calling
`subscribe()` while the socket is down puts the join into the client's
send buffer. If the outage outlasts the join timeout, the channel
errors. Once the socket reconnects, the channel is rejoined with a new
ref, and then the buffer was flushed, so the old join went out after the
new one. Phoenix answers a second join on a topic by shutting down the
first, so the server would close the channel that just rejoined and keep
a join ref the client ignores.

The send buffer now holds the messages instead of closures. When it is
flushed, a join that no channel on its topic is waiting for is dropped.
That covers a channel that resent its join or was removed. The path that
resolves a `customAccessToken` already cleared the buffer, so only
clients without one were affected.

## Already handled

These cases from the swift PRs needed no change. Each now has a
regression test:

- After a socket drop and rejoin, an error reply, a `phx_error` or the
timeout of the original join does not affect the rejoined channel and
adds no status. Resending the join removes the old reply binding and
cancels its timer, and channel events with an old join ref are dropped.
- `connect()` returns after one failed attempt during an outage, while
the reconnect timer keeps retrying. A `subscribe()` made during the
outage reports `timedOut` after its timeout instead of hanging, and
joins once the server is back.

## Tests

- New `test/stale_join_test.dart` with the five cases above. The outage
test fails without the fix.
- The buffered push test in `socket_test.dart` now buffers a
`phx_leave`, because a bare `phx_join` with no channel behind ... (continued)

11 of 11 new or added lines in 1 file covered. (100.0%)

10177 of 11057 relevant lines covered (92.04%)

5.49 hits per line