This time with less bugs. It turned out that the crash was due to a
synchronisation issue between the pending reads and the actual shard
polling mechanism.
Essentially the pending reads would be cancelled via a simple bool but
there would still be a pass left and thus we would have a single
pending read left before or after running the polling mechanism and
this would cause a race condition.
Now the pending read mechanism is properly waited for before returning
control back to the caller.
This should allow an easier way to query the Discord protocol gateway
latency, defined by the difference HEARTBEAT_ACK between and the last
sent HEARTBEAT.
This setting configures how long before a timeout event is emitted
internally and disconnects the websocket. Since some users were
experiencing issues with the gateway not responding, this should help
mitigate the issue for those with poor PCs.
This is a complete redesign of the old voice code.
A list of major changes is as follows:
* The voice websocket will now automatically reconnect with
exponential back-off just like the regular Client does.
* Removal of the stream player concept.
* Audio now gracefully pauses and resumes when a disconnect is found.
* Introduce a discord.AudioSource concept to abstract streams
* Flatten previous stream player functionality with the
VoiceClient, e.g. player.stop() is now voice_client.stop()
* With the above re-coupling this means you no longer have to
store players anywhere.
* The after function now requires a single parameter, the error,
if any existed. This will typically be None.
A lot of this design is experimental.
This shows the Shard ID in more places, along with a gateway trace and
session ID. Also helps show the RESUME/IDENTIFY/RESUMED/READY flow a
bit more instead of it looking like the connection has zombied out.
Apparently we should always try to RESUME first and if we get
INVALIDATE_SESSION then we should IDENTIFY instead. This is the
preferred way to do RESUMEs.
The library worked with the assumption that whenever the recv call for
the websocket would lead to a closure, the close method would be called
to signal closure and as a result our close method would be called.
This assumption turned out to be false as the websockets library would
instead call an internal function named close_connection instead. So to
solve our problem we need to override this function instead of close.
This should have a more uniform approach to rate limit handling. Instead
of queueing every request, wait until we receive a 429 and then block
the requesting bucket until we're done being rate limited. This should
reduce the number of 429s done by the API significantly (about 66% avg).
This also consistently checks for 502 retries across all requests.
Tobotimus
Rapptz
khazhyk
cookie
Jake