RedisChannelLayer.send channel capacity checks against real internal redis queue, instead of expected asgi channel

[thomas-riccardi]

ASGI Channels for the same RedisChannelLayer instance (

channels_redis/channels_redis/core.py

Line 239 in 54d935b

self.client_prefix,

) are factorized into one redis queue (also sometimes called a channel, which is confusing).
The ASGI channels capacity is check against the internal redis queue length, not the asgi channel.

Related to #95.

It is not only an issue if we have different rules per channel, but also and primarily an issue if we have many asgi channels for one RedisChannelLayer instance: some asgi channels may get ChannelFull because of other asgi channels being non-empty.

I'm not sure what to do to fix this, but it could at least be documented in the README (https://github.com/django/channels_redis#channel_capacity).

[andrewgodwin]

Documenting this in the README is the correct approach I feel. I don't think there's any other sensible way around this without throwing away the significant efficiency gains of reading from a single list.

[thomas-riccardi]

Do you propose to stop channel-specific capacity altogether ? (we should then remove the examples from the README)
Indeed I don't see any other reasonable way.

[andrewgodwin]

No, the backpressure is valuable to stop Redis filling up, and I think that part still works fine if it's shared across channels.