* Use separate stream handler builders for easy override and testing.
* Fix worker pool execution that was silently failing since it only
expected connection receivers.
Sparked by code in #24162.
* Introduce a connection coder responsible for encoding Cable messages
as WebSocket messages, defaulting to `ActiveSupport::JSON` and duck-
typing to any object responding to `#encode` and `#decode`.
* Consolidate encoding responsibility to the connection. No longer
explicitly JSON-encode from channels or other sources. Pass Cable
messages as Hashes to `#transmit` and rely on it to encode.
* Introduce stream encoders responsible for decoding pubsub messages.
Preserve the currently raw encoding, but make it easy to use JSON.
Same duck type as the connection encoder.
* Revert recent data normalization/quoting (#23649) which treated
`identifier` and `data` values as nested JSON objects rather than as
opaque JSON-encoded strings. That dealt us an awkward hand where we'd
decode JSON strings… or not, but always encode as JSON. Embedding
JSON object values directly is preferably, no extra JSON encoding,
but that should be a purposeful protocol version change rather than
ambiguously, inadvertently supporting multiple message formats.
This change makes ping into a message type, which
makes the whole protocol a lot more consistent.
Also fixes hacks on the client side to make this all
work.
Before this commit, the `unsubscribed` callbacks in Action Cable server
side channels were never called. This is because when a WebSocket
"goodbye" message was sent from the client, the Action Cable server
didn't properly clean up after the now closed WebSocket. This means that
memory could possibly skyrocket with this behavior, since part of this
commit is to properly remove closed subscriptions from the global
subscriptions hash. Say you have 10,000 users currently connected, and
then all 10,000 disconnect -- before this patch, Action Cable would
still hold onto information (and Ruby objects!) for all of these now
dead connections.
The `WorkerTest`'s `Receiver` is imporsonating an `ActionCable::Connection::Base`, but
just delegates the logger to `ActionCable.logger`.
This creates a mismatch as the connection requires the logger to be a
`TaggedLoggerProxy`'ied logger, while the server doesn't.
Thus to ensure an exception isn't raised when the worker tries to call `tag`
other tests have to assign a proxied logger to their test server.
Instead of forcing change on other tests, have Receiver adhere to the connection
contract and use a `TaggedLoggerProxy`.
As a consequence remove more setup from the tests.
Instead of depending on ApplicationCable::Connection being defined at initialize
we should inject it in the Railtie.
Thus we can kill more setup in the tests too.
We were explicitly referencing Rails.root in ActionCable::Server::Configuration.initialize,
thereby coupling ourselves to Rails.
Instead add `app/channels` to Rails' app paths and assign the existent files
to `channel_paths`.
Users can still append to those load paths with `<<` and `push` in `config/application.rb`.
This means we can remove the custom `Dir` lookup in `channel_paths` and the Rails
and root definitions in the tests.
Also, address the possibility of the listen thread dying and needing to
be respawned. As a bonus, we now defer construction of the thread until
we are first given something to monitor.
