Why

River Schema v2 compatibility so we can start getting rid of the compatibility layer.

What changed

Client

Significantly restructured the river.v2.*.

• Nearly all asyncio.Lock's have been removed
• Background tasks are now cleaned up correctly, which should reduce the amount of noise around tasks being GC'd while they're still running
• Move websocket state ownership from Client transport down into Client Session. This significantly simplifies transparent reconnects, since we don't need to coordinate state shifting.
• There are four background tasks:
  • Heartbeat checker. This keeps track of whether we've gotten any heartbeats from the server recently.
  • "Close session" checker. This keeps track of whether we've exceeded the grace period for transparent reconnects.
  • Buffered message sender. One downside with the previous implementation was that we had a race on sending messages over the websocket, since we had no coordination between calls to send_message when called concurrently, so we'd end up with seq bound and incremented but then messages would become out-of-order by the time the ws.send method returned. Now we just atomically append to a deque as we push onto the send buffer, and when we get a confirmation from the ws library we move the message over to an ack buffer (also a deque) waiting for incoming messages from the server to allow us to drain.

Codegen

• init is now required for all method types, input is now optional. This simplifies rpc and stream codegen.
• Fixed some bugs where init was being used in input position and vice versa.
• Swapping f"'{foo}'" encoding for f"{repr(foo)}", giving greater safety post-generation (to avoid situations where f"'{None}'" gets rendered as "None". f"{repr(None)}" would render as None which would fail typechecks.)

Test plan

Manually ran codegen against v2 generated schemas, everything typechecks.