WebSocket transport does not allow healing connections

[tom-sherman]

In other words, the session does not allow the websocket to reopen after it closes - the session is immediately aborted.

A way to reproduce this is with using reconnecting-websocket or partysocket.

import { newWebSocketRpcSession } from "capnweb";
import { WebSocket as ReconnectingWebsocket } from "partysocket";

const sesssion = newWebSocketRpcSession(
  new ReconnectingWebsocket("ws://localhost:8001/rpc")
);

Is this intentional behaviour of the default websocket transport? If so, would a new ReconnectingWebsocket transport be acceptable or should this be left to userland? I'm not sure what your thoughts are about adding new transports to core.

[kentonv]

Hmm, what are you imagining should happen to the session state on a reconnect? That is, what happens to stubs that were passed over the connection before the disconnect?

Possibilities include:

• They are all just broken. Only the "main" stub is reconnected.
• The client reconnects to the same server as before, where the session state still exists in memory, so that it can be resumed without losing anything.
• On reconnect, the system automatically recreates all stubs by replaying the same calls that created them the first time.

[LoganDark]

I can't even tell when the WebSocket connection breaks. If I could, then I'd be able to call newWebSocketRpcSession again, get a new stub, and do my own replay.

I can listen to the WebSocket's close and error events, but that doesn't catch all cases where the (private) transport calls #recievedError.

I had to make my own transport for the server side (because of Bun), so it looks like I'll have to do that for the client too.

Edit: Looks like onRpcBroken can be used for this.

[kentonv]

Yes, you can use onRpcBroken() to detect when the connection is lost. And from that you should be able to implement various approaches to handling this.

I could imagine having something more batteries-included, but it's not entirely clear what the behavior should be (see my previous comment).