Reconnect same host plugin process and client

[amiramw]

Fixes problem of plugins not working after short disconnection.
Keep disconnected host plugin process alive for at least one minute and
reconnect it if the client is the same.

Signed-off-by: Amiram Wingarten [email protected]

[akosyakov]

We could accept it as a temporary fix. I don't like it much since it will cause other issue, i.e. you close a window but process tree is hanging for one minute occupying different ports and so on. So if someone open a new window within one minute and tries to launch something it is not going to work. Maybe timeout can be smaller like 5-10s. @amiramw for how long you have connection issues?

I also believe we already have general 30s timeout over websocket connection. Not sure that we need a specific one for plugin system. I would prefer that if websocket connection is gone then everything gone as well and is not kept for another minute.

[amiramw]

@akosyakov We are facing disconnections that take few seconds. I reduced timeout to 10 seconds and added an ENV variable so that we can override it if needed. If you want I can make it even lower. Also I fixed the other code review comments.
Even as temporary solution for 0.7.0 this would be a major improvement in theia productivity for us.
I think that for few seconds of disconnection it is good not to be forced to execute all plugins and vscode extensions again.
For later version I can take a look into a full solution that will work also after the timeout and also after theia restart.

[akosyakov]

I think that for few seconds of disconnection it is good not to be forced to execute all plugins and vscode extensions again.

It is error prone. If a ws connection is gone then some data can be lost which should have been sent, like text edits, so state of plugin on backend is different from frontend state leading to all kind of functionality bugs. If you want to really reuse backend plugins you should make sure that they sync state with frontend when a client reconnects. It is not an easy task as well.

Usually we preserve ws connection for 30s on network issues that data are delivered eventually. In this case it's possible to reuse plugins. If a ws connection is not preserved the easiest strategy is to kill a plugin and start a new one to make sure that it is in sync with frontend state.

On your network, apparently, somehow ws connection is got killed on frontend or backend too early by network errors.

[amiramw]

@akosyakov I see your point. Maybe timeout small enough can make these inconsistencies pretty rare.
We have an issue with the way nginx ingress controller handles route changes which cause short disconnection.
It can be reproduced locally with proxy server copied from here: #4590. Terminate host plugin process is triggered immediately when killing the proxy.

Anyway as a temporary solution for 0.7.0 I would appreciate if this could be merged. I'll try to investigate in parallel the second approach.

[amiramw]

@akosyakov @benoitf If this is acceptable as temporary solution I would be happy if you could merge it before 0.7.0 release today 🙏

[akosyakov]

ok, let's try ... i just afraid that we open up another can of issues here instead of solving the real issue