Client.end never resolves if the socket was already destroyed

[mastermatt]

I've been debugging this issue kibae/pg-logical-replication#20 and narrowed it down to the Client's end method not resolving (or calling the callback) under certain error conditions.

Minimal reproduction:

const client = new Client(connOpts)
await client.connect()

setImmediate(() => {
  client.connection.stream.resetAndDestroy() // simulate a network outage or the db going offline e.g. ECONNRESET

  setImmediate(async () => {
    await client.end() // <-- never resolves
  })
})

Notably, by the time .end is called, the connection and its stream (a new.Socket instance) have already emitted their close and end events. However, .end relies on an end event from the connection to resolve, unless the connection never connected to begin with.
I believe the check to have .end resolve early should not only look at the _connecting flag, but should also check if the socket has been destroyed.

node-postgres/packages/pg/lib/client.js

Lines 605 to 606 in adbe86d

	// if we have never connected, then end is a noop, callback immediately
	if (!this.connection._connecting) {

I believe these issues are related:

1. Connection hang on the server side after ended #2329
2. Bug in pool.end () never returns when trying to close cleanly #2341
3. pg version 8.7.1 hangs on await db.end() but before version 8 doesn't #2648

[fmmoret]

Timing is uncanny -- ran into this as well today on "pg": "8.7.1"

[fmmoret]

Uncertain but this open PR might do it: https://github.com/brianc/node-postgres/pull/2717/files

[mastermatt]

@fmmoret yes, it would fix this. I'm surprised I didn't come across that one when I was searching for existing issues. The cause is described differently, but it's the same underlaying issues of the stream having already emitted its 'close' event.

[mastermatt]

My local testing shows this was resolved with 8.10.0.