Watcher done callback may be invoked twice on unexpected connection loss

[bverhoeven]

Describe the bug

When a watch connection closes abruptly, the done callback may be invoked multiple times because of reentrancy in the doneCallOnce() function.

Client Version
1.1.2

Server Version
v1.32.2-gke.1182003

To Reproduce

1. Use Watch to start a stream (e.g. watch pods in a namespace).
2. Simulate a premature close on the connection (e.g. by destroying the socket on the server side).
3. The done callback is called with "AbortError: The user aborted a request"
4. The done callback is called again, but now with "Error: Premature close"

Expected behavior

The done callback should only be called once per watch termination.

Example Code

The code below can be used to replicate the issue on a local machine.

import http from 'http'
import { KubeConfig, Watch } from '@kubernetes/client-node'

const MOCK_PORT = 8333
const MOCK_URL = `http://localhost:${MOCK_PORT}`

const minimalServer = http.createServer((req, res) => {
    res.writeHead(200, {
        'Content-Type': 'application/json',
        'Transfer-Encoding': 'chunked',
    })

    res.flushHeaders()
    res.destroy() // Prematurely close the connection
})

minimalServer.listen(MOCK_PORT, () => {
    const kubeConfig = new KubeConfig()

    kubeConfig.loadFromClusterAndUser(
        {
            name: 'mock-cluster',
            server: MOCK_URL,
            skipTLSVerify: true,
        },
        { name: 'mock-user' }
    )

    const watch = new Watch(kubeConfig)

    watch.watch(
        `/api/v1/namespaces/default/pods`,
        {},
        () => { },
        (err) => console.log('done()', err)
    )
})

Output:

done() AbortError: The user aborted a request.
    at abort (/home/bas/projects/watch-repro/node_modules/node-fetch/lib/index.js:1458:16)
    at AbortSignal.abortAndFinalize (/home/bas/projects/watch-repro/node_modules/node-fetch/lib/index.js:1473:4)
    at [nodejs.internal.kHybridDispatch] (node:internal/event_target:827:20)
    at AbortSignal.dispatchEvent (node:internal/event_target:762:26)
    at runAbort (node:internal/abort_controller:449:10)
    at abortSignal (node:internal/abort_controller:435:3)
    at AbortController.abort (node:internal/abort_controller:468:5)
    at PassThrough.doneCallOnce (file:///home/bas/projects/watch-repro/node_modules/@kubernetes/client-node/dist/watch.js:33:28)
    at PassThrough.emit (node:events:519:35)
    at emitErrorNT (node:internal/streams/destroy:170:8) {
  type: 'aborted'
}
done() Error: Premature close
    at IncomingMessage.<anonymous> (/home/bas/projects/watch-repro/node_modules/node-fetch/lib/index.js:1748:18)
    at Object.onceWrapper (node:events:621:28)
    at IncomingMessage.emit (node:events:507:28)
    at emitCloseNT (node:internal/streams/destroy:148:10)
    at process.processTicksAndRejections (node:internal/process/task_queues:89:21) {
  code: 'ERR_STREAM_PREMATURE_CLOSE'
}

Environment (please complete the following information):

• OS: Linux
• Node.js version 23.11.0
• Cloud runtime: GCP

Additional context

This was discovered while debugging a broader issue with ListWatch failing to reconnect on non-410 errors (#2385). In those cases, doneCallOnce emits two errors:

• AbortError: The user aborted a request.
• Error: Premature close

Although doneCallOnce is intended to emit an error only once, it currently calls controller.abort() before setting doneCalled = true. This seems to trigger an AbortError in certain circumstances, which leads to a second invocation of doneCallOnce, since the doneCalled flag hasn’t been set yet.

As a result, both the original error and the AbortError are reported.

javascript/src/watch.ts

Lines 46 to 53 in 88184dc

	let doneCalled: boolean = false;
	const doneCallOnce = (err: any) => {
	if (!doneCalled) {
	controller.abort();
	doneCalled = true;
	done(err);
	}
	};

I was asked to create a separate issue and am working on a PR to add tests and fix the issue.

[bverhoeven]

Fixed in #2389