[miniflare] Hyperdrive proxy crashes the Node process when the client socket errors

[Pduhard]

What versions & operating system are you using?

  System:
    OS: Linux 7.0 Ubuntu 24.04.4 LTS 24.04.4 LTS (Noble Numbat)
    CPU: (32) x64 AMD Ryzen 9 9950X 16-Core Processor
    Memory: 16.45 GB / 30.45 GB
    Container: Yes
    Shell: 5.9 - /usr/bin/zsh
  Binaries:
    Node: 22.14.0
    npm: 10.9.2
    pnpm: 11.24.0
  npmPackages:
    miniflare: 4.20260730.0 => 4.20260730.0

Please provide a link to a minimal reproduction

https://github.com/Pduhard/miniflare-hyperdrive-crash-repro

Describe the Bug

The local Hyperdrive proxy pipes the workerd client socket to the database socket. Only the database side gets an error listener.

In packages/miniflare/src/plugins/hyperdrive/hyperdrive-proxy.ts on main:

• createPlainTCPConnection attaches dbSocket.on("error", ...)
• setupTLSConnection attaches tlsSocket.on("error", ...)
• clientSocket gets nothing, at any of the four places it is piped: #handleConnection, handlePostgresTlsConnection, handleMySQLTlsConnection, setupTLSConnection. It also gets nothing before those pipes exist: on the TLS paths, #handleConnection writes to the database, reads the reply and runs a TLS handshake first, and a client that resets during that window crashes the process the same way

pipe() does not cover this. It attaches an onerror handler to the destination, but that handler removes itself and re-emits once no other listener is left. So when the client socket errors after the pipes are set up, nothing is listening, and the whole Node process exits. That takes down wrangler dev, getPlatformProxy(), or a vitest run.

Steps to reproduce

git clone https://github.com/Pduhard/miniflare-hyperdrive-crash-repro
cd miniflare-hyperdrive-crash-repro
npm install
npm run repro:half-close
npm run repro:reset

repro:half-close runs a TCP server that speaks just enough of the Postgres protocol to negotiate TLS, then streams data. A Worker connects through a Hyperdrive binding and half-closes its socket while that data is still arriving. Nothing reaches into Miniflare internals.

repro:reset covers the other window: it talks to the proxy port directly and resets the connection while the proxy is still negotiating TLS with the database, before any pipe exists.

Expected: the connection is torn down and Miniflare keeps running.
Observed: the process exits 1 on an uncaught error.

Which configurations reach this code

Since #13390 the proxy is skipped when sslmode=disable, so that mode no longer reaches it. require, prefer, verify-ca and verify-full still do, which is the ordinary setup for developing locally against a hosted Postgres that requires TLS.

version	sslmode=disable	sslmode=require
4.20260424.0	exits 1	exits 1
4.20260730.0	20 of 20 attempts survive	exits 1

The same code is on main today.

Please provide any relevant error logs

node:events:496
      throw er; // Unhandled 'error' event
      ^

Error: write EPIPE
    at WriteWrap.onWriteComplete [as oncomplete] (node:internal/stream_base_commons:87:19)
Emitted 'error' event on Socket instance at:
    at Socket.onerror (node:internal/streams/readable:1028:14)
    at Socket.emit (node:events:518:28)
    at emitErrorNT (node:internal/streams/destroy:170:8)
    at emitErrorCloseNT (node:internal/streams/destroy:129:3)
    at process.processTicksAndRejections (node:internal/process/task_queues:90:21) {
  errno: -32,
  code: 'EPIPE',
  syscall: 'write'
}

Socket.onerror at readable.js:1028 is the listener pipe() installed on the destination, re-emitting because nothing else was listening. Depending on whether the write lands just before or just after the peer's FIN, the same failure also shows up as Error: This socket has been ended by the other party at Socket.writeAfterFIN.

---

PR: #15769. It attaches the listener as soon as the connection is accepted and at each pipe site, with a regression test for each of the two windows in packages/miniflare/test/plugins/hyperdrive/proxy.spec.ts. Both tests fail on main.

[workers-devprod]

Automated triage

Issue assessment

• Actionability: ✅ Actionable.
• Type: Bug.
• Reproduction: Automatic attempt requested.
• Labels:
  • package:miniflare (high)
  • product:hyperdrive (high)

Actionability and missing information

Miniflare's Hyperdrive proxy lacks an error listener on the client socket at its four pipe sites, so a client-side socket error becomes an unhandled error event that crashes the Node process; report includes minimal repro, versions, environment, and error logs.

Reproduction reason: The report provides concrete steps, version comparisons, and consistent error output for a crash that still occurs in current releases, so automatic reproduction can validate it.

Reproduction

• Result: ✅ Reproduced.
• Summary: Reproduced [miniflare] Hyperdrive proxy crashes the Node process when the client socket errors #15768 with miniflare 4.20260730.0: a Worker closing its Hyperdrive socket mid-stream while a mock Postgres TLS server (sslmode=require) streams crashes the Node process running the Hyperdrive proxy with an uncaught EPIPE on the client socket, matching the reported exit-1 crash.

Evidence and attempted steps

Steps attempted:

1. Set up the fixture project with the exact reported dependency version
   • Command: cd project && pnpm install # pins miniflare 4.20260730.0 exactly
2. Run the repro: streaming mock Postgres TLS server on 127.0.0.1:15432; Miniflare Hyperdrive binding with sslmode=require (proxy in-process, same wiring wrangler dev uses); dispatch a Worker fetch whose env.DB.connect() socket half-closes and closes mid-stream while the mock database keeps streaming
   • Command: node repro.mjs
3. Verify reliability across consecutive runs
   • Command: for i in 1 2 3 4 5; do node repro.mjs > /tmp/final-$i.log 2>&1; echo "run $i: exit=$?"; done
4. Verify the shipped artifact end-to-end from a clean state
   • Command: rm -rf node_modules && pnpm install --frozen-lockfile && pnpm reproduce

Expected behavior: The client-side connection is torn down gracefully and the Node process running Miniflare (wrangler dev / getPlatformProxy() / a Miniflare Node-API process) keeps running; the socket error is handled rather than crashing the process.

Observed behavior: In 5/5 consecutive runs (plus one clean-state run after a frozen reinstall), the Node process running Miniflare exited 1 during the Worker request with the exact reported signature: an uncaught 'error' event on the Hyperdrive proxy's client socket — "Error: write EPIPE" (errno: -32, syscall: 'write'), "Emitted 'error' event on Socket instance at: at Socket.onerror (node:internal/streams/readable:1028:14)" — the same re-emit site (readable.js:1028, the listener installed by pipe()) and same error as the issue's log.

Artifact: Download reproduction archive (Cloudflare Access required).

Treat archive contents as untrusted. Inspect them before running anything.

---

This summary was generated by an LLM and may contain errors. Please verify its findings.

Triage bot commands and options

Repository collaborators with write or admin access can trigger actions by commenting on this issue:

• @workers-devprod reproduce: Attempt to reproduce the issue using its description and recent comments.