node v20, closing server and re-listening between requests causes ECONNRESET

[isaacs]

Repeating a request to a server after it's been closed and re-opened seems to try to re-use the prior connection, resulting in ECONNRESET.

Reproduction

Steps to reproduce the behavior:

Using node 20.6.1

// test.mjs
import { createServer } from 'node:http'
import fetch from 'node-fetch'

const server = createServer((req, res) => {
  res.writeHead(200, {
    // connection: 'close',
    'content-type': 'application/json',
  })
  res.end(
    JSON.stringify({
      hello: req.url === '/' ? 'world' : String(req.url).slice(1),
    })
  )
})

server.listen({ host: '127.0.0.1', port: 65200 })

{
  const res = await fetch('http://127.0.0.1:65200/')
  console.error(res.status, res.headers)
  console.log(await res.text())
}

server.close()
server.listen({ host: '127.0.0.1', port: 65200 })

{
  const res = await fetch('http://127.0.0.1:65200/')
  console.error(res.status, res.headers)
  console.log(await res.text())
}

server.close()

Expected behavior

Connection is closed, so it is not re-used.

Screenshots

Your Environment

software	version
node-fetch	3.3.2
node	20.6.1
npm	10.0.0
Operating System	Darwin moxy.lan 22.6.0 Darwin Kernel Version 22.6.0: Wed Jul 5 22:22:05 PDT 2023; root:xnu-8796.141.3~6/RELEASE_ARM64_T6000 arm64

Additional context

Worked fine in node 18. I haven't dug into what might have changed, but perhaps the connection close isn't being surfaced in the same way?

Node's built-in http.request() works in this scenario without any problem.

[zalox]

I also encountered this bug now when attempting to do consecutive calls to an API. One workaround I found was to include a await new Promise(resolve => resolve(setTimeout)); before the await fetch. So there is probably a synchronization bug somewhere.

The same code base works as expected with node 18, so I switched to that for now.

[Moumouls]

My team encountered the same issue. In our test suite for impacted test we provided a new agent without keep alive.

Btw, think it's not really an issue of node-fetch, but a new behavior of Node 19 and 20. Since Node 19 the global default agent use keep alive.

The workaround is providing a custom agent for impacted test calls.

import { Agent } from 'http'
import fetch from 'node-fetch'

fetch(myUrl, { agent: new Agent() })

Technically this kind of issue should only occur in test purpose. Another solution is to change the port between each test.

[isaacs]

Technically this kind of issue should only occur in test purpose. Another solution is to change the port between each test.

If you make repeated requests to a server in production, and the connection closes between requests, it could happen in production as well. Servers do restart sometimes in real life.