When sockets get reused for multiple requests to endpoints that support chunked transfer encoding, the code in node-fetch that's designed to detect malformed chunked responses gets confused and throws false-positive 'premature close' errors.

Sockets get reused in such a way when you use an http.Agent or an https.Agent in ‘keep alive’ mode.

Reproduction

Steps to reproduce the behavior:

1. Add this logging to node-fetch (diff link)
2. Create an HTTPS Agent with keepAlive: true and maxSockets of 5.
3. Make 6 requests in a row for an endpoint that supports chunked transfer encoding

In that repro, 6 requests are made with a limit of 5 sockets in keep-alive mode. One of those sockets gets reused for the 6th request. The presence of the 6th request as a pending listener for that reused socket confuses the logic in node-fetch that is designed to detect corrupt chunked responses.

Here's a repro project that you can clone that does all of the above: https://github.com/steveluscher/repro-node-fetch-premature-close-with-agent

Expected behavior

No premature close error should be thrown, since the chunked responses are well-formed.

Your Environment

Additional context

This bug is the root of solana-foundation/solana-web3.js#1418.

cc/ @achingbrain.