turbo-http: ECONNRESET when handling requests asynchronously
I’m seeing an issue where if
- the request handler is asynchronous (in that it doesn’t respond to the request in the same event loop tick)
- the server receives new requests while a previous one is still being handled
I will randomly get an ECONNRESET error on the client without any finish, end, close or error events on the server req.socket object.
About this issue
- Original URL
- State: closed
- Created 6 years ago
- Comments: 15 (15 by maintainers)
Commits related to this issue
- Set default backlog size to 511 Match Node's default backlog of 511 sockets (see https://nodejs.org/dist/latest-v9.x/docs/api/net.html#net_server_listen) This helps address https://github.com/mafint... — committed to pietermees/turbo-net by pietermees 6 years ago
- Set default backlog size to 511 (#9) Match Node's default backlog of 511 sockets (see https://nodejs.org/dist/latest-v9.x/docs/api/net.html#net_server_listen) This helps address https://github.com/m... — committed to mafintosh/turbo-net by pietermees 6 years ago
With some further digging I think I found the origin of the issue.
TLDR: the server runs out of backlog for new connections
It’s indeed possible to reproduce the issue with Node’s native
httpmodule, but only with much larger request counts.The reason is the
backlogparameter toserver.listen. Node uses511by default, although this is further limited by the OS’sSOMAXCONNsetting. On my Mac that defaults to128. If I setSOMAXCONN=1024I can get rid of the error.The reason why the problem happens much sooner with
turbo-httpis thatturbo-netsets the backlog size to a fixed5. This means that only 5 connections can be queued up to be accepted by the server before new connections get rejected. By sending a lot of simultaneous requests, it’s easy to cross that limit.Here’s the line in
turbo-netcausing this (note the TODO about researching backlog): https://github.com/mafintosh/turbo-net/blob/master/src/turbo_net.c#L190