Requests to local CouchDB have 40ms latency

Question

Requests to a CouchDB instance on the same machine have about 40ms latency when HTTP keepalive is used by the client, about 0-1ms latency when the client does not use keepalive. What causes this latency?

Answer 1

This is caused by Nagle's algorithm. CouchDB sends the HTTP headers and the response body in separate calls, causing the kernel to not deliver the response body for 40ms.

As a workaround, you can use this LD_PRELOAD shim to completely turn off Nagle's algorithm for the database process.