HTTP write timeout closes HTTP reads without logging

[rdivia]

Influxdb2 2.7.5 on Linux RHEL 8.9. We had a influxd client (Grafana) with complex queries (GET or POST), failing if these took more than 16-18 seconds to complete (after ~16 seconds the client would get EOF on the HTTP channel to influxd). Making the queries simpler (14- seconds) all worked OK. Problem confirmed with local influx queries (via shell “influx” command), same symptoms, same timings. Turned on all log options, up to debug level: failing queries reported completed OK after 16+ seconds, no trace of problems with the queries - but the client would still get the EOF.

To make a long story short (several days of investigations), we noticed that the “http-write-timeout” configuration parameter was set to 15 seconds, very close to the timeout with our failing queries. After increasing it, all worked as expected, problem solved.

Questions:

Why does a write-related parameter affect read operations? Note: docs strongly recommend to set this parameter according to local working conditions (which we did).

Why this timeout firing and causing the query channel to be closed (but not the query itself, that one - at least according to the logs - completed OK) was not mentioned at all in the logs of influxd?

[davidby-influx]

The write timeout is passed directly to the Go net/http.Server. It is not the timeout for a write to InfluxDB, but the timeout for writing the response to an http request.

Thus, this timeout, when reached, will affect the http connection, not InfluxDB's query.

The documentation is in error.