this is affected some CSIT jobs and makes them spend a lot of wasted time.

This email thread sort of explains it:

https://lists.opendaylight.org/pipermail/integration-dev/2017-January/008967.html

to reproduce you can bring up a three node cluster. then kill the leader
node and one of the followers. the remaining node that stays up will be in
Candidate state. next send a simple rest request (e.g. /restconf/streams )
to this node. The request will be open and pending for 20s before it's
answered.

in csit where there tend to be a lot of these requests, when this
is seen it can severely add to the runtime of the test.

If possible, even if the response needs to be an error, the response should
come back right away. It seems there is some internal timer/timeout
getting in the way here, since it's always 20s.