This issue is a follow-up to
If connections are contended towards a slow target, it may make sense
to set connectionTimeout and connectionRequestTimeout to values much lower than
the receiveTimeout. Expected client behavior is to receive an error if a connection
does not become available within connectionRequestTimeout. Current behavior however
is that the error is only received after up to receiveTimeout has passed, when a
current request to the target has finished and the connection is released or returned
to the pool.
This causes a possible build-up of pending requests in memory for the duration of receiveTimeout instead of connectionRequestTimeout.
See github PR: https://github.com/apache/cxf/pull/466
The reference solution in the PR works well, but it may not be the most elegant one due to my currently limited understanding of the overall design of the connection handling code.