This avoids a regression for issues like #334 where high speed repeated connections eventually run the HTTP client out of sockets because all of theirs end up in time_wait.
Maybe the trade-off here is suboptimal, but if both choices will fail then we prefer fewer changes until the root cause is solved.