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. Rebased-From: 1a25a7edf82706c3152e2d978d320ec465a34de1 7d2cb485116636595250fce4ea4eab16a877479b Github-Pull: #5674