Every deploy restarting six consumers at once, triggering a rebalance that stops consumption for forty seconds.
the deploy: six workers restarted simultaneously.
the broker: a rebalance per departure and per arrival,
twelve in a few seconds, each pausing the
group.
measured: 41 seconds of no consumption per deploy,
four deploys a day.
the fix: restart them one at a time with a pause, and
accept a slower deploy. 41 seconds → 6.Rolling the restarts is the obvious answer and it is slower, which is why it was not done — a deploy step that takes two minutes instead of twenty seconds is a real cost paid on every release. Static group membership is the other option and it trades a different thing: a genuinely dead consumer is not noticed until its session times out.