We've noticed that our mirrors will semi regularly have problems due to
old stale works. For example using old ssl certs or having connection
problems to round robin backend services. In all cases restarting the
service (killing old workers) seems to fix things. Try to force this to
automatically happen by setting a reasonable connection limit per worker
before we recycle them.
Change-Id: Ic377f48d1a5a3eecbcb183327c9255134c4364ab
33c744d923