Fix changes worker timeout cleanup
Previously if we timed out waiting for the next message the changes
reader would end up just exiting with an error. Unfortunately the
ibrowse worker doesn't bother noticing that its streaming target has
died and will wait in perpetuity. If the main replication process
happens to be waiting on this HTTP worker it'll block indefinitely and
never make progress in the replication.
This change just ensures that the ibrowse worker is killed which will
cause the main replication pid to restart.
This particular bug has been observed on the Oculus clusters at a fairly
low rate so the cost of restarting a replication shouldn't be an issue.
BugzId: 47971
1 file changed
