Ensure that ibrowse streams are ended properly
I found a situation where we had live lock on a running application due
to an ibrowse request that hadn't been properly terminated. This
manifested as a cesation of updates to the _active_tasks information.
Debugging this lead me to see that the main couch_replicator pid was
stuck on a call to get_pending_changes. This call was stuck because the
ibrowse_http_client process being used was stuck waiting for a changes
request to complete.
This changes request as it turns out had been abandoned by the
couch_replicator_changes_reader. The changes reader was then stuck
trying to do a gen_server:call/2 back to the main couch_replicator
process with the report_seq_done message.
Given all this, it became apparent that the changes feed improperly
ending its ibrowse streams was the underlying culprit. Issuing a call to
ibrowse:stream_next/1 with the abandoned ibrowse stream id resulted in
the replication resuming.
This bug was introduced in this commit:
bfa020b43be20c54ab166c51f5c6e55c34d844c2
BugzId: 47306
This is a cherry-pick of:
https://github.com/cloudant/couch_replicator/commit/f9db37a9b293f5f078681e7539fd35a92eb3adec
1 file changed