Filtered replication failure on changes_timeout
I have a filtered replication routine that always fails with the error message showing "changes_timeout". Without filters it works fine, though. I may not have understood this correctly, but I think this indicates that the filter is running too long causing the destination server to not send a heartbeat signal, thus triggering the timeout? My database is rather large with upwards of 10k documents.
I also notice that on replication, the changes heartbeat is set to 10 secs. Is there any way to change this value through either configuration or query string? Or is the heartbeat value irrelevant to the problem I'm having?
Any suggestions are welcome.
Hi mkultra,
I'm now fairly sure this has been fixed in later versions. If you have access to a sufficiently large dataset, you might try something like this:
1. Install couchdb (not couchbase) from the git repository master. It should be version 1.3 or 1.2.
2. Replicate your dataset to this couchdb.
3. Try filtered replication (preferably using _replicator db) without doing any of the tricks you mentioned before. I'm fairly sure it'll work.
I just wonder when are they going to release the 1.2 version. It is much nicer than 1.1.x
bump.
I'd like to hear what the devs have to say about this too. Here are my experiences with this...
We have production databases with 100's of thousands to millions of docs. Filtered replication has been an issue for us as well. If you tail the changes feed with a filter, *sometimes*, on timeout the result will be "{"results":[" and thats it. The json object does not get closed out.
It seems that if this result is fed into the replicator, the replicator will break. I'm not sure about the heartbeat theory like mentioned above. Maybe both of these are part of the problem.
So, we do our own filtering server side, catch this botched result when it occurs, and feed the results (when we have them) into the _replicate db to let couch pick up the rest.
We use couch v1.1.1. Not sure if it's been fixes in newer versions... thats why this thread is interesting. I'm hoping to find some answers.