[ https://issues.apache.org/jira/browse/COUCHDB-1258?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13090466#comment-13090466 ]
Paul Joseph Davis commented on COUCHDB-1258: -------------------------------------------- Sounds reasonable. A map function that emits no keys coupled with a lot of deleted documents means that the partial_update frequency will be quite large. couch_work_queue's behave quite poorly when the dequeue side is faster than the queue side and end up basically not bulking updates together. A quick test would be to guard the partial_update message to only fire at most every 1000 writes. A quick hack to store calls in the process dict in do_writes should give a quick idea if it's the partial updates. A should_flush for every partial update seems a bit extreme. Although I can't think of a better solution that doesn't just make an equivalent to should_flush that's more specific to this case. > eheap_alloc OOM errors when attempting to build selected views > -------------------------------------------------------------- > > Key: COUCHDB-1258 > URL: https://issues.apache.org/jira/browse/COUCHDB-1258 > Project: CouchDB > Issue Type: Bug > Affects Versions: 1.1 > Environment: CentOS 5.6, 1GB RAM (approx 800MB free) > CouchDb 1.1.0, Erlang R14B-03, js lib 1.7.0-8 EPEL RPM build > Couch database exhibiting this behaviour: > {"db_name":"activity_new","doc_count":593274,"doc_del_count":4743352,"update_seq":10559287,"purge_seq":0,"compact_running":false,"disk_size":3366396013,"instance_start_time":"1314097423985726","disk_format_version":5,"committed_update_seq":10559287} > Reporter: James Cohen > Priority: Critical > > We spotted OOM errors crashing our CouchDb instance when attempting to > rebuild selected views. CouchDb was dying with the following messages (worth > noting that the type reported varies between heap/old_heap > eheap_alloc: Cannot allocate 478288480 bytes of memory (of type "heap"). > eheap_alloc: Cannot allocate 597860600 bytes of memory (of type "heap"). > eheap_alloc: Cannot allocate 747325720 bytes of memory (of type "old_heap"). > eheap_alloc: Cannot allocate 597860600 bytes of memory (of type "old_heap"). > By modifying the view I was able to find a view that could consistently crash > the server and another that ran fine. They are as follows: > Runs out of memory v.quickly > { > "_id": "_design/cleanup", > "_rev": "5-e004fbab278355e9d08763877e5a8295", > "views": { > "byDate": { > "map": "function(doc) { if (! doc.action) emit([doc.date], doc); }" > } > } > } > Runs fine with minimal memory usage (returns 88128 docs in the view) > { > "_id": "_design/cleanup", > "_rev": "6-3823be6b72ca2441e235addfece6900c", > "views": { > "byDate": { > "map": "function(doc) { if (doc.action) emit([doc.date], doc); }" > } > } > } > The only difference between the two is the negation of the if conditional. > memory usage was monitored with top on the machine while the view was being > built. Under correct behaviour I could see beam.smp using just 3 or 4% of the > server's memory. With the view that causes problems that memory usage > increased until the RAM/swap on the server was exhausted (as you can see from > the error messages around 500/700MB) -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira