[ https://issues.apache.org/jira/browse/SOLR-9925?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16063663#comment-16063663 ]
Brandon Chapman edited comment on SOLR-9925 at 6/26/17 8:04 PM: ---------------------------------------------------------------- We have seen this same issue on Solr 6.4.2. If you send deleteByQuery, {noformat}"_root_ :THE_ROOT"{noformat} with an add for the same document in the same request, you can trigger this easily. In our case we had to use a deleteByQuery due to deleteById not working correctly on sharded collections. Our only solution is to go back to a non-sharded collection so we can use deleteById. was (Author: bchapman): We have seen this same issue on Solr 6.4.2. If you send deleteByQuery {noformat}"_root_ :THE_ROOT"{noformat}, with an add for the same document in the same request, you can trigger this easily. In our case we had to use a deleteByQuery due to deleteById not working correctly on sharded collections. Our only solution is to go back to a non-sharded collection so we can use deleteById. > Child documents missing from replicas during parallel delete+add > ---------------------------------------------------------------- > > Key: SOLR-9925 > URL: https://issues.apache.org/jira/browse/SOLR-9925 > Project: Solr > Issue Type: Bug > Security Level: Public(Default Security Level. Issues are Public) > Affects Versions: 5.5.2, 6.3 > Environment: Java 1.8 (OpenJDK) on both CentOS 6.7 and Ubuntu 16.04.1 > Reporter: Dan Sirotzke > Attachments: generate.py, run.sh > > > When pushing documents to Solr in parallel, doing a delete-by-query and then > add for the same set of IDs within each thread results in some of the > replicas missing some of the child documents. All the parent documents are > successfully replicated. > This appears to trigger some sort of race condition, since: > * Documents are never missing from the leader. > * Documents _might_ be missing from the replicas. > * When they are missing, the number and which documents are different for > each replica and each run. > * It happens more easily with large documents; my test script needs a huge > number of documents to trigger it a small number of times, whereas it happens > ~5% of the time on our dataset. > * We're currently on Solr 5.5.2, but I've also managed to trigger it on 6.3.0 > * When not running anything in parallel, this doesn't occur. > Quick aside, since this is surely the first thing that will jump out: We > can't just do an update due to to the uniqueKey/\_root\_ issue behind > SOLR-5211. -- This message was sent by Atlassian JIRA (v6.4.14#64029) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org