[ https://issues.apache.org/jira/browse/SOLR-12011?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16383232#comment-16383232 ]
Shalin Shekhar Mangar commented on SOLR-12011: ---------------------------------------------- bq. Should we? Replica only sends PrepRecoveryOp to the leader after success updates its term. So I think a live-fetch on the leader's side will be enough. And I'm afraid that looping at that call can cause an endless loop. ( I'm not sure about this point ) So I traced the code and a live fetch on the leader works fine today but as a side-effect. We set the term equal to max for a recoverying replica (using ZkShardTerms.startRecovering() method) in ZkController.publish *before* we publish the replica state to the overseer queue. So if the leader (during prep recovery) sees replica state as recoverying then Zookeeper also guarantees that it will see the max term published before the recoverying state was published. I think we should make this behavior clear via a code comment. The changes in SolrCmdDistributor fix a different bug, no? Describe the problem here in this issue and how it is solved. Otherwise extract it to its own ticket. bq. Latest patch added changes for RestoreCoreOp and SplitOp where an empty core is added new data I don't understand this change. It seems to disallow restore or split if the slice has more than 1 replicas? > Consistence problem when in-sync replicas are DOWN > -------------------------------------------------- > > Key: SOLR-12011 > URL: https://issues.apache.org/jira/browse/SOLR-12011 > Project: Solr > Issue Type: Bug > Security Level: Public(Default Security Level. Issues are Public) > Components: SolrCloud > Reporter: Cao Manh Dat > Assignee: Cao Manh Dat > Priority: Major > Attachments: SOLR-12011.patch, SOLR-12011.patch, SOLR-12011.patch, > SOLR-12011.patch > > > Currently, we will meet consistency problem when in-sync replicas are DOWN. > For example: > 1. A collection with 1 shard with 1 leader and 2 replicas > 2. Nodes contain 2 replicas go down > 3. The leader receives an update A, success > 4. The node contains the leader goes down > 5. 2 replicas come back > 6. One of them become leader --> But they shouldn't become leader since they > missed the update A > A solution to this issue : > * The idea here is using term value of each replica (SOLR-11702) will be > enough to tell that a replica received the latest updates or not. Therefore > only replicas with the highest term can become the leader. > * There are a couple of things need to be done on this issue > ** When leader receives the first updates, its term should be changed from 0 > -> 1, so further replicas added to the same shard won't be able to become > leader (their term = 0) until they finish recovery > ** For DOWN replicas, the leader should also need to check (in DUP.finish()) > that those replicas have term less than leader before return results to users > ** Just by looking at term value of replica, it is not enough to tell us > that replica is in-sync with leader or not. Because that replica might not > finish the recovery process. We need to introduce another flag (stored on > shard term node on ZK) to tell us that replica finished recovery or not. It > will look like this. > *** {"code_node1" : 1, "core_node2" : 0} — (when core_node2 start recovery) > ---> > *** {"core_node1" : 1, "core_node2" : 1, "core_node2_recovering" : 1} — > (when core_node2 finish recovery) ---> > *** {"core_node1" : 1, "core_node2" : 1} -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org