[ https://issues.apache.org/jira/browse/CASSANDRA-5902?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15058454#comment-15058454 ]
Aleksey Yeschenko commented on CASSANDRA-5902: ---------------------------------------------- This is fine, conceptually. That said, if our working assumption is that this situation should be rare (and I agree with the assumption here), then we might go a slightly simpler route: - for hints files for which the node is no longer part of the ring, instead of regular delivery, write the new hints locally instead, just to their new destinations - and have them replay later via the regular hints replay path - on the receiving end, if the hint no longer belongs, write a hint on that node for all the replicas instead, and have them be delivered later via the regular replay path We do pay a bit more price in file IO - in redundant writes, but writing a hint for multiple replicas is relatively cheap (serialised just once, using the same space in the buffer), and this should be rare. As a bonus, {{Hint::apply()}} remains local only, without calls to {{StorageProxy}}, plus we could avoid further extending {{AWRH}}. Please let me know if I'm just being silly and none of this makes any sense - I haven't head enough tea today. > Dealing with hints after a topology change > ------------------------------------------ > > Key: CASSANDRA-5902 > URL: https://issues.apache.org/jira/browse/CASSANDRA-5902 > Project: Cassandra > Issue Type: Sub-task > Components: Coordination > Reporter: Jonathan Ellis > Assignee: Branimir Lambov > Priority: Minor > Fix For: 3.0.x, 3.x > > > Hints are stored and delivered by destination node id. This allows them to > survive IP changes in the target, while making "scan all the hints for a > given destination" an efficient operation. However, we do not detect and > handle new node assuming responsibility for the hinted row via bootstrap > before it can be delivered. > I think we have to take a performance hit in this case -- we need to deliver > such a hint to *all* replicas, since we don't know which is the "new" one. > This happens infrequently enough, however -- requiring first the target node > to be down to create the hint, then the hint owner to be down long enough for > the target to both recover and stream to a new node -- that this should be > okay. -- This message was sent by Atlassian JIRA (v6.3.4#6332)