[ https://issues.apache.org/jira/browse/HBASE-20894?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16545671#comment-16545671 ]
Mike Drob commented on HBASE-20894: ----------------------------------- Want to check with folks to see if there is agreement that this is a reasonable approach to take. A few design questions: * Should I be building a separate IO Engine implementation to do this instead of trying to handle it inline? * Is it ok to put the messy PB logic in the persist/retrieve methods, or should that go to various classes with toPB/fromPB methods in those? I see some examples of both in our code. * What is the difference for PB between writeTo and writeDelimitedTo (and the corresponding read methods) * Are my protobuf message definitions fine or do they need to be organized differently? I haven't spent too much thought on these. Regarding my previous question, I think recording cache size and IO Engine class seems fine, but tracking the backing map class is probably not necessary. Also, maybe we can simplify the logic and not worry about the old serialization types - it's "just" a cache hint anyway so nothing critical lost if it doesn't come up with the RS. > Move BucketCache from java serialization to protobuf > ---------------------------------------------------- > > Key: HBASE-20894 > URL: https://issues.apache.org/jira/browse/HBASE-20894 > Project: HBase > Issue Type: Task > Components: BucketCache > Affects Versions: 2.0.0 > Reporter: Mike Drob > Priority: Major > Fix For: 3.0.0 > > Attachments: HBASE-20894.WIP-2.patch, HBASE-20894.WIP.patch > > > We should use a better serialization format instead of Java Serialization for > the BucketCache entry persistence. > Suggested by Chris McCown, who does not appear to have a JIRA account. -- This message was sent by Atlassian JIRA (v7.6.3#76005)