[ https://issues.apache.org/jira/browse/KUDU-2711?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Grant Henke resolved KUDU-2711. ------------------------------- Resolution: Fixed Fix Version/s: 1.10.0 Resolved via: * [https://github.com/apache/kudu/commit/c74e67830] * [https://github.com/apache/kudu/commit/717349b4e] * [https://github.com/apache/kudu/commit/586e957f7] * [https://github.com/apache/kudu/commit/58f189dac] > Improve performance of GetTableLocations for large tables > --------------------------------------------------------- > > Key: KUDU-2711 > URL: https://issues.apache.org/jira/browse/KUDU-2711 > Project: Kudu > Issue Type: Improvement > Components: master, perf > Affects Versions: 1.9.0 > Reporter: Todd Lipcon > Assignee: Will Berkeley > Priority: Major > Fix For: 1.10.0 > > > Currently the response for TableLocations returns a separate copy of the > TSInfoPB for every replica. These PBs are relatively large and include string > UUIDs, hostnames, etc. For large tables, a response may have up to 3000 > replicas. In most clusters, the table itself will be on far fewer than 3000 > unique hosts, so it makes sense to return the replicas as indexes into a > response-level list rather than separately duplicate the TS info in each of > the replica PBs. > Additionally, we can consider a reader-writer lock on TSDescriptor so that > these requests can proceed more in parallel. -- This message was sent by Atlassian JIRA (v7.6.3#76005)