[ https://issues.apache.org/jira/browse/HBASE-12976?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Lars Hofhansl updated HBASE-12976: ---------------------------------- Attachment: 12976.txt Trivial but important change. Client can set caching as desired, the size of chunks between client and server is limited to 2MB (unless a single row is larger than 2MB). In my tests that made a *huge* difference in region server health and GC behavior (avoided full GCs due to many handlers allocating too many big chunks for responses, while not impacting performance). > Default hbase.client.scanner.max.result.size > -------------------------------------------- > > Key: HBASE-12976 > URL: https://issues.apache.org/jira/browse/HBASE-12976 > Project: HBase > Issue Type: Bug > Reporter: Lars Hofhansl > Fix For: 2.0.0, 1.0.1, 0.94.27, 0.98.11 > > Attachments: 12976.txt > > > Setting scanner caching is somewhat of a black art. It's hard to estimate > ahead of time how large the result set will be. > I propose we hbase.client.scanner.max.result.size to 2mb. That is good > compromise between performance and buffer usage on typical networks (avoiding > OOMs when the caching was chosen too high). > To an HTable client this is completely transparent. -- This message was sent by Atlassian JIRA (v6.3.4#6332)