[ https://issues.apache.org/jira/browse/HBASE-13472?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14496458#comment-14496458 ]
Andrew Purtell edited comment on HBASE-13472 at 4/15/15 4:16 PM: ----------------------------------------------------------------- Also, as part of this work, let's measure benefit of IN_MEMORY data placement with block preload. One easy comparison would be number of IOPS (with histogram of disk op latencies as appendix) with this mode versus normal operation, HFile access and WAL considered separately. Let's make sure we have coverage for these metrics. I seem to recall removing unused HFile metrics related to this, we don't need to reintroduce them, could just be a patch for the experiment. For quantifying client visible effects, consider YCSB2 (https://github.com/apurtell/YCSB2/tree/new_hbase_client, perhaps with improvements to this client as well, as it is still under development) was (Author: apurtell): Also, as part of this work, let's measure benefit of IN_MEMORY data placement with block preload. One easy comparison would be number of IOPS (with histogram of disk op latencies as appendix) with this mode versus normal operation, HFile access and WAL considered separately. For quantifying client visible effects, consider YCSB2 (https://github.com/apurtell/YCSB2/tree/new_hbase_client, perhaps with improvements to this client as well, as it is still under development) > Polish IN_MEMORY table behavior > ------------------------------- > > Key: HBASE-13472 > URL: https://issues.apache.org/jira/browse/HBASE-13472 > Project: HBase > Issue Type: Task > Reporter: Andrew Purtell > > For a long time we've been able to support a mode of operation that keeps as > much table data as possible in memory, so HBase can be used as an 'in-memory' > DB with fully durable WAL and write-behind persistence of table data. However: > - There are a set of relevant schema options (IN_MEMORY, CACHE_ON_WRITE, > PREFETCH_BLOCKS_ON_OPEN, block encoding), so set up isn't simple. We should > have a shortcut that sets all this up in one place. I'm thinking a utility > class with static helpers that configure a table descriptor with all of the > needed bits. (Other ideas?) > - We don't have a safety valve. An in-memory table can become too large, > where it falls out of blockcache and performs poorly without warning because > it's become too big. Consider table quota support with an option for region > size limits as % of total heap consumed by regions for a given table. Warn at > soft limit. Refuse writes if over hard limit. > Follow on work can investigate options hooking up to offheap work. That's not > in scope here. -- This message was sent by Atlassian JIRA (v6.3.4#6332)