[
https://issues.apache.org/jira/browse/HBASE-700?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12607490#action_12607490
]
Andrew Purtell commented on HBASE-700:
--------------------------------------
Hi Stack.
HBASE-62 is a generalization of the changes to HTableDescriptor under
consideration. So actually it makes sense to implement this and then have the
regionservers watch for certain known key-value pairs that specify table or
column store parameters. (Commenting on this issue specifically I don't think
JSON is warranted as an encoding for table and column metadata. Simple single
value strings for keys and values should be enough. However user metadata could
be formatted however the user desires.)
HBASE-34 could be specified at the table level. Regarding this issue I read
Bryan's concerns about exposing tuning parameters but suggest that people who
are tuning parameters exposed as proposed should know what they are doing, and,
if not, will soon learn better.
HBASE-43 seems pretty trivial, in a sense: In order to apply HTableDescriptor
updates, the client would need to disable the table, tell the master to update
the descriptor, and then reenable the table. So all of the pending edits would
be flushed for the disable. Then, when reenabled, the regionservers could note
the read only attribute and simply reject edits to the columns, and then both
they and whatever mapreduce job running over the mapfiles could coexist happily.
I think all of this could be rolled into one change set. Want to tie these all
to HBASE-42, or open a new JIRA?
> hbase.io.index.interval need be configuratable in column family
> ----------------------------------------------------------------
>
> Key: HBASE-700
> URL: https://issues.apache.org/jira/browse/HBASE-700
> Project: Hadoop HBase
> Issue Type: Improvement
> Components: regionserver
> Affects Versions: 0.1.2
> Reporter: LN
> Priority: Minor
>
> setting parameter hbase.io.index.interval to smaller can improve hbase
> reading performance significantly, esp. in large value size column families.
> however, small hbase.io.index.interval cause more memory usage, because all
> index will read into memory when loading a mapfile.
> in my test env, i set hbase.io.index.interval to 1, after inserting about 3M
> samll size records to a table(about 1.5G in hadoop file), the regionserver
> throws OOME. then i found total size of map file index is 350M. however,
> i can't adjust hbase.io.index.interval to a larger one, like 32, because
> other big cell size tables need it be 1.
> so, i think make hbase.io.index.interval a column family property should be
> very important for performance tuning.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
