[
https://issues.apache.org/jira/browse/HBASE-10018?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13831365#comment-13831365
]
Nicolas Liochon commented on HBASE-10018:
-----------------------------------------
bq. Does the prefetch have to be done inline
Not really. It's just a little bit more complicated if we start a thread.
bq. What to do though if 1M regions
It does not care: it reads the next 10 regions, whatever the table size.
Ihmo, when you have a lot of region it does not really work, when you have just
a few it's not really necessary.
Excepted in the scenario mentioned by Elliott above.
bq. Could just flip the configuration to turn off prefetch in 0.98?
I would prefer to be more radical: just remove it. Or at least, make it
optional, deactivated by default, on .96 and .98.
Thinking more about it, that's my prefered scenario. If nobody disagrees, I
will do that: "make it optional, deactivated by default, on .96 and .98."
> Change the location prefetch
> ----------------------------
>
> Key: HBASE-10018
> URL: https://issues.apache.org/jira/browse/HBASE-10018
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.98.0, 0.96.0
> Reporter: Nicolas Liochon
> Assignee: Nicolas Liochon
> Priority: Minor
> Fix For: 0.98.0, 0.96.1
>
>
> Issues with prefetching are:
> - we do two calls to meta: one for the exact row, one for the prefetch
> - it's done in a lock
> - we take the next 10 regions. Why 10, why the 10 next?
> - is it useful if the table has 100K regions?
> Options are:
> - just remove it
> - replace it with a reverse scan: this would save a call
>
--
This message was sent by Atlassian JIRA
(v6.1#6144)
