Okay, that getSplits part is specifically where my code was involved.
My use case was one of salted rowkeys. We are storing documents that
have a guid as the id and the creation date of the document is important
for scanning. When we tested having a rowkey format of
<creationtimestamp>+<guid>, the RegionServer hotspots became
problematic, so we decided to salt the rowkey by using the first digit
of the guid: <hexchar>+<creationtimestamp>+<guid>. This gives us nice
distribution of inserts throughout the cluster, but of course, it makes
scanning a contiguous date range much more complicated.
The code I have allows us to write a MR that takes a list of prefixes
(e.g. the hexchar) and a list of ranges (e.g. the desired timestamps)
and construct a master Scan object that contains any configuration such
as filters or cache settings, and a series of Scan objects that
constitute the Cartesian product of the ranges. Then, it passes those
in to a custom getSplits that ensures only the needed regions
participate in the Map.
If this sounds like it might be useful, I'll work on getting it cleaned
up and posted somewhere so you can review it and maybe glean it for
ideas. If you are already past that point then I apologize for not
checking into this sooner. :)
-Daniel
On 9/11/10 7:09 PM, John Sichi wrote:
Hi Daniel,
I'm almost done with this for HIVE-1226; the remaining step I need to finish is
to get the filter passed down during getSplits, since the HBase getSplits
implementation takes care of figuring out which regions contain the row in
question.
JVS
On Sep 11, 2010, at 7:00 PM, Daniel Einspanjer wrote:
I was trying to spend a little time this weekend catching up with the current
state of HBase integration for Hive. One thing that I haven't seen mentioned
is how exactly Hive scans an HBase table during a SELECT.
Does Hive have logic that allows it to intelligently scan only the
participating regions during a SELECT query that uses the rowkey? If not, I
recently wrote some code that allows a MapReduce job to effectively select the
regions based on a list of start/end rowkey ranges. If this might be useful to
the Hive integration, I could create a Jira and take a look at trying to set up
a patch.
Daniel Einspanjer
Metrics Architect
Mozilla Corporation
