Details
-
Improvement
-
Status: Closed
-
Major
-
Resolution: Invalid
-
0.95.1
-
None
-
None
-
None
Description
I noticed that the current split behavior is rather suboptimal with regard to compactions. On large regions, HFile size limit triggers a split. Split is followed by major compaction to get rid of the partial reference files. However, HFile size limit is surpassed after compaction most of the time.
So, first we rewrite a lot of data into a new file. Then we say "Oh look! A large file!", split the region and rewrite everything again.
Perhaps region split should be based on store size, or incoming compaction size - large enough compaction should be converted into splits.
Thoughts? I think basing off store size is a simple fix, and will code it up soon if there are no objections