Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Invalid
-
None
-
None
-
None
-
None
Description
The API gap for bulk HFile loading was discussed on the mailing list but it didn't make it into a JIRA. It also came up on HBASE-5498.
See http://search-hadoop.com/m/eEUHK1s4fo81/bulk+loading+and+RegionObservers
The salient detail:
A simple and straightforward course of action is to give the CP the option of rewriting the submitted store file(s) before the regionserver attempts to validate and move them into the store. This is similar to how CPs are hooked into compaction: CPs hook compaction by allowing one to wrap the scanner that is iterating over the store files. So the wrapper gets a chance to examine the KeyValues being processed and also has an opportunity to modify or drop them.
Similarly for incoming HFiles for bulk load, the CP could be given a scanner iterating over those files, if you had a RegionObserver installed. You would be given the option in effect to rewrite the incoming HFiles before they are handed over to the RegionServer for addition to the region.
I think this is a reasonable approach to interface design, because the fact you are given a scanner highlights the bulk nature of the input. However I think there should be two hooks here: one that allows for a simple yes/no answer as to whether the bulk load should proceed; and one that allows for a more expensive filtering or transformation or whatever via scanner-like interface. Bulk loads could be potentially very large so requiring a scan over them always is not a good idea.