[jira] [Commented] (HBASE-18294) Flush is based on data size instead of heap size

["Eshcar Hillel (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-18294?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16071594#comment-16071594 ] 

Eshcar Hillel commented on HBASE-18294:
---------------------------------------

Considering heap size for triggering global heap pressure reduction is fine, however, we expect the system to get to this point rarely.
As you mention in your document disregarding heap size can cause blocking, excessive GC and even OOME.
Following HBASE-16747 there are 3 cases where considering data size and not heap size
(1) when choosing biggest region to flush following global heap pressure
(2) when choosing large stores to flush within the region
(3) when deciding to trigger  region level flush (exceeding the 128MB threshold).

Lets consider the on-heap scenario first:
For (1)+(2) it is already decided that there is a flush and now we are looking to choose biggest memory consumers (biggest region/largest store). Therefore we should compare heap size. Otherwise, we may choose a smaller region/store simply because its data is bigger but it may actually take less heap space.  
For (3) the configurable threshold gives the admin a way to compute how many regions the RS can handle with a given memory. That was documented in blogs etc. Now with the change we actually have no way to predict what is the maximum memory size of a region. This may come as a surprise to some admins.

bq. For the off heap based memstores also the data size alone based region flush is better.
bq. Offheap memstores need this change so that we are able to decide the region flush based on data size alone as the whole data is offheap.
Then lets make a different policy for off-heap cases. Enforcing the same policy for on/off-heap doesn't seem to be optimal for both.

I suggest to make different policies for on-heap data (to consider heap size in all the cases above), and off-heap that considers heap/data size any way you think is best for off-heap. 

> Flush is based on data size instead of heap size
> ------------------------------------------------
>
>                 Key: HBASE-18294
>                 URL: https://issues.apache.org/jira/browse/HBASE-18294
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Eshcar Hillel
>            Assignee: Eshcar Hillel
>
> A region is flushed if its memory component exceed a threshold (default size is 128MB).
> A flush policy decides whether to flush a store by comparing the size of the store to another threshold (that can be configured with hbase.hregion.percolumnfamilyflush.size.lower.bound).
> Currently the implementation (in both cases) compares the data size (key-value only) to the threshold where it should compare the heap size (which includes index size, and metadata).



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)