Hbase takes a long time to recover after a restart due to disabled splitting.

[Guy Groulx]

DTM currently disables splits/moves during transactions.
I think there is an issue where it does not get reenabled.

What that means is that on hbase restarts, hbase goes into move/split/compaction mode where many many regions are offline until this is done. (We've seen up to an hour).

[Oliver Bucaojit (oliver-bucaojit)]

The split delay functionality has been checked-in on Jan 8. Similar to the 0.94 functionality, the split operation is delayed as long as there is a transaction in the pending or active list of the region.

So to test and confirm that the splitting is getting reenabled, it would be good to test with this new change and without the workaround of setting the “hbase.hregion.max.filesize” property.

The balancer work is still in progress and so invoking the balancer manually when the system has no transactions is still necessary.

[Adriana Fuentes (adriana-fuentes)]

Needs to be tested by QA or Guy to check if this situation reoccurs since Oliver committed the functionality that was missing and that might have been the cause for this bug.

[Guy Groulx (guy-groulx)]

The split delay seems to work much better on .98 than on .94. We are not seeing the long split times on hbase restarts anymore.