[
https://issues.apache.org/jira/browse/HBASE-18105?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16181268#comment-16181268
]
stack commented on HBASE-18105:
-------------------------------
Rollback in ProcedureV2 is a fuzzy rather than hard concept. It doesn't always
make sense in the Procedure context. For example, in assign procedure, there is
no rollback of an assign; i.e. re-assign back to the old location. Instead, we
press-on and assign 'anywhere' if we are unable to assign to the original
target. It is better that the procedure 'succeed' even if the end-point is not
exactly what was asked for rather than rollback and fail. That said, some
procedures make use of rollback. Others use rollback to roll back some steps --
e.g. the first few in a Procedure -- but for the later steps, they'll consider
themselves committed and will try to achieve the endpoint.
So rollback is a bit messy in Pv2.
For split/merge, purge PONR. We don't have this in AMv2. But if during split,
there is an error, can we figure an alternate route to the end-point ; i.e. two
daughters assigned when split....
If not possible, just fail the procedure. Lets get some experience with the
failure types and fix them as they come up in subsequent issues.
Hope this helps [~easyliangjob]
> [AMv2] Split/Merge need cleanup; currently they diverge and do not fully
> embrace AMv2 world
> -------------------------------------------------------------------------------------------
>
> Key: HBASE-18105
> URL: https://issues.apache.org/jira/browse/HBASE-18105
> Project: HBase
> Issue Type: Sub-task
> Components: Region Assignment
> Affects Versions: 2.0.0
> Reporter: stack
> Assignee: Yi Liang
> Fix For: 2.0.0
>
>
> Region Split and Merge work on the new AMv2 but they work differently. This
> issue is about bringing them back together and fully embracing the AMv2
> program.
> They both have issues mostly the fact that they carry around baggage no
> longer necessary in the new world of assignment.
> Here are some of the items:
> Split and Merge metadata modifications are done by the Master now but we have
> vestige of Split/Merge on RS still; e.g. when we SPLIT, we ask the Master
> which asks the RS, which turns around, and asks the Master to run the
> operation. Fun. MERGE is all done Master-side.
>
> Clean this up. Remove asking RS to run SPLIT and remove RegionMergeRequest,
> etc. on RS-side. Also remove PONR. We don’t Points-Of-No-Return now we are up
> on Pv2. Remove all calls in Interfaces; they are unused. Make RS still able
> to detect when split, but have it be a client of Master like anyone else.
> Split is Async but does not return procId
> Split is async. Doesn’t return the procId though. Merge does. Fix. Only hard
> part here I think is the Admin API does not allow procid return.
> Flags
> Currently OFFLINE is determined by looking either at the master instance of
> HTD (isOffline) and/or at the RegionState#state. Ditto for SPLIT. For MERGE,
> we rely on RegionState#state. Related is a note above on how split works --
> there is a split flag in HTD when there should not be.
>
> TODO is move to rely on RegionState#state exclusively in Master.
> From Split/Merge Procedures need finishing in
> https://docs.google.com/document/d/1eVKa7FHdeoJ1-9o8yZcOTAQbv0u0bblBlCCzVSIn69g/edit#heading=h.4b60dc1h4m1f
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
