z-york commented on pull request #2237:
URL: https://github.com/apache/hbase/pull/2237#issuecomment-676829830
I have been watching this for a while, and I still don't have good clarity
on the following case.
>
> > [update] I will follow your suggestion and throws a exception if we find
the meta table is not empty/partial
>
> Yes let's do this first to prevent damaging a cluster.
@Apache9 Can you explain how a partial meta table will damage a cluster? I
have been looking at the code for the InitMetaProcedure and the only thing that
I can see that is not idempotent is the directory creation for the NS (and the
delete of the meta table, but that is being used as the fix for making things
idempotent). If InitMetaProcedure is idempotent, there should be no need to
delete the directory.
If we restart a cluster and schedule an InitMetaProcedure:
1. Don't delete the meta table if it exists
2. It will then move to Assign (which makes sense, to trigger an IMP it
cannot already be assigned on a RS).
3. Then it will move to create NS (which if I understand the logic
correctly):
a. it can remove the create directory if they already exist
b. still overwrite the row in the meta table since it should be the same
(or check to see if the row exists first if we want to avoid an overwrite).
After that case, if there is any inconsistency in the meta table, that can
be handled via HBCK as usual.
A couple tenants we should always have:
1. Any potential data loss or corrupting action should be taken by HBCK
where the operator is making a conscious choice on how to fix a scenario.
2. As more information is being stored in hbase:meta, it becomes harder (or
at least more time consuming) to regenerate the information from the
filesystem. hbase:meta should only be deleted/completely regenerated in HBCK.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
