[
https://issues.apache.org/jira/browse/HBASE-14623?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15206757#comment-15206757
]
Stephen Yuan Jiang commented on HBASE-14623:
--------------------------------------------
[~stack], you are right that we need to add tests (and run chaos monkey tests)
to see the benefit of this change before committing the change.
Since you have some concern on this JIRA, let me give you some background and
explanation (based on customer issue and internal chaos monkey tests):
Today, we have a dedicated WAL for META system table because we don't want to
other regions to wait for META to recovery and want to make sure that META
region recovers first. That change is about 3 years ago. Since then, we saw
usage of HBASE exploded and we have more system table that are also very
important. From time to time (in product env), we saw log split takes long
time or even fails due to resource limitation. If a system table region is in
the part of log in the split (or in the log that are waiting to be split), the
functionality of the cluster would be affected. Therefore, we thought of
dedicated WAL for system table approach to speed up system table recovery.
In terms of benefit to region assignment, we have already had special case for
META region assignment in the Assignment manager code (for user regions, if
assignment fails after some retry, it would give up and queue it and come back
to try again later; for meta region, the retry is forever until assignment
succeeds). From internally chaos money testing, we often see that a system
table region fails after maximum retries (the number is configurable, 10 is the
default max retry attempts), without system table regions online, we cannot do
a lot of thing and sometime test fails/times out. At that time, there is some
discussion on whether to give system table the infinite retry treatment as META
region. When I worked on designing new proc-v2 based assignment, I exchanged
this idea with [~mbertozzi], he had similar thoughts on this. We discussed
that the recovery for region assignment priority should be META, then system
table, and then user table regions. When [~tedyu] tried to fix a customer
issue in HBASE-14190 and he had this idea on dedicate system table WAL. After
discussion, we decided to have this JIRA to separate the work of HBASE-14190 to
make it cleaner. Having dedicated WAL(s) for system tables would help assign
system table region in higher priority than user table region, and hence speed
up system table region assignment.
Once we agree that there is a need to have dedicated WAL(s) for system tables,
then we can discuss the approach to implement this JIRA. I think [~tedyu]'s
approach is reasonable - to be safe and minimize the impact on this change, not
touch META WAL handling and use a separate logroller for system table.
However, I am open to more discussion on the approach. (By the way, my
personal opinion is that we should have sys:meta / sys:namespace / etc. instead
of hbase.meta /hbase:namespace / etc., but too late to make this change :-);
the same is that we have to keep .meta for META table WAL even it is a system
table)
> Implement dedicated WAL for system tables
> -----------------------------------------
>
> Key: HBASE-14623
> URL: https://issues.apache.org/jira/browse/HBASE-14623
> Project: HBase
> Issue Type: Sub-task
> Components: wal
> Reporter: Ted Yu
> Assignee: Ted Yu
> Labels: wal
> Fix For: 2.0.0
>
> Attachments: 14623-v1.txt, 14623-v2.txt, 14623-v2.txt, 14623-v2.txt,
> 14623-v2.txt, 14623-v3.txt, 14623-v4.txt
>
>
> As Stephen suggested in parent JIRA, dedicating separate WAL for system
> tables (other than hbase:meta) should be done in new JIRA.
> This task is to fulfill the system WAL separation.
> Below is summary of discussion:
> For system table to have its own WAL, we would recover system table faster
> (fast log split, fast log replay). It would probably benefit
> AssignmentManager on system table region assignment. At this time, the new
> AssignmentManager is not planned to change WAL. So the existence of this JIRA
> is good for overall system, not specific to AssignmentManager.
> There are 3 strategies for implementing system table WAL:
> 1. one WAL for all non-meta system tables
> 2. one WAL for each non-meta system table
> 3. one WAL for each region of non-meta system table
> Currently most system tables are one region table (only ACL table may become
> big). Choices 2 and 3 basically are the same.
> From implementation point of view, choices 2 and 3 are cleaner than choice 1
> (as we have already had 1 WAL for META table and we can reuse the logic).
> With choice 2 or 3, assignment manager performance should not be impacted and
> it would be easier for assignment manager to assign system table region (eg.
> without waiting for user table log split to complete for assigning system
> table region).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
