[
https://issues.apache.org/jira/browse/HBASE-26726?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andrew Kyle Purtell updated HBASE-26726:
----------------------------------------
Description:
We have encountered two issues with region warmup before the assignment manager
gracefully moves a region from one regionserver to another.
The first instance is HBASE-26722. Part of the failure chain is temporary
double assignment like conditions where both the source and destination
regionservers think they have exclusive rights to storefile management (as they
should) but warmup opens a region before it is closed and then both
regionservers take compaction related actions. While this can be remediated
with more care to this case, it is unclear if warmup affords significant
advantage. The motivation of the original commit in 2015 introducing this
feature was avoidance of blockcache misses once region ownership transfer was
advertised to clients. Depending on use case and additional default-false
schema options (like preload) this could be valuable. Or not.
The second instance is in place upgrade from HBase 1 to HBase 2. In a scenario
where regionservers have been replaced by HBase 2 versions, but the master is
still HBase 1, the region warmup RPC request fails. This is not particularly
harmful but indicates it will not be useful during the transition period.
We think it would be good to allow for warmup before move to be optionally
disabled by a site configuration setting. In particular there have been many
unrelated changes committed since 2015 and expectations of invariants in the
contribution of the warmup-on-move feature have been invalidated. HBASE-26722
may be the only case, or it might not. This is a fairly trivial change.
was:
We have encountered two issues with region warmup before the assignment manager
gracefully moves a region from one regionserver to another.
The first instance is HBASE-26722. Part of the failure chain is temporary
double assignment like conditions where both the source and destination
regionservers think they have exclusive rights to storefile management (as they
should) but warmup opens a region before it is closed. While this can be
remediated with more care to this case, it is unclear if warmup affords
significant advantage. The motivation of the original commit in 2015
introducing this feature was avoidance of blockcache misses once region
ownership transfer was advertised to clients. Depending on use case and
additional default-false schema options (like preload) this could be valuable.
Or not.
The second instance is in place upgrade from HBase 1 to HBase 2. In a scenario
where regionservers have been replaced by HBase 2 versions, but the master is
still HBase 1, the region warmup RPC request fails. This is not particularly
harmful but indicates it will not be useful during the transition period.
We think it would be good to allow for warmup before move to be optionally
disabled by a site configuration setting. In particular there have been many
unrelated changes committed since 2015 and expectations of invariants in the
contribution of the warmup-on-move feature have been invalidated. HBASE-26722
may be the only case, or it might not. This is a fairly trivial change.
> Allow disable of region warmup before graceful move
> ---------------------------------------------------
>
> Key: HBASE-26726
> URL: https://issues.apache.org/jira/browse/HBASE-26726
> Project: HBase
> Issue Type: Improvement
> Components: master, Region Assignment
> Reporter: Andrew Kyle Purtell
> Assignee: Andrew Kyle Purtell
> Priority: Minor
> Fix For: 2.5.0, 2.6.0, 3.0.0-alpha-3
>
>
> We have encountered two issues with region warmup before the assignment
> manager gracefully moves a region from one regionserver to another.
> The first instance is HBASE-26722. Part of the failure chain is temporary
> double assignment like conditions where both the source and destination
> regionservers think they have exclusive rights to storefile management (as
> they should) but warmup opens a region before it is closed and then both
> regionservers take compaction related actions. While this can be remediated
> with more care to this case, it is unclear if warmup affords significant
> advantage. The motivation of the original commit in 2015 introducing this
> feature was avoidance of blockcache misses once region ownership transfer was
> advertised to clients. Depending on use case and additional default-false
> schema options (like preload) this could be valuable. Or not.
> The second instance is in place upgrade from HBase 1 to HBase 2. In a
> scenario where regionservers have been replaced by HBase 2 versions, but the
> master is still HBase 1, the region warmup RPC request fails. This is not
> particularly harmful but indicates it will not be useful during the
> transition period.
> We think it would be good to allow for warmup before move to be optionally
> disabled by a site configuration setting. In particular there have been many
> unrelated changes committed since 2015 and expectations of invariants in the
> contribution of the warmup-on-move feature have been invalidated. HBASE-26722
> may be the only case, or it might not. This is a fairly trivial change.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
