[
https://issues.apache.org/jira/browse/HBASE-20111?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16415886#comment-16415886
]
Josh Elser commented on HBASE-20111:
------------------------------------
{quote}{{checkSplit}} already has check w.r.t. system tables:
{quote}
I think I'd disagree with you, although I see what lead you to this opinion. I
read this code as the intent of {{splittable}} is more about overt details of
this region which are not tied to the lifecycle of this region: e.g. split
policy, not a 'special' system table. {{checkSplit}} more encapsulates logic
about whether we can (right now) split the region.
{quote}It seems {{shouldSplit}} should take into account the value of
{{bestSplitRow}} - if {{bestSplitRow}} is null, {{shouldSplit}} should be false.
{quote}
This is handled in the Master which I think is the right place for it to
happen. You're right that if this RPC would return that the region is
splittable, the user did not provide a row to split upon, and the region could
not provide a row to split upon, that should be an assertion failure.
{code:java}
if (node != null) {
try {
if (bestSplitRow == null || bestSplitRow.length == 0) {
LOG.info("splitKey isn't explicitly specified, " + " will try to find
a best split key from RS");
}
// Always set bestSplitRow request as true here,
// need to call Region#checkSplit to check it splittable or not
GetRegionInfoResponse response =
Util.getRegionInfoResponse(env, node.getRegionLocation(),
node.getRegionInfo(), true);
if(bestSplitRow == null || bestSplitRow.length == 0) {
bestSplitRow = response.hasBestSplitRow() ?
response.getBestSplitRow().toByteArray() : null;
}
splittable = response.hasSplittable() && response.getSplittable();
if (LOG.isDebugEnabled()) {
LOG.debug("Splittable=" + splittable + " " + node.toShortString());
}
} catch (IOException e) {
splittableCheckIOE = e;
}
}
if (!splittable) {
IOException e = new
DoNotRetryIOException(regionToSplit.getShortNameToLog() + " NOT splittable");
if (splittableCheckIOE != null) e.initCause(splittableCheckIOE);
throw e;
}
if(bestSplitRow == null || bestSplitRow.length == 0) {
throw new DoNotRetryIOException("Region not splittable because
bestSplitPoint = null, "
+ "maybe table is too small for auto split. For force split, try
specifying split row");
}{code}
IMO, what we have already covers this pretty well since the SplitTableProcedure
could still proceed with a user-provided split key if the Region could not
compute one for some reason.
That said, the {{isClosing()}} check done in {{HRegion.checkSplit()}} does
sound like it should be lifted out of this method and set explicitly on
{{isSplittable}}..
> Able to split region explicitly even on shouldSplit return false from split
> policy
> ----------------------------------------------------------------------------------
>
> Key: HBASE-20111
> URL: https://issues.apache.org/jira/browse/HBASE-20111
> Project: HBase
> Issue Type: Bug
> Reporter: Rajeshbabu Chintaguntla
> Assignee: Rajeshbabu Chintaguntla
> Priority: Critical
> Fix For: 2.0.0
>
> Attachments: HBASE-20111.001.branch-2.0.patch, HBASE-20111.patch,
> HBASE-20111_test.patch, HBASE-20111_v2.patch, HBASE-20111_v3.patch
>
>
> Currently able to split the region explicitly even when the split policy
> returns from shouldSplit.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
