Barry Oglesby created GEODE-333:
-----------------------------------
Summary: Indexes sometimes are no longer used after a rebalance
Key: GEODE-333
URL: https://issues.apache.org/jira/browse/GEODE-333
Project: Geode
Issue Type: Bug
Components: querying
Reporter: Barry Oglesby
Assignee: Barry Oglesby
My test defines the following index:
{noformat}
<index name="customer_displayName" from-clause="/customer"
expression="displayName"/>
{noformat}
And the following query:
{noformat}
select * from /customer where displayName='some-display-name'
{noformat}
Query execution uses the index during execution before a rebalance takes place.
In some cases (see below) the index is no longer used after a rebalance.
When the issue occurs, the {{IndexManager getBestMatchIndex}} method returns no
indexes for the {{index_iter1.displayName}} expression even though there is one.
Here is some debugging that shows null being returned:
{noformat}
IndexManager getBestMatchIndex FUNCTIONAL index for index_iter1.displayName
found null from [Index [ Name=customer_displayName Type =FUNCTIONAL
IdxExp=displayName From=/customer Proj=*]imports : null]
{noformat}
This is due to this clause in the getBestMatchIndex method:
{noformat}
index = prIndex.getBucketIndex();
if (index == null) {
continue;
}
{noformat}
The {{getBucketIndex}} method asks the first value in the {{bucketIndexes}} Map
for its indexes. If the first value is empty, null is returned.
In this case, the first value is empty because the index at that bucket
({{\_B__customer_47}}) was removed during the rebalance.
Here is some debugging that shows it. The first value in the {{bucketIndexes}}
Map is for {{\_B__customer_47}}. This bucket is removed during the rebalance,
so its List of indexes initially contains an index, then it doesn't. This
causes {{getBucketIndex}} to return null and causes {{getBestMatchIndex}} to
short-circuit.
{noformat}
PartitionedIndex Bucket Indexes size=113; first
entry=BucketRegion[path='/__PR/_B__customer_47;serial=149;primary=false;indexUpdater=null]=[Index
[ Name=customer_displayName Type =FUNCTIONAL IdxExp=displayName From=/customer
Proj=*]]
PartitionedIndex removeFromBucketIndexes /__PR/_B__customer_47->Index [
Name=customer_displayName Type =FUNCTIONAL IdxExp=displayName From=/customer
Proj=*]
PartitionedIndex Bucket Indexes size=113; first
entry=BucketRegion[path='/__PR/_B__customer_47;serial=149;primary=false;indexUpdater=null]=[]
{noformat}
A potential fix is to remove any entries from the {{bucketIndexes}} Map that
are empty in the {{removeFromBucketIndexes}} method like:
{noformat}
public void removeFromBucketIndexes(Region r, Index index)
{
synchronized(this.bucketIndexes) {
List<Index> indexes = this.bucketIndexes.get(r);
if(indexes != null) {
indexes.remove(index);
if (indexes.isEmpty()) {
this.bucketIndexes.remove(r);
}
}
}
}
{noformat}
Here is some debugging that shows the first value in the {{bucketIndexes}} Map
is initially for {{\_B\_\_customer_68}}. Then {{\_B\_\_customer_68}} is removed
and {{\_B\_\_customer\_46}} takes over as first in the value in the
{{bucketIndexes}} Map.
{noformat}
PartitionedIndex Bucket Indexes size=113; first
entry=BucketRegion[path='/__PR/_B__customer_68;serial=209;primary=true;indexUpdater=null]=[Index
[ Name=customer_displayName Type =FUNCTIONAL IdxExp=displayName From=/customer
Proj=*]]
PartitionedIndex removeFromBucketIndexes removing entry for
/__PR/_B__customer_68
PartitionedIndex Bucket Indexes size=95; first
entry=BucketRegion[path='/__PR/_B__customer_46;serial=229;primary=false;indexUpdater=null]=[Index
[ Name=customer_displayName Type =FUNCTIONAL IdxExp=displayName From=/customer
Proj=*]]
{noformat}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
