[
https://issues.apache.org/jira/browse/PHOENIX-3451?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15665868#comment-15665868
]
James Taylor commented on PHOENIX-3451:
---------------------------------------
[~comnetwork] - good, helpful analysis, but your conclusion isn't quite right.
We want to use the pk position according to the GROUP BY because the rows
returned from the server which are *aggregated* rows and the sort will be done
on the client. The problem appears to be that our hasEqualityConstraints isn't
taking this into account. It should be determining if there's an equality
constraint for the GROUP BY expressions in those positions (instead of treating
those as positions in the original schema).
Probably the easiest fix would be to hold on to the OrderPreservingTracker.Info
in a list for the GroupByCompiler. Then in the OrderByCompiler, we could look
at the List<OrderPreservingTracker.Info> from the GroupBy and essentially index
it by position, but then translate it based on
OrderPreservingTracker.Info.pkPosition. That would be the correct index to use
when calling hasEqualityConstraints. If there's no list (i.e. it's not an
aggregation), we'd just use the pkPosition to directly index the ScanRanges as
we're doing now.
> Secondary index and query using distinct: LIMIT doesn't return the first rows
> -----------------------------------------------------------------------------
>
> Key: PHOENIX-3451
> URL: https://issues.apache.org/jira/browse/PHOENIX-3451
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.8.0
> Reporter: Joel Palmert
> Assignee: chenglei
> Attachments: PHOENIX-3451.diff
>
>
> This may be related to PHOENIX-3452 but the behavior is different so filing
> it separately.
> Steps to repro:
> CREATE TABLE IF NOT EXISTS TEST.TEST (
> ORGANIZATION_ID CHAR(15) NOT NULL,
> CONTAINER_ID CHAR(15) NOT NULL,
> ENTITY_ID CHAR(15) NOT NULL,
> SCORE DOUBLE,
> CONSTRAINT TEST_PK PRIMARY KEY (
> ORGANIZATION_ID,
> CONTAINER_ID,
> ENTITY_ID
> )
> ) VERSIONS=1, MULTI_TENANT=TRUE, REPLICATION_SCOPE=1, TTL=31536000;
> CREATE INDEX IF NOT EXISTS TEST_SCORE ON TEST.TEST (CONTAINER_ID, SCORE DESC,
> ENTITY_ID DESC);
> UPSERT INTO test.test VALUES ('org2','container2','entityId6',1.1);
> UPSERT INTO test.test VALUES ('org2','container1','entityId5',1.2);
> UPSERT INTO test.test VALUES ('org2','container2','entityId4',1.3);
> UPSERT INTO test.test VALUES ('org2','container1','entityId3',1.4);
> UPSERT INTO test.test VALUES ('org2','container3','entityId7',1.35);
> UPSERT INTO test.test VALUES ('org2','container3','entityId8',1.45);
> EXPLAIN
> SELECT DISTINCT entity_id, score
> FROM test.test
> WHERE organization_id = 'org2'
> AND container_id IN ( 'container1','container2','container3' )
> ORDER BY score DESC
> LIMIT 2
> OUTPUT
> entityId5 1.2
> entityId3 1.4
> The expected out out would be
> entityId8 1.45
> entityId3 1.4
> You will get the expected output if you remove the secondary index from the
> table or remove distinct from the query.
> As described in PHOENIX-3452 if you run the query without the LIMIT the
> ordering is not correct. However, the 2first results in that ordering is
> still not the onces returned by the limit clause, which makes me think there
> are multiple issues here and why I filed both separately. The rows being
> returned are the ones assigned to container1. It looks like Phoenix is first
> getting the rows from the first container and when it finds that to be enough
> it stops the scan. What it should be doing is getting 2 results for each
> container and then merge then and then limit again.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
