[
https://issues.apache.org/jira/browse/HIVE-21857?focusedWorklogId=265034&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-265034
]
ASF GitHub Bot logged work on HIVE-21857:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 21/Jun/19 22:17
Start Date: 21/Jun/19 22:17
Worklog Time Spent: 10m
Work Description: jcamachor commented on pull request #671: HIVE-21857
URL: https://github.com/apache/hive/pull/671#discussion_r296410050
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/stats/HiveRelMdDistinctRowCount.java
##########
@@ -61,17 +62,11 @@ private HiveRelMdDistinctRowCount() {
@Override
public Double getDistinctRowCount(RelNode rel, RelMetadataQuery mq,
ImmutableBitSet groupKey,
RexNode predicate) {
- if (rel instanceof HiveTableScan) {
- return getDistinctRowCount((HiveTableScan) rel, mq, groupKey, predicate);
- }
- /*
- * For now use Calcite' default formulas for propagating NDVs up the Query
- * Tree.
- */
- return super.getDistinctRowCount(rel, mq, groupKey, predicate);
+ return NumberUtil.multiply(mq.getRowCount(rel),
Review comment:
The logic in the super method was:
```
// REVIEW zfong 4/19/06 - Broadbase code does not take into
// consideration selectivity of predicates passed in. Also, they
// assume the rows are unique even if the table is not
boolean uniq = RelMdUtil.areColumnsDefinitelyUnique(mq, rel, groupKey);
if (uniq) {
return NumberUtil.multiply(mq.getRowCount(rel),
mq.getSelectivity(rel, predicate));
}
return null;
```
I cannot figure out the logic behind it... if columns are definitely unique,
return distinct row count, otherwise return null?
In any case, I thought about this again and I changed the logic. I left
implementation as it was (I removed the method because it was just calling to
the super method) and I changed slightly the logic in FilterEstimator so it can
handle null values. Basically, null is unknown and you just push those
predicates in the AND / OR to the very end since you do not know how selective
they are. This may happen when we find an operator in the tree that is not
recognized yet in RelMdDistinctRowCount (e.g. DruidQuery).
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
Issue Time Tracking
-------------------
Worklog Id: (was: 265034)
Time Spent: 1h 10m (was: 1h)
> Sort conditions in a filter predicate to accelerate query processing
> --------------------------------------------------------------------
>
> Key: HIVE-21857
> URL: https://issues.apache.org/jira/browse/HIVE-21857
> Project: Hive
> Issue Type: New Feature
> Components: CBO
> Reporter: Jesus Camacho Rodriguez
> Assignee: Jesus Camacho Rodriguez
> Priority: Major
> Labels: pull-request-available
> Attachments: HIVE-21857.01.patch, HIVE-21857.02.patch,
> HIVE-21857.03.patch, HIVE-21857.04.patch
>
> Time Spent: 1h 10m
> Remaining Estimate: 0h
>
> Following approach similar to
> http://db.cs.berkeley.edu/jmh/miscpapers/sigmod93.pdf .
> To reorder predicates in AND conditions, we could rank each of elements in
> the clauses in increasing order based on following formula:
> {code}
> rank = (selectivity - 1) / cost per tuple
> {code}
> Similarly, for OR conditions:
> {code}
> rank = (-selectivity) / cost per tuple
> {code}
> Selectivity can be computed with FilterSelectivityEstimator. For cost per
> tuple, we will need to come up with some heuristic based on how expensive is
> the evaluation of the functions contained in that predicate. Custom UDFs
> could be annotated.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
