[ https://issues.apache.org/jira/browse/DRILL-6836?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Boaz Ben-Zvi updated DRILL-6836: -------------------------------- Fix Version/s: (was: 1.16.0) 1.17.0 > Eliminate StreamingAggr for COUNT DISTINCT > ------------------------------------------ > > Key: DRILL-6836 > URL: https://issues.apache.org/jira/browse/DRILL-6836 > Project: Apache Drill > Issue Type: Improvement > Components: Execution - Relational Operators, Query Planning & > Optimization > Affects Versions: 1.14.0 > Reporter: Boaz Ben-Zvi > Assignee: Boaz Ben-Zvi > Priority: Minor > Fix For: 1.17.0 > > > The COUNT DISTINCT operation is often implemented with a Hash-Aggr operator > for the DISTINCT, and a Streaming-Aggr above to perform the COUNT. That > Streaming-Aggr does the counting like any aggregation, counting each value, > batch after batch. > While very efficient, that counting work is basically not needed, as the > Hash-Aggr knows the number of distinct values (in the in-memory partitions). > Hence _a possible small performance improvement_ - eliminate the > Streaming-Aggr operator, and notify the Hash-Aggr to return a COUNT (these > are Planner changes). The Hash-Aggr operator would need to generate the > single Float8 column output schema, and output that batch with a single > value, just like the Streaming -Aggr did (likely without generating code). > In case of a spill, the Hash-Aggr still needs to read and process those > partitions, to get the exact distinct number. > The expected improvement is the elimination of the batch by batch output > from the Hash-Aggr, and the batch by batch, row by row processing of the > Streaming-Aggr. -- This message was sent by Atlassian JIRA (v7.6.3#76005)