[
https://issues.apache.org/jira/browse/DRILL-6836?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16701142#comment-16701142
]
Boaz Ben-Zvi commented on DRILL-6836:
-------------------------------------
Regarding a parallel plan, where there are multiple Hash-Aggr (2nd phase)
working in parallel, and a single Streaming-Aggr at the root fragment to return
the final count – this would +work the same way+, as the actual plan (see
below) places a Streaming-Aggr (01-01) on top of each 2nd phase Hash-Aggr
(01-02), to perform the COUNT, and the root Streaming-Aggr (00-02) is actually
performing a SUM. Here is an example plan (trimmed for readability):
{code:java}
00-00 Screen : rowType = RecordType(BIGINT EXPR$0): rowcount = 1.0,
cumulative cost = {}, id = 1698
00-01 Project(EXPR$0=[$0]) : rowType = RecordType(BIGINT EXPR$0): rowcount
= 1.0, cumulative cost = {}, id = 1697
00-02 StreamAgg(group=[{}], EXPR$0=[$SUM0($0)]) : rowType =
RecordType(BIGINT EXPR$0): rowcount = 1.0, cumulative cost = {}, id = 1696
00-03 UnionExchange : rowType = RecordType(BIGINT EXPR$0): rowcount =
1.0, cumulative cost = {}, id = 1695
01-01 StreamAgg(group=[{}], EXPR$0=[COUNT($0)]) : rowType =
RecordType(BIGINT EXPR$0): rowcount = 1.0, cumulative cost = {}, id = 1694
01-02 HashAgg(group=[{0}]) : rowType = RecordType(ANY full_name):
rowcount = 1.0, cumulative cost = {}, id = 1693
01-03 Project(full_name=[$0]) : rowType = RecordType(ANY
full_name): rowcount = 2.5, cumulative cost = {}, id = 1692
01-04 HashToRandomExchange(dist0=[[$0]]) : rowType =
RecordType(ANY full_name, ANY E_X_P_R_H_A_S_H_F_I_E_L_D): rowcount = 2.5,
cumulative cost = {}, id = 1691
02-01 UnorderedMuxExchange : rowType = RecordType(ANY
full_name, ANY E_X_P_R_H_A_S_H_F_I_E_L_D): rowcount = 2.5, cumulative cost =
{}, id = 1690
03-01 Project(full_name=[$0], E_X_P=[hash32AsDouble($0,
1301011)]) : rowType = RecordType(ANY, ANY E_X_P): rowcount = 2.5, cumulative
cost = {}, id = 1689
03-02 HashAgg(group=[{0}]) : rowType = RecordType(ANY
full_name): rowcount = 2.5, cumulative cost = {}, id = 1688
03-03 Scan(table=[[]], groupscan=[EasyGroupScan
[selectionRoot=file:, numFiles=9, files=[]) : rowType = : rowcount =,
cumulative cost = {}, id = 1687
{code}
> 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: 1.16.0
>
>
> 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)
