You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@drill.apache.org by "Jinfeng Ni (JIRA)" <ji...@apache.org> on 2014/07/28 21:49:41 UTC
[jira] [Commented] (DRILL-1212) Remove extra exchange operator when
generate a two phase aggregation plan.
[ https://issues.apache.org/jira/browse/DRILL-1212?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14076694#comment-14076694 ]
Jinfeng Ni commented on DRILL-1212:
-----------------------------------
Here is part of plan containing one extra hash exchange operator.
01-02 Sort(sort0=[$0], dir0=[ASC])
01-03 HashToRandomExchange(dist0=[[$0]])
02-01 HashAgg(group=[{0}], high_line_count=[SUM($1)], low_line_count=[SUM($2)])
02-02 HashToRandomExchange(dist0=[[$0]])
03-01 HashAgg(group=[{0}], high_line_count=[SUM($1)], low_line_count=[SUM($2)])
Operator 01-03 is duplicate, since operator 02-02 will hash-distributed the data using the same keys.
> Remove extra exchange operator when generate a two phase aggregation plan.
> --------------------------------------------------------------------------
>
> Key: DRILL-1212
> URL: https://issues.apache.org/jira/browse/DRILL-1212
> Project: Apache Drill
> Issue Type: Bug
> Reporter: Jinfeng Ni
>
> We may see an extra hash exchange operator, when optimizer choose a two phase aggregation plan, for a query with group by GB_Key order by OB_Key and GB_key is same as Ob_key.
> One example of such plan with extra exchange is TPCH Q12:
> select
> l.l_shipmode,
> sum(case
> when o.o_orderpriority = '1-URGENT'
> or o.o_orderpriority = '2-HIGH'
> then 1
> else 0
> end) as high_line_count,
> sum(case
> when o.o_orderpriority <> '1-URGENT'
> and o.o_orderpriority <> '2-HIGH'
> then 1
> else 0
> end) as low_line_count
> from
> cp.`tpch/orders.parquet` o,
> cp.`tpch/lineitem.parquet` l
> where
> o.o_orderkey = l.l_orderkey
> and l.l_shipmode in ('TRUCK', 'REG AIR')
> and l.l_commitdate < l.l_receiptdate
> and l.l_shipdate < l.l_commitdate
> and l.l_receiptdate >= date '1994-01-01'
> and l.l_receiptdate < date '1994-01-01' + interval '1' year
> group by
> l.l_shipmode
> order by
> l.l_shipmode;
> The issue is caused by the rule does not propagate the hash distribution up, and hence introduce unnecessary hash exchange.
>
--
This message was sent by Atlassian JIRA
(v6.2#6252)