[
https://issues.apache.org/jira/browse/HIVE-18454?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16327141#comment-16327141
]
Zoltan Haindrich commented on HIVE-18454:
-----------------------------------------
my guess is that the join codes which are responsible for estimating the
rowcounts for cross joins are work by the edge type; which is BROADCAST in this
case
> Incorrect rownum estimation in joins
> ------------------------------------
>
> Key: HIVE-18454
> URL: https://issues.apache.org/jira/browse/HIVE-18454
> Project: Hive
> Issue Type: Sub-task
> Components: Statistics
> Reporter: Zoltan Haindrich
> Priority: Major
>
> I've seen this probably earlier...row counts seems to be off the
> charts...120000 rows estimated when the table has only 10 rows
> {code}
> create table s (x int);
> insert into s values
> (1),(2),(3),(4),(5),
> (6),(7),(8),(9),(10);
> create table tu(id_uv int,id_uw int,u int);
> create table tv(id_uv int,v int);
> create table tw(id_uw int,w int);
> from s
> insert overwrite table tu
> select x,x,x
> where x<=6 or x=10
> insert overwrite table tv
> select x,x
> where x<=3 or x=10
> insert overwrite table tw
> select x,x
> ;
> set hive.explain.user=true;
> explain analyze
> select sum(u*v*w) from tu
> join tv on (tu.id_uv=tv.id_uv)
> join tw on (tu.id_uw=tw.id_uw)
> where w>9 and u>1 and v>3;
> desc formatted tv;
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
