[
https://issues.apache.org/jira/browse/HIVE-8031?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14146874#comment-14146874
]
Mostafa Mokhtar commented on HIVE-8031:
---------------------------------------
Issue is that NDV is not scaled with row count as a results the CE of the join
ends up being too low.
Another issue is that the join of date_dim x store_sales doesn't go through the
PK/FK path HiveRelMdUniqueKeys.getUniqueKeys(ProjectRelBase, boolean) is of
type HepRelVertex not HiveTableScanRel.
> CBO should use per column join selectivity not NDV when applying exponential
> backoff.
> -------------------------------------------------------------------------------------
>
> Key: HIVE-8031
> URL: https://issues.apache.org/jira/browse/HIVE-8031
> Project: Hive
> Issue Type: Bug
> Components: CBO
> Affects Versions: 0.14.0, 0.13.1
> Reporter: Mostafa Mokhtar
> Assignee: Harish Butani
> Fix For: 0.14.0
>
>
> Currently CBO uses NDV not join selectivity in computeInnerJoinSelectivity
> which results in in-accurate estimate number of rows.
> I looked at the plan for TPC-DS Q17 after the latest set of changes and I am
> concerned that the estimate of rows for the join of store_sales and
> store_returns is so low, as you can see the estimate is 8461 rows for joining
> 1.2795706667449066E8 with 1.2922108035889767E7.
> {code}
> HiveJoinRel(condition=[AND(=($130, $3), =($129, $15))],
> joinType=[inner]): rowcount = 1079.1345153548855, cumulative cost =
> {8.271845957931738E10 rows, 0.0 cpu, 0.0 io}, id = 517
> HiveJoinRel(condition=[=($0, $38)], joinType=[inner]):
> rowcount = 6.669190301841249E7, cumulative cost = {4.300510912631623E10 rows,
> 0.0 cpu, 0.0 io}, id = 402
> HiveTableScanRel(table=[[catalog_sales]]): rowcount =
> 4.3005109025E10, cumulative cost = {0}, id = 2
> HiveFilterRel(condition=[in($15, '2000Q1', '2000Q2',
> '2000Q3')]): rowcount = 101.31622746185853, cumulative cost = {0.0 rows, 0.0
> cpu, 0.0 io}, id = 181
> HiveTableScanRel(table=[[d3]]): rowcount = 73049.0,
> cumulative cost = {0}, id = 3
> HiveJoinRel(condition=[AND(AND(=($3, $61), =($2, $60)),
> =($9, $67))], joinType=[inner]): rowcount = 8461.27236667537, cumulative cost
> = {8.26517592150266E10 rows, 0.0 cpu, 0.0 io}, id = 515
> HiveJoinRel(condition=[=($27, $0)], joinType=[inner]):
> rowcount = 1.2795706667449066E8, cumulative cost = {8.251088004031622E10
> rows, 0.0 cpu, 0.0 io}, id = 417
> HiveTableScanRel(table=[[store_sales]]): rowcount =
> 8.2510879939E10, cumulative cost = {0}, id = 5
> HiveFilterRel(condition=[=($15, '2000Q1')]): rowcount =
> 101.31622746185853, cumulative cost = {0.0 rows, 0.0 cpu, 0.0 io}, id = 173
> HiveTableScanRel(table=[[d1]]): rowcount = 73049.0,
> cumulative cost = {0}, id = 0
> HiveJoinRel(condition=[=($0, $24)], joinType=[inner]):
> rowcount = 1.2922108035889767E7, cumulative cost = {8.332595810316228E9 rows,
> 0.0 cpu, 0.0 io}, id = 424
> HiveTableScanRel(table=[[store_returns]]): rowcount =
> 8.332595709E9, cumulative cost = {0}, id = 7
> HiveFilterRel(condition=[in($15, '2000Q1', '2000Q2',
> '2000Q3')]): rowcount = 101.31622746185853, cumulative cost = {0.0 rows, 0.0
> cpu, 0.0 io}, id = 177
> HiveTableScanRel(table=[[d2]]): rowcount = 73049.0,
> cumulative cost = {0}, id = 1
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
