[
https://issues.apache.org/jira/browse/HIVE-23903?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
mahesh kumar behera updated HIVE-23903:
---------------------------------------
Description:
In case of anti join, bloom filter can be created on left side also ("IN
(keylist right table)").But the filter should be "not-in" ("NOT IN (keylist
right table)") as we want to select the records from left side which are not
present in the right side. But it may cause wrong result as bloom filter may
have false positive and thus simply adding not is not correct, special handling
is required for "NOT IN".
[https://github.com/jmhodges/opposite_of_a_bloom_filter/]
was:
Currently hive does not support Anti join. The query for anti join is converted
to left outer join and null filter on right side join key is added to get the
desired result. This is causing
# Extra computation — The left outer join projects the redundant columns from
right side. Along with that, filtering is done to remove the redundant rows.
This is can be avoided in case of anti join as anti join will project only the
required columns and rows from the left side table.
# Extra shuffle — In case of anti join the duplicate records moved to join
node can be avoided from the child node. This can reduce significant amount of
data movement if the number of distinct rows( join keys) is significant.
# Extra Memory Usage - In case of map based anti join , hash set is sufficient
as just the key is required to check if the records matches the join
condition. In case of left join, we need the key and the non key columns also
and thus a hash table will be required.
For a query like
{code:java}
select wr_order_number FROM web_returns LEFT JOIN web_sales ON
wr_order_number = ws_order_number WHERE ws_order_number IS NULL;{code}
The number of distinct ws_order_number in web_sales table in a typical 10TB
TPCDS set up is just 10% of total records. So when we convert this query to
anti join, instead of 7 billion rows, only 600 million rows are moved to join
node.
In the current patch, just one conversion is done. The pattern of
project->filter->left-join is converted to project->anti-join. This will take
care of sub queries with “not exists” clause. The queries with “not exists” are
converted first to filter + left-join and then its converted to anti join. The
queries with “not in” are not handled in the current patch.
>From execution side, both merge join and map join with vectorized execution
>is supported for anti join.
> Support "not-in" for bloom filter
> ---------------------------------
>
> Key: HIVE-23903
> URL: https://issues.apache.org/jira/browse/HIVE-23903
> Project: Hive
> Issue Type: Bug
> Reporter: mahesh kumar behera
> Assignee: mahesh kumar behera
> Priority: Major
> Labels: pull-request-available
>
> In case of anti join, bloom filter can be created on left side also ("IN
> (keylist right table)").But the filter should be "not-in" ("NOT IN (keylist
> right table)") as we want to select the records from left side which are not
> present in the right side. But it may cause wrong result as bloom filter may
> have false positive and thus simply adding not is not correct, special
> handling is required for "NOT IN".
> [https://github.com/jmhodges/opposite_of_a_bloom_filter/]
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
