hive-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Maciek Kocon (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (HIVE-12334) Partition Map Join
Date Wed, 04 Nov 2015 15:25:28 GMT

     [ https://issues.apache.org/jira/browse/HIVE-12334?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Maciek Kocon updated HIVE-12334:
--------------------------------
    Description: 
Logically and functionally bucketing and partitioning are quite similar - both provide mechanism
to segregate and separate the table's data based on its content. Thanks to that significant
further optimisations like [partition] PRUNING or [bucket] MAP JOIN are possible.
The difference seems to be imposed by design where the PARTITIONing is open/explicit while
BUCKETing is discrete/implicit.
Partitioning seems to be very common if not a standard feature in all current RDBMS while
BUCKETING seems to be HIVE specific only.
In a way BUCKETING could be also called by "hashing" or simply "IMPLICIT PARTITIONING".

Regardless of the fact that these two are recognised as two separate features available in
Hive there should be nothing to prevent leveraging same existing query/join optimisations
across the two.

PARTITION MAPJOIN
Use the same type of optimization as in BUCKETED MAP JOIN when PARTITIONED tables being joined
are partitioned on the join columns:

If table A has set partitioning on KEY column and table B is partitioned on KEY column, the
following join
SELECT /*+ MAPJOIN(b) */ a.key, a.value
FROM a JOIN b ON a.key = b.key
can be done on the mapper only. Instead of fetching B completely for each mapper of A, only
the required partitions are fetched. For the query above, the mapper processing partition
key='part_key_value' for A will only fetch partition for key='part_key_value' of B.

  was:
Logically and functionally bucketing and partitioning are quite similar - both provide mechanism
to segregate and separate the table's data based on its content. Thanks to that significant
further optimisations like [partition] PRUNING or [bucket] MAP JOIN are possible.
The difference seems to be imposed by design where the PARTITIONing is open/explicit while
BUCKETing is discrete/implicit.
Partitioning seems to be very common if not a standard feature in all current RDBMS while
BUCKETING seems to be HIVE specific only.
In a way BUCKETING could be also called by "hashing" or simply "IMPLICIT PARTITIONING".

Regardless of the fact that these two are recognised as two separate features available in
Hive there should be nothing to prevent leveraging same existing query/join optimisations
across the two.


①[Sort Merge] PARTITION Map join (no progress yet)
Enable Bucket Map Join or better, the Sort Merge Bucket Map Join equivalent optimisations
when PARTITIONING is used exclusively or in combination with BUCKETING.

For JOIN conditions where partitioning criteria are used respectively:
            ⋮ 
FROM TabA JOIN TabB
   ON TabA.partCol1 = TabB.partCol2
   AND TabA.partCol2 = TabB.partCol2

the optimizer could/should choose to treat it the same way as with bucketed tables: ⋮ 
FROM TabC
  JOIN TabD
     ON TabC.clusteredByCol1 = TabD.clusteredByCol2
   AND TabC.clusteredByCol2 = TabD.clusteredByCol2

and use either Bucket Map Join or better, the Sort Merge Bucket Map Join. The latter would
require capability to create sorted partitions first.

This is based on fact that same way as buckets translate to separate files, the partitions
essentially provide the same mapping.
When data locality is known the optimizer could focus only on joining corresponding partitions
rather than whole data sets.

②BUCKET pruning (taken care by [HIVE-11525|https://issues.apache.org/jira/browse/HIVE-11525])
Enable partition PRUNING equivalent optimisation for queries on BUCKETED tables

Simplest example is for queries like:
"SELECT … FROM x WHERE colA=123123"
to read only the relevant bucket file rather than all file-buckets that belong to a table.


> Partition Map Join
> ------------------
>
>                 Key: HIVE-12334
>                 URL: https://issues.apache.org/jira/browse/HIVE-12334
>             Project: Hive
>          Issue Type: Improvement
>          Components: Logical Optimizer, Physical Optimizer, SQL
>    Affects Versions: 0.13.0, 0.14.0, 0.13.1, 1.0.0, 1.1.0
>            Reporter: Maciek Kocon
>              Labels: gsoc2015
>
> Logically and functionally bucketing and partitioning are quite similar - both provide
mechanism to segregate and separate the table's data based on its content. Thanks to that
significant further optimisations like [partition] PRUNING or [bucket] MAP JOIN are possible.
> The difference seems to be imposed by design where the PARTITIONing is open/explicit
while BUCKETing is discrete/implicit.
> Partitioning seems to be very common if not a standard feature in all current RDBMS while
BUCKETING seems to be HIVE specific only.
> In a way BUCKETING could be also called by "hashing" or simply "IMPLICIT PARTITIONING".
> Regardless of the fact that these two are recognised as two separate features available
in Hive there should be nothing to prevent leveraging same existing query/join optimisations
across the two.
> PARTITION MAPJOIN
> Use the same type of optimization as in BUCKETED MAP JOIN when PARTITIONED tables being
joined are partitioned on the join columns:
> If table A has set partitioning on KEY column and table B is partitioned on KEY column,
the following join
> SELECT /*+ MAPJOIN(b) */ a.key, a.value
> FROM a JOIN b ON a.key = b.key
> can be done on the mapper only. Instead of fetching B completely for each mapper of A,
only the required partitions are fetched. For the query above, the mapper processing partition
key='part_key_value' for A will only fetch partition for key='part_key_value' of B.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message