[jira] [Updated] (HIVE-12336) Sort Merge Partition Map Join

["Maciek Kocon (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HIVE-12336?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Maciek Kocon updated HIVE-12336:
--------------------------------
    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 SORT MERGE BUCKETED MAP JOIN for partitioned tables.
The sort-merge join optimization could be performed when PARTITIONED tables being joined are sorted and partitioned on the join columns.

The corresponding partitions are joined with each other at the mapper. If both A and B have partitions set on their columns KEY, the following join
SELECT /*+ MAPJOIN(b) */ a.key, a.value
FROM A a JOIN B b ON a.key = b.key
can be done on the mapper only. The mapper for the partition key='201512' for A will traverse the corresponding partition for B. Traversing is possible if the corresponding partitions are sorted on the same columns.

  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.


> Sort Merge Partition Map Join
> -----------------------------
>
>                 Key: HIVE-12336
>                 URL: https://issues.apache.org/jira/browse/HIVE-12336
>             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 SORT MERGE BUCKETED MAP JOIN for partitioned tables.
> The sort-merge join optimization could be performed when PARTITIONED tables being joined are sorted and partitioned on the join columns.
> The corresponding partitions are joined with each other at the mapper. If both A and B have partitions set on their columns KEY, the following join
> SELECT /*+ MAPJOIN(b) */ a.key, a.value
> FROM A a JOIN B b ON a.key = b.key
> can be done on the mapper only. The mapper for the partition key='201512' for A will traverse the corresponding partition for B. Traversing is possible if the corresponding partitions are sorted on the same columns.



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