[
https://issues.apache.org/jira/browse/HIVE-20720?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Daniel Dai updated HIVE-20720:
------------------------------
Description:
Currently JdbcStorageHandler does not split input in Tez. The reason is
numSplit of JdbcInputFormat.getSplits can only pass via "mapreduce.job.maps" in
Tez. And "mapreduce.job.maps" is not a valid param if authorizer(eg.
SQLStdAuth) is in use. User ends up always use 1 split.
We need to rely on this new feature if we want to support multi-splits. Here is
my proposal:
1. Specify partitionColumn/numPartitions, and optional lowerBound/upperBound in
tblproperties if user want to split jdbc data source. In case
lowerBound/upperBound is not specified, JdbcStorageHandler will run max/min
query to get this in planner. We can currently limit partitionColumn to only
numeric/date/timestamp column for simplicity
2. If partitionColumn/numPartitions are not specified, don't split input
3. Splits are equal intervals without respect to data distribution
4. There is also a "hive.sql.query.split" flag vetos the split (can be set
manually or automatically by calcite)
5. If partitionColumn is not defined, but numPartitions is defined, use
original limit/offset logic (however, don't rely on numSplit).
was:
Currently JdbcStorageHandler does not split input in Tez. The reason is
numSplit of JdbcInputFormat.getSplits can only pass via "mapreduce.job.maps" in
Tez. And "mapreduce.job.maps" is not a valid param if ranger is in use. User
ends up always use 1 split.
We need to rely on this new feature if we want to support multi-splits. Here is
my proposal:
1. Specify partitionColumn/numPartitions, and optional lowerBound/upperBound in
tblproperties if user want to split jdbc data source. In case
lowerBound/upperBound is not specified, JdbcStorageHandler will run max/min
query to get this in planner. We can currently limit partitionColumn to only
numeric/date/timestamp column for simplicity
2. If partitionColumn/numPartitions are not specified, don't split input
3. Splits are equal intervals without respect to data distribution
4. There is also a "hive.sql.query.split" flag vetos the split (can be set
manually or automatically by calcite)
5. If partitionColumn is not defined, but numPartitions is defined, use
original limit/offset logic (however, don't rely on numSplit).
> Add partition column option to JDBC handler
> -------------------------------------------
>
> Key: HIVE-20720
> URL: https://issues.apache.org/jira/browse/HIVE-20720
> Project: Hive
> Issue Type: New Feature
> Components: StorageHandler
> Reporter: Daniel Dai
> Assignee: Daniel Dai
> Priority: Major
> Attachments: HIVE-20720.1.patch
>
>
> Currently JdbcStorageHandler does not split input in Tez. The reason is
> numSplit of JdbcInputFormat.getSplits can only pass via "mapreduce.job.maps"
> in Tez. And "mapreduce.job.maps" is not a valid param if authorizer(eg.
> SQLStdAuth) is in use. User ends up always use 1 split.
> We need to rely on this new feature if we want to support multi-splits. Here
> is my proposal:
> 1. Specify partitionColumn/numPartitions, and optional lowerBound/upperBound
> in tblproperties if user want to split jdbc data source. In case
> lowerBound/upperBound is not specified, JdbcStorageHandler will run max/min
> query to get this in planner. We can currently limit partitionColumn to only
> numeric/date/timestamp column for simplicity
> 2. If partitionColumn/numPartitions are not specified, don't split input
> 3. Splits are equal intervals without respect to data distribution
> 4. There is also a "hive.sql.query.split" flag vetos the split (can be set
> manually or automatically by calcite)
> 5. If partitionColumn is not defined, but numPartitions is defined, use
> original limit/offset logic (however, don't rely on numSplit).
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
