[
https://issues.apache.org/jira/browse/SPARK-59347?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anish Mahto updated SPARK-59347:
--------------------------------
Description:
Part of the contract for the SCD2 version map is the keys in the version map
should track exactly with the data column names.
Complexity arises when case insensitivity is used, and the incoming microbatch
has a different case-spelling for certain columns than the persisted target/aux
tables.
When `Scd2ForeachBatchHandler` does a [union with the incoming microbatch rows
and the existing target/aux table
rows|https://github.com/apache/spark/blob/master/sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd2ForeachBatchHandler.scala#L104-L106],
due to ordering of the union operators, the resulting DF will respect the
case-spelling on the incoming microbatch.
But when this dataframe is eventually merged back into the target/aux tables,
the existing case-spelling on those tables will win. This is no longer
acceptable since the version map is constructed by referencing column names in
the microbatch dataframe. If those column names change, the version map's keys
become inconsistent.
The proposal is to union the target schema onto the microbatch dataframe during
microbatch preprocessing, and before we eventually construct the version map.
was:
Part of the contract for the SCD2 version map is the keys in the version map
should track exactly with the data column names.
Complexity arises when case insensitivity is used, and the incoming microbatch
has a different case-spelling for certain columns than the persisted target/aux
tables.
When `Scd2ForeachBatchHandler` does a [union with the incoming microbatch rows
and the existing target/aux table
rows|https://github.com/apache/spark/blob/master/sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd2ForeachBatchHandler.scala#L104-L106],
due to ordering of the union operators, the resulting DF will respect the
case-spelling on the incoming microbatch.
But when this dataframe is eventually merged back into the target/aux tables,
the existing case-spelling on those tables will win. This is no longer
acceptable since the version map is constructed by referencing column names in
the dataframe. If those column names change, the version map's keys need to
also change for consistency.
The proposal is two-fold:
1. The 3-way union should setup target table's df as the first operator, so
that from this point forward rows always respect the persisted target table's
case-spellings
2. Rename the in-memory version map keys on `preprocessedBatchDf` to use the
same case-spelling as the target table, right before the union operation
Then both before and after the union, the version map keys match 1:1 with data
column names, under whatever case-sensitivity scheme is configured.
> SCD2 Ignore-null support; union microbatch with target table during
> preprocessing
> ---------------------------------------------------------------------------------
>
> Key: SPARK-59347
> URL: https://issues.apache.org/jira/browse/SPARK-59347
> Project: Spark
> Issue Type: Sub-task
> Components: Declarative Pipelines
> Affects Versions: 4.4.0
> Reporter: Anish Mahto
> Priority: Major
>
> Part of the contract for the SCD2 version map is the keys in the version map
> should track exactly with the data column names.
> Complexity arises when case insensitivity is used, and the incoming
> microbatch has a different case-spelling for certain columns than the
> persisted target/aux tables.
> When `Scd2ForeachBatchHandler` does a [union with the incoming microbatch
> rows and the existing target/aux table
> rows|https://github.com/apache/spark/blob/master/sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd2ForeachBatchHandler.scala#L104-L106],
> due to ordering of the union operators, the resulting DF will respect the
> case-spelling on the incoming microbatch.
> But when this dataframe is eventually merged back into the target/aux tables,
> the existing case-spelling on those tables will win. This is no longer
> acceptable since the version map is constructed by referencing column names
> in the microbatch dataframe. If those column names change, the version map's
> keys become inconsistent.
> The proposal is to union the target schema onto the microbatch dataframe
> during microbatch preprocessing, and before we eventually construct the
> version map.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
