Work on TRAFODION-2234 started by Anoop Sharma.
> turn aligned row format for tables to ON by default
> ---------------------------------------------------
>                 Key: TRAFODION-2234
>                 URL: https://issues.apache.org/jira/browse/TRAFODION-2234
>             Project: Apache Trafodion
>          Issue Type: Improvement
>            Reporter: Anoop Sharma
>            Assignee: Anoop Sharma
>            Priority: Minor
> Columns in Trafodion tables are stored in 2 formats:
> -- regular hbase format where each column is stored as one cell
> -- aligned format where the whole row is packed and stored in one cell
> Aligned row provides performance boost during inserts and selects by
> retrieving one cell from hbase instead of multiple cells. As the number
> of columns in a table increase, perf of aligned format gets better.
> There are some limitations with aligned format:
> -- selection predicates cannot be pushed down to hbase region server
> -- all columns need to be retrieved and updated as packed row in a cell
> -- columns cannot be dropped without reloading the table
> Over time, these limitations will be removed by use of user defined filters
> and coprocessors to select/project rows at hbase region level.
> During perf runs, the pros for aligned format outweigh the cons.
> This jira is being filed to change the default from hbase format row
> to aligned format row. Code for both aligned and hbase format already
> exists and is being used.
> A table can always be created in either of these 2 formats by explicitly
> specifying the format during create time.
> The default can also be changed to off or on by inserting the appropriate
> value in the system defaults table.
> Turning on aligned format as default will be done in 2 phases:
> -- in phase 1, aligned default will be turning on during dev regressions run
>   until it has stabilized.
> -- in phase 2, system default will be changed to aligned. All table created
> without an explicit format specification will be created in aligned format.
> Metadata, repository, privilege and histogram tables will always be 
> created in hbase format. This is needed for backward compatibility.
> Any component or application that doesn't want to depend on the system
> default must explicitly specify the row format in their create ddl.

This message was sent by Atlassian JIRA

Reply via email to