bersprockets opened a new pull request #23392: [SPARK-26450][SQL] Avoid rebuilding map of schema for every column in projection URL: https://github.com/apache/spark/pull/23392 ## What changes were proposed in this pull request? When creating some unsafe projections, Spark rebuilds the map of schema attributes once for each expression in the projection. Some file format readers create one unsafe projection per input file, others create one per task. ProjectExec also creates one unsafe projection per task. As a result, for wide queries on wide tables, Spark might build the map of schema attributes hundreds of thousands of times. This PR changes two functions to reuse the same AttributeSeq instance when creating BoundReference objects for each expression in the projection. This avoids the repeated rebuilding of the map of schema attributes. ### Benchmarks The time saved by this PR depends on size of the schema, size of the projection, number of input files (or number of file splits), number of tasks, and file format. I chose a couple of example cases. In the following tests, I ran the query ```sql select * from table where id1 = 1 ``` Matching rows are about 0.2% of the table. #### Orc table 6000 columns, 500K rows, 34 input files baseline | pr | improvement ----|----|---- 1.772306 min | 1.487267 min | 16.082943% #### Orc table 6000 columns, 500K rows, *17* input files baseline | pr | improvement ----|----|---- 1.656400 min | 1.423550 min | -14.057595% #### Orc table 60 columns, 50M rows, 34 input files baseline | pr | improvement ----|----|---- 0.299878 min | 0.290339 min | 3.180926% #### Parquet table 6000 columns, 500K rows, 34 input files baseline | pr | improvement ----|----|---- 1.478306 min | 1.373728 min | 7.074165% Note: The parquet reader does not create an unsafe projection. However, the filter operation in the query causes the planner to add a ProjectExec, which does create an unsafe projection for each task. So these results have nothing to do with Parquet itself. #### Parquet table 60 columns, 50M rows, 34 input files baseline | pr | improvement ----|----|---- 0.245006 min | 0.242200 min | 1.145099% #### CSV table 6000 columns, 500K rows, 34 input files baseline | pr | improvement ----|----|---- 2.390117 min | 2.182778 min | 8.674844% #### CSV table 60 columns, 50M rows, 34 input files baseline | pr | improvement ----|----|---- 1.520911 min | 1.510211 min | 0.703526% ## How was this patch tested? SQL unit tests Python core and SQL test
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
