Skip to content
Snippets Groups Projects
Commit d6cbec75 authored by hyukjinkwon's avatar hyukjinkwon Committed by Wenchen Fan
Browse files

[SPARK-18943][SQL] Avoid per-record type dispatch in CSV when reading

## What changes were proposed in this pull request?

`CSVRelation.csvParser` does type dispatch for each value in each row. We can prevent this because the schema is already kept in `CSVRelation`.

So, this PR proposes that converters are created first according to the schema, and then apply them to each.

I just ran some small benchmarks as below after resembling the logics in https://github.com/apache/spark/blob/7c33b0fd050f3d2b08c1cfd7efbff8166832c1af/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVFileFormat.scala#L170-L178 to test the updated logics.

```scala
test("Benchmark for CSV converter") {
  var numMalformedRecords = 0
  val N = 500 << 12
  val schema = StructType(
    StructField("a", StringType) ::
    StructField("b", StringType) ::
    StructField("c", StringType) ::
    StructField("d", StringType) :: Nil)

  val row = Array("1.0", "test", "2015-08-20 14:57:00", "FALSE")
  val data = spark.sparkContext.parallelize(List.fill(N)(row))
  val parser = CSVRelation.csvParser(schema, schema.fieldNames, CSVOptions())

  val benchmark = new Benchmark("CSV converter", N)
  benchmark.addCase("cast CSV string tokens", 10) { _ =>
    data.flatMap { recordTokens =>
      parser(recordTokens, numMalformedRecords)
    }.collect()
  }
  benchmark.run()
}
```

**Before**

```
CSV converter:                           Best/Avg Time(ms)    Rate(M/s)   Per Row(ns)   Relative
------------------------------------------------------------------------------------------------
cast CSV string tokens                        1061 / 1130          1.9         517.9       1.0X
```

**After**

```
CSV converter:                           Best/Avg Time(ms)    Rate(M/s)   Per Row(ns)   Relative
------------------------------------------------------------------------------------------------
cast CSV string tokens                         940 / 1011          2.2         459.2       1.0X
```

## How was this patch tested?

Tests in `CSVTypeCastSuite` and `CSVRelation`

Author: hyukjinkwon <gurwls223@gmail.com>

Closes #16351 from HyukjinKwon/type-dispatch.
parent f2ceb2ab
No related branches found
No related tags found
No related merge requests found
Loading
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment