[BUG] cast(9.95 as decimal(3,1)), actual: 9.9, expected: 10.0

[gerashegalov]

Repro

 ~/dist/spark-3.3.0-bin-hadoop3/bin/spark-shell  \
   --conf spark.plugins=com.nvidia.spark.SQLPlugin \
   --conf spark.rapids.sql.test.enabled=true \
   --conf spark.rapids.sql.explain=ALL 
   --jars dist/target/rapids-4-spark_2.12-24.06.0-SNAPSHOT-cuda11.jar 

scala> val df = Seq(9.95).toDF.coalesce(1)
df: org.apache.spark.sql.Dataset[org.apache.spark.sql.Row] = [value: double]

scala> spark.conf.set("spark.rapids.sql.enabled", true)

scala> df.selectExpr("cast(value as decimal(3,1))").collect()
24/05/13 11:34:42 WARN GpuOverrides: 
*Exec <ProjectExec> will run on GPU
  *Expression <Alias> cast(value#1 as decimal(3,1)) AS value#5 will run on GPU
    *Expression <Cast> cast(value#1 as decimal(3,1)) will run on GPU
  *Exec <CoalesceExec> will run on GPU
    ! <LocalTableScanExec> cannot run on GPU because GPU does not currently support the operator class org.apache.spark.sql.execution.LocalTableScanExec
      @Expression <AttributeReference> value#1 could run on GPU

res2: Array[org.apache.spark.sql.Row] = Array([9.9])

scala> spark.conf.set("spark.rapids.sql.enabled", false)

scala> df.selectExpr("cast(value as decimal(3,1))").collect()
res4: Array[org.apache.spark.sql.Row] = Array([10.0])

Related to #9682

[mattahrens]

Scope: make casting of floats to decimals feature flag off by default and update documentation accordingly with this example. Ref: spark.rapids.sql.castFloatToDecimal.enabled. Check if supported operators in tools needs to be updated.

[ttnghia]

Filed a cudf issue: rapidsai/cudf#15862

[gerashegalov]

Also take a look at the discussion around #9682 (comment) and above. 9.95 is not representable as double

$ jshell 
|  Welcome to JShell -- Version 21.0.2
|  For an introduction type: /help intro

jshell> new BigDecimal(9.95)
$8 ==> 9.949999999999999289457264239899814128875732421875

jshell> new BigDecimal(9.949999999999999289457264239899814128875732421875)
$9 ==> 9.949999999999999289457264239899814128875732421875

jshell> new BigDecimal(9.95).setScale(1, BigDecimal.ROUND_HALF_UP)
$11 ==> 9.9

jshell> new BigDecimal(String.valueOf(9.95)).setScale(1, BigDecimal.ROUND_HALF_UP)
$1 ==> 10.0

jshell> new BigDecimal("9.95").setScale(1, BigDecimal.ROUND_HALF_UP)
$12 ==> 10.0

[ttnghia]

Okay after reading through the issue #9682 then I realize that this issue is also one instance of it.

[thirtiseven]

I tried the float => string => decimal path (it is very easy to implement in plugin), it can pass the Spark UT, but still some differences from the known limits of ryu float to string and a very edge cases in string to decimal #10890. I will post a pr and share some results for review next week, but not sure if the diffs are acceptable or original way can match it better.

[ttnghia]

That sounds good. I've also found a way to implement in C++ which is very efficient but not sure if it will pass integration test. I'll verify that and will post a PR too.

If having a chance, please list the related tests that I can run to verify.

[ttnghia]

@thirtiseven Please test #10917 to see if you have any test fails. On my local env, all the unit tests and integration tests passed but I'm not sure if I missed anything.