Feature/decimal support #982
Conversation
Co-authored-by: Steve Suh <[email protected]>
|
Can we get some traction on this one? It looks like this PR is ready and we have a lot of code that uses decimals that will be difficult to write tests around without this. |
| /// </summary> | ||
| /// <param name="s">The stream to write</param> | ||
| /// <param name="value">The decimal to write</param> | ||
| public static void Write(Stream s, decimal value) => Write(s, value.ToString()); |
There was a problem hiding this comment.
Should use ToString(CultureInfo.InvariantCulture) if we are using a string on the wire.
| @@ -267,6 +267,9 @@ private ISocketWrapper GetConnection() | |||
| case 'd': | |||
| returnValue = SerDe.ReadDouble(inputStream); | |||
| break; | |||
| case 'm': | |||
| returnValue = decimal.Parse(SerDe.ReadString(inputStream)); | |||
There was a problem hiding this comment.
Use decimal.Parse(SerDe.ReadString(inputStream), CultureInfo.InvariantCulture) to ensure we are using invariant culture on the wire.
| Row row = df.Collect().First(); | ||
| Assert.Equal(decimal.MinValue, row[0]); | ||
| Assert.Equal(decimal.MaxValue, row[1]); | ||
| Assert.Equal(decimal.Zero, row[2]); |
There was a problem hiding this comment.
I haven't gotten to dive deep into whether this is an issue yet, but want to bring it to attention just in case:
There was a time when we were comparing SQL Server output to Spark SQL output trying to migrate a pipeline to Synapse, and when attempting to diff two tables, found an issue with a double.
SQL Server uses, presumably, C#'s (and JavaScript, which the Python Notebook table preview in Synapse uses)'s conception of floats: -0.0 == 0.0, but the JVM/Spark in some cases compares by bit and differentiates because of the signed bit: -0.0 != 0.0.
It's resolved in later versions of Spark's DataFrames, and may not apply in the case of [decimal]String, so it may not be problematic.
- https://issues.apache.org/jira/browse/SPARK-26021
- https://www.mail-archive.com/[email protected]/msg283973.html
- https://issues.apache.org/jira/browse/SPARK-32110
- [SPARK-32110][SQL] normalize special floating numbers in HyperLogLog++ apache/spark#30673
- [BUG] -0.0 vs 0.0 is a hot mess NVIDIA/spark-rapids#294
There was a problem hiding this comment.
This is because internally, BigDecimal uses BigInteger, and BigInteger also only has a single concept of zero. A BigInteger behaves as a two's-complement integer, and two's-complement only has a single zero.
|
Any updates on this PR? |
|
Hey @GoEddie , are you still working on this? |
|
Hi @AFFogarty, I had given up really as no one seemed to be reviewing pr’s but am happy to get it up to date again. ed |
We are excited to review your PR.
So we can do the best job, please check:
Fixes #nnnnin your description to cause GitHub to automatically close the issue(s) when your PR is merged.This implements #818
On the Apache Spark side decimal is implemented using a java.math.BigDecimal and the only way to construct that is using a int/long - there is no way to construct it with any value larger than max long unless you use the string constructor so I pass a string back and forth between .NET and the JVM, hope that is ok.