rich7420 opened a new issue, #6044:
URL: https://github.com/apache/datafusion-comet/issues/6044
### Describe the bug
Native variance and standard deviation lose accuracy for large, nearby
DOUBLE values. For `10000000000000000` and `10000000000000002`, `var_pop`
returns 2 instead of Spark's 1. This occurs with and without `GROUP BY`.
### Steps to reproduce
With Comet and native shuffle enabled, write one Parquet file to preserve
input order:
```scala
import spark.implicits._
spark.conf.set("spark.sql.adaptive.enabled", "false")
val path =
java.nio.file.Files.createTempDirectory("variance-repro").resolve("data").toString
Seq((0, 10000000000000000d), (0, 10000000000000002d))
.toDF("g", "v").coalesce(1).write.parquet(path)
spark.read.parquet(path).createOrReplaceTempView("variance_repro")
```
```sql
SELECT var_pop(v), var_samp(v), stddev_pop(v), stddev_samp(v)
FROM variance_repro;
```
Repeat with `GROUP BY g`.
### Expected behavior
| Expression | Spark | Comet |
| --- | --- | --- |
| var_pop | 1.0 | 2.0 |
| var_samp | 2.0 | 4.0 |
| stddev_pop | 1.0 | 1.4142135623730951 |
| stddev_samp | 1.4142135623730951 | 2.0 |
### Additional context
Reproduced on main `b7f35b6ac`, Spark 3.5.9 and 4.1.3, with native Partial
and Final aggregates asserted.
`welford::variance_update` computes `delta * (value - new_mean)`. Spark's
`CentralMomentAgg` computes `delta * (delta - delta / new_count)`. Rounding the
large `new_mean` before subtraction makes Comet add 4 to M2 where Spark adds 2.
The shared helper also serves `corr`, whose Spark formula differs from
`CentralMomentAgg`; a fix needs to preserve that distinction. Unlike the
tolerance cases in #1375 and #392, this example has a 100% relative variance
error.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
