Github user srowen commented on a diff in the pull request:
https://github.com/apache/spark/pull/19438#discussion_r143324827
--- Diff:
sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/util/QuantileSummariesSuite.scala
---
@@ -58,7 +58,7 @@ class QuantileSummariesSuite extends SparkFunSuite {
if (data.nonEmpty) {
val approx = summary.query(quant).get
// The rank of the approximation.
- val rank = data.count(_ < approx) // has to be <, not <= to be exact
+ val rank = data.count(_ <= approx)
--- End diff --
This is trying to recover the rank of the element that was picked as the
quantile. I think that's problematic when the input repeats the value chosen as
the quantile. Consider estimating the median of [1,2,2,2,2,2,2,2,3]. If the
method correctly picks 2, depending on whether you define this test as < or <=,
you conclude that it picked rank 1 or 8 of 9 as the median. Any reasonable test
of whether that rank is near the expected 5 will fail either way in some cases.
One reasonable fix it to actually use the average of `data.count(_ <
approx)` and `data.count(_ <= approx)` as the implied rank that was chosen.
Do you know which test case failed without this change?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
