hanover-fiste commented on a change in pull request #31965:
URL: https://github.com/apache/spark/pull/31965#discussion_r602134395
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/jdbc/JDBCRelation.scala
##########
@@ -118,13 +119,28 @@ private[sql] object JDBCRelation extends Logging {
s"Upper bound: ${boundValueToString(upperBound)}.")
upperBound - lowerBound
}
- // Overflow and silliness can happen if you subtract then divide.
- // Here we get a little roundoff, but that's (hopefully) OK.
- val stride: Long = upperBound / numPartitions - lowerBound / numPartitions
+
+ // Overflow can happen if you subtract then divide. For example:
+ // (Long.MaxValue - Long.MinValue) / (numPartitions - 2).
+ // Also, using fixed-point decimals here to avoid possible inaccuracy from
floating point.
+ val strideUpperCalculation = (upperBound / BigDecimal(numPartitions))
+ .setScale(18, RoundingMode.HALF_EVEN)
+ val strideLowerCalculation = (lowerBound / BigDecimal(numPartitions))
+ .setScale(18, RoundingMode.HALF_EVEN)
+
+ val preciseStride = strideUpperCalculation - strideLowerCalculation
+ val stride = preciseStride.toLong
+
+ // Determine the number of strides the last partition will fall short of
compared to the
+ // supplied upper bound. Take half of those strides, and then add them to
the lower bound
+ // for better distribution of the first and last partitions.
+ val lostNumOfStrides = (preciseStride - stride) * numPartitions / stride
+ val lowerBoundWithStrideAlignment = lowerBound +
+ ((lostNumOfStrides / 2) * stride).setScale(0,
RoundingMode.HALF_UP).toLong
Review comment:
Definitely! I'll add a couple examples.
--
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
