[
https://issues.apache.org/jira/browse/DRILL-6307?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16453304#comment-16453304
]
ASF GitHub Bot commented on DRILL-6307:
---------------------------------------
Github user paul-rogers commented on a diff in the pull request:
https://github.com/apache/drill/pull/1228#discussion_r184244479
--- Diff:
exec/java-exec/src/main/java/org/apache/drill/exec/record/RecordBatchSizer.java
---
@@ -277,18 +286,29 @@ public boolean isRepeatedList() {
/**
* This is the average per entry width, used for vector allocation.
*/
- public int getEntryWidth() {
+ private int getEntryWidthForAlloc() {
int width = 0;
if (isVariableWidth) {
- width = getNetSizePerEntry() - OFFSET_VECTOR_WIDTH;
+ width = getAllocSizePerEntry() - OFFSET_VECTOR_WIDTH;
// Subtract out the bits (is-set) vector width
- if (metadata.getDataMode() == DataMode.OPTIONAL) {
+ if (isOptional) {
width -= BIT_VECTOR_WIDTH;
}
+
+ if (isRepeated && getValueCount() == 0) {
+ return (safeDivide(width, STD_REPETITION_FACTOR));
+ }
}
- return (safeDivide(width, cardinality));
+ return (safeDivide(width, getEntryCardinalityForAlloc()));
+ }
+
+ /**
+ * This is the average per entry cardinality, used for vector
allocation.
+ */
+ private float getEntryCardinalityForAlloc() {
+ return getCardinality() == 0 ? (isRepeated ? STD_REPETITION_FACTOR :
1) :getCardinality();
--- End diff --
Makes sense, but why would a batch be empty unless that path hit EOF?
Otherwise, the batch might be due to an empty input file. We'd just skip it and
move to the next batch until we find one with data. Any reason the "get next
batch" code can just loop to be "get next non-empty batch" instead? Otherwise,
we can't really do any effective batch sizing as we have no data to go on...
> Handle empty batches in record batch sizer correctly
> ----------------------------------------------------
>
> Key: DRILL-6307
> URL: https://issues.apache.org/jira/browse/DRILL-6307
> Project: Apache Drill
> Issue Type: Bug
> Components: Execution - Flow
> Affects Versions: 1.13.0
> Reporter: Padma Penumarthy
> Assignee: Padma Penumarthy
> Priority: Major
> Fix For: 1.14.0
>
>
> when we get empty batch, record batch sizer calculates row width as zero. In
> that case, we do not do accounting and memory allocation correctly for
> outgoing batches.
> For example, in merge join, for outer left join, if right side batch is
> empty, we still have to include the right side columns as null in outgoing
> batch.
> Say first batch is empty. Then, for outgoing, we allocate empty vectors with
> zero capacity. When we read the next batch with data, we will end up going
> through realloc loop. If we use right side row width as 0 in outgoing row
> width calculation, number of rows we will calculate will be higher and later
> when we get a non empty batch, we might exceed the memory limits.
> One possible workaround/solution : Allocate memory based on std size for
> empty input batch. Use allocation width as width of the batch in number of
> rows calculation.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
