maedhroz commented on code in PR #3054:
URL: https://github.com/apache/cassandra/pull/3054#discussion_r1467202036
##########
src/java/org/apache/cassandra/index/sai/iterators/KeyRangeIntersectionIterator.java:
##########
@@ -77,7 +77,19 @@ protected PrimaryKey computeNext()
if (index != alreadyAvanced)
{
KeyRangeIterator range = ranges.get(index);
- PrimaryKey nextKey = nextOrNull(range, highestKey);
+ PrimaryKey nextKey = range.getCurrent();
+
+ // Note that we will either have a data model that
produces SKINNY primary keys or a data model
+ // that produces some combination of WIDE and STATIC
prikary keys.
+ if (nextKey.kind() == PrimaryKey.Kind.WIDE ||
nextKey.kind() == highestKey.kind())
+ // We can always skip if the target is of the same
kind or this range is non-static.
+ nextKey = nextOrNull(range, highestKey);
+ else if (nextKey.kind() == PrimaryKey.Kind.STATIC &&
nextKey.compareTo(highestKey) < 0)
+ // For a range of static keys, only skip if we'e
advanced to a new partition, and when we
+ // do, skip to an actual static key. We may otherwise
skip too far, as static row IDs always
+ // precede non-static ones in on-disk postings lists.
+ nextKey = nextOrNull(range, highestKey.toStatic());
Review Comment:
While I wasn't quite able to settle on a two-loop solution that I liked,
I've tried to a.) further document our assumptions, b.) create some reasonable
guardrails (see the new intersection compatibility check in `buildIterator()`
and `Kind#isIntersectable()`), and c.) expand the randomized tests even further.
I'm convinced, at this point, that the patch is correct, and that even a
rewrite of this intersection logic at a later time would have sufficient tests
ensuring it wouldn't make the same mistakes :)
Let me know what you think...
--
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]
