github-actions[bot] commented on code in PR #66447:
URL: https://github.com/apache/doris/pull/66447#discussion_r3713571033
##########
fe/fe-core/src/main/java/org/apache/doris/cloud/catalog/CloudTabletRebalancer.java:
##########
@@ -1083,6 +1097,22 @@ void fillBeToTablets(Long be, Long tableId, Long partId,
Long indexId, Long tabl
beToTabletsOfIndex.computeIfAbsent(be, ignored ->
ConcurrentHashMap.newKeySet()).add(tabletId);
}
+ private Function<Long, Set<Long>> newGlobalTabletSetFactory(Map<Long,
Set<Long>> previousBeToTablets) {
+ Map<Long, Set<Long>> previousRoute = previousBeToTablets == null
+ ? Collections.emptyMap() : previousBeToTablets;
+ return be -> {
+ Set<Long> previousTablets = previousRoute.get(be);
+ int initialCapacity = previousTablets == null ? 0 :
previousTablets.size();
Review Comment:
[P2] Bound stale cardinality before allocating the new set
`statRouteInfo()` rebuilds from the current catalog, so dropping a large
table or database can reduce a surviving BE from millions of tablets to one
between rounds. This passes the old count unchanged to `newKeySet`; [JDK 17
sizes the lazily created backing table for that
count](https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/util/concurrent/ConcurrentHashMap.html)
on the first `add`. The rebuild then holds old-large plus new-large instead of
old-large plus new-small. If allocation fails before the temporary maps are
published, `Daemon.run()` retries while the old map and same stale hint remain,
so this can become a repeated OOME. The added test already rebuilds one tablet
from previous sizes 3/5 but overrides the allocator, masking this behavior.
Please bound or validate stale hints (or pre-count the current per-BE
cardinality) and cover a sharp-shrink case.
--
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]
