Piotr Nowojski created FLINK-12912:
--------------------------------------
Summary: Incorrect handling of task.checkpoint.alignment.max-size
when one checkpoint subsumes another one
Key: FLINK-12912
URL: https://issues.apache.org/jira/browse/FLINK-12912
Project: Flink
Issue Type: Bug
Components: Runtime / Network
Affects Versions: 1.8.0, 1.7.2, 1.6.4
Reporter: Piotr Nowojski
Assignee: Piotr Nowojski
{{BarrierBuffer#numQueuedBytes}} which is used to evaluate {{
task.checkpoint.alignment.max-size}} limit, is not correctly handled if one
checkpoint subsumes another one.
The max size limit is checked against a sum of {{numQueuedBytes}} and
{{bufferBlocker.getBytesBlocked()}}. The {{getBytesBlocked}} keeps tracks of
the alignment size of the only most latest checkpoint. The bug is
{{BarrierBuffer#releaseBlocksAndResetBarriers()}} method, where while handling
first subsumed checkpoint in the branch:
{code:java}
if (currentBuffered == null) {
// common case: no more buffered data
currentBuffered =
bufferBlocker.rollOverReusingResources();
if (currentBuffered != null) {
currentBuffered.open();
}
}
{code}
we clear the {{bufferBlocker.getBytesBlocked()}} counter, while we do not
update {{numQueuedBytes}} counter.
For example when first checkpoint approached to 99.9% of max alignment size
when it was subsumed, due to this bug calculated alignment size drops to 0
bytes. For subsequent subsumed checkpoints {{numQueuedBytes}} is correctly
updated.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
