Rui Fan created FLINK-40269:
-------------------------------
Summary: Unaligned checkpoint restore may fail after rescale when
a two-input task reads from the same upstream
Key: FLINK-40269
URL: https://issues.apache.org/jira/browse/FLINK-40269
Project: Flink
Issue Type: Bug
Components: Runtime / Checkpointing
Reporter: Rui Fan
We found a restore failure in Flink 2.3.0 when restoring from an unaligned
checkpoint after rescaling the job.
This issue likely exists since the early unaligned-checkpoint-rescaling support
around Flink 1.13/1.14.
It can happen when one two-input task has both inputs derived from the same
upstream stream:
{code:java}
source
|-- rebalance() --> CoMap input 1
|-- keyBy(...) --> CoMap input 2{code}
During rescale restore, Flink can confuse the two input edges and route channel
state with the wrong mapping. The job then fails during restore with an error
like:
{code:java}
text Cannot select SubtaskConnectionDescriptor{inputSubtaskIndex=1,
outputSubtaskIndex=0};
known channels are [SubtaskConnectionDescriptor{inputSubtaskIndex=0,
outputSubtaskIndex=0}]{code}
Reproducer branch:
https://github.com/1996fanrui/flink/tree/repro/unaligned-channel-remap
Commit `2711b807112e50b4bdabbd1ca181f800e8968a15` adds a minimal ITCase that
reproduces the failure 100% before the runtime fix:
UnalignedCheckpointRescaleSameUpstreamITCase.java
The proposed fix is in commit `9b8560e69818e25ac20e19875aca55bdea208fcc`.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
