Bowen Li created FLINK-39946:
--------------------------------
Summary: Unaligned checkpoint restore can mis-map union input
channel state after reordering UIDed sources
Key: FLINK-39946
URL: https://issues.apache.org/jira/browse/FLINK-39946
Project: Flink
Issue Type: Bug
Components: Runtime / Checkpointing
Affects Versions: 1.19.1
Reporter: Bowen Li
When a Flink job restores from an unaligned checkpoint, input channel state is
restored by positional input gate/channel indexes rather than by a stable edge
identity.
This can break when a job has multiple source operators feeding a union and the
source declarations are reordered between job versions. Even if the source
operators have stable UIDs, the generated {{JobEdge}} / input gate order can
still change because initial source traversal is based on transient stream node
ordering. After restore, in-flight channel state from one logical source may be
applied to another logical input gate.
This is especially problematic for jobs that:
* use unaligned checkpoints or aligned checkpoints that fall back to unaligned,
* restore from checkpoint/savepoint state containing in-flight channel data,
* reorder UIDed union inputs, for example by reordering Kafka source
declarations.
Expected behavior: if union inputs have stable operator identities, restoring
channel state should preserve the logical source-to-input-gate mapping across
source declaration reorders.
Observed behavior: source operator UIDs remain stable, but union input gate
positions can change, so channel state restore is tied to the wrong logical
input.
For this specific bug, it is an unaligned-checkpoint channel-state restore
issue, not a normal aligned-checkpoint restore issue.
Why:
* Aligned checkpoints do not persist in-flight input channel buffers; barriers
align first, so there is no channel state to remap on restore.
* Unaligned checkpoints do persist in-flight channel state, and Flink restores
that by positional {{{}gateIdx/inputChannelIdx{}}}. That is where reordered
union inputs can restore the wrong logical channel state.
One caveat: an “aligned” job can still hit it if it uses aligned checkpoint
timeout and falls back to unaligned. In that case the affected checkpoint
contains channel state, so the same bug applies.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
