aglinxinyuan opened a new issue, #7246:
URL: https://github.com/apache/texera/issues/7246
### What happened?
A Loop End with two inbound links into its single input port has **no
runtime guard**. Unlike Loop Start, it does not fail at `StartWorkflow` — it
runs and silently double-consumes.
Every reader on a materialized input port replays that port's states
independently, so a second link delivers the loop state twice per iteration:
the user's `update` runs twice and the back-edge is written twice, producing a
wrong iteration count with no error.
### Root cause
The only runtime check lives in
`WorkflowExecutionManager.loopStartPortUris`, and it sits under
`filter(_.isLoopStart)`:
```scala
region.getOperators.filter(_.isLoopStart).map { op =>
require(cfg.storagePairs.size == 1, s"Loop Start input port $gpid expected
exactly one reader URI, ...")
```
So the `require` only ever sees Loop Start operators. There is no
`isLoopEnd` marker on `PhysicalOp` and no equivalent check for the Loop End
side.
#7154 declares `disallowMultiLinks` on the shared `LoopOpDesc` input port,
which stops the GUI drawing a second link into **either** loop operator — but
that is a GUI guard. Plans built programmatically (the case in #6966) bypass it
entirely, and for Loop End nothing downstream catches the mistake.
### Expected behavior
| | Loop Start | Loop End |
|---|---|---|
| GUI (after #7154) | 2nd link refused | 2nd link refused |
| Programmatic plan, 2 links | fails at `StartWorkflow` | **runs and
double-consumes** ← the gap |
A matching runtime check for Loop End would close it — e.g. an `isLoopEnd`
flag on `PhysicalOp` (mirroring `isLoopStart`) plus a `require` on its input
port's reader count, or a general single-reader assertion applied to any port
whose descriptor declares `disallowMultiLinks`. The latter is more general and
would cover future single-input operators too.
Raised in review of #7154.
--
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]
