[
https://issues.apache.org/jira/browse/FLINK-38286?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ramin Gharib updated FLINK-38286:
---------------------------------
Description:
*Summary*
The MAP function exhibits non-deterministic behavior when duplicate keys are
provided, returning different results across environments and test runs. This
breaks reproducibility and can cause CI failures.
h3. *Description*
h4. Problem
The `MAP` function in Flink Table API/SQL produces inconsistent results when
duplicate keys are provided. For example, `MAP[f0, f0, f0, f1]` where `f0=1`
and `f1=2` should deterministically return `\{1=2}` (last value wins), but
sometimes returns `\{1=1}` instead.
h4. Root Cause
The issue lies in the code generation logic in `ScalarOperatorGens.scala`
(lines ~1510-1530). The current implementation uses:
{code:java}
val keyElements = elements
.grouped(2)
.map { case Seq(key, value) => (key, value) }
.toSeq
.groupBy(_._1)
.map(_._2.last)
.keys
.toSeq{code}
The problem is that `groupBy` returns a `Map`, and when we extract `.keys` and
`.values`, the iteration order is {_}non-deterministic{_}. This breaks the
correspondence between `keyArray[i]` and `valueArray[i]` in the generated code.
h3. Steps to Reproduce
1. Run the `MapFunctionITCase` test with constant folding disabled
2. Execute the specific test case: `map(f0, f0, f0, f1)` where `f0=1, f1=2`
3. Observe that results vary between runs/environments
*Test Code:*
{code:java}
// In MapFunctionITCase.java
resultSpec(
map($("f0"), $("f0"), $("f0"), $("f1")),
"MAP[f0, f1]",
Collections.singletonMap(1, 2), // Expected: {1=2}
DataTypes.MAP(INT().notNull(), INT().notNull()).notNull()
){code}
*Expected Behavior*
- `MAP[1, 1, 1, 2]` should consistently return `\{1=2}` (last value wins)
- Results should be deterministic across all environments
*Actual Behavior*
- Sometimes returns `\{1=2}` ✅
- Sometimes returns `\{1=1}` ❌
- Non-deterministic failures in CI environments
was:
*Summary*
The MAP function exhibits non-deterministic behavior when duplicate keys are
provided, returning different results across environments and test runs. This
breaks reproducibility and can cause CI failures.
h3. *Description*
h4. Problem
The `MAP` function in Flink Table API/SQL produces inconsistent results when
duplicate keys are provided. For example, `MAP[f0, f0, f0, f1]` where `f0=1`
and `f1=2` should deterministically return `\{1=2}` (last value wins), but
sometimes returns `\{1=1}` instead.
h4. Root Cause
The issue lies in the code generation logic in `ScalarOperatorGens.scala`
(lines ~1510-1530). The current implementation uses:
```scala
val keyElements = elements
.grouped(2)
.map \{ case Seq(key, value) => (key, value) }
.toSeq
.groupBy(_._1)
.map(_._2.last)
.keys
.toSeq
```
The problem is that `groupBy` returns a `Map`, and when we extract `.keys` and
`.values`, the iteration order is {_}non-deterministic{_}. This breaks the
correspondence between `keyArray[i]` and `valueArray[i]` in the generated code.
h3. Steps to Reproduce
1. Run the `MapFunctionITCase` test with constant folding disabled
2. Execute the specific test case: `map(f0, f0, f0, f1)` where `f0=1, f1=2`
3. Observe that results vary between runs/environments
*Test Code:*
{code:java}
// In MapFunctionITCase.java
resultSpec(
map($("f0"), $("f0"), $("f0"), $("f1")),
"MAP[f0, f1]",
Collections.singletonMap(1, 2), // Expected: {1=2}
DataTypes.MAP(INT().notNull(), INT().notNull()).notNull()
){code}
*Expected Behavior*
- `MAP[1, 1, 1, 2]` should consistently return `\{1=2}` (last value wins)
- Results should be deterministic across all environments
*Actual Behavior*
- Sometimes returns `\{1=2}` ✅
- Sometimes returns `\{1=1}` ❌
- Non-deterministic failures in CI environments
> MAP function with duplicate keys produces non-deterministic results
> -------------------------------------------------------------------
>
> Key: FLINK-38286
> URL: https://issues.apache.org/jira/browse/FLINK-38286
> Project: Flink
> Issue Type: Bug
> Components: Table SQL / API
> Reporter: Ramin Gharib
> Priority: Major
> Labels: pull-request-available
>
> *Summary*
> The MAP function exhibits non-deterministic behavior when duplicate keys are
> provided, returning different results across environments and test runs. This
> breaks reproducibility and can cause CI failures.
> h3. *Description*
> h4. Problem
> The `MAP` function in Flink Table API/SQL produces inconsistent results when
> duplicate keys are provided. For example, `MAP[f0, f0, f0, f1]` where `f0=1`
> and `f1=2` should deterministically return `\{1=2}` (last value wins), but
> sometimes returns `\{1=1}` instead.
> h4. Root Cause
> The issue lies in the code generation logic in `ScalarOperatorGens.scala`
> (lines ~1510-1530). The current implementation uses:
>
> {code:java}
> val keyElements = elements
> .grouped(2)
> .map { case Seq(key, value) => (key, value) }
> .toSeq
> .groupBy(_._1)
> .map(_._2.last)
> .keys
> .toSeq{code}
> The problem is that `groupBy` returns a `Map`, and when we extract `.keys`
> and `.values`, the iteration order is {_}non-deterministic{_}. This breaks
> the correspondence between `keyArray[i]` and `valueArray[i]` in the generated
> code.
> h3. Steps to Reproduce
> 1. Run the `MapFunctionITCase` test with constant folding disabled
> 2. Execute the specific test case: `map(f0, f0, f0, f1)` where `f0=1, f1=2`
> 3. Observe that results vary between runs/environments
> *Test Code:*
>
> {code:java}
> // In MapFunctionITCase.java
> resultSpec(
> map($("f0"), $("f0"), $("f0"), $("f1")),
> "MAP[f0, f1]",
> Collections.singletonMap(1, 2), // Expected: {1=2}
> DataTypes.MAP(INT().notNull(), INT().notNull()).notNull()
> ){code}
>
>
>
> *Expected Behavior*
> - `MAP[1, 1, 1, 2]` should consistently return `\{1=2}` (last value wins)
> - Results should be deterministic across all environments
> *Actual Behavior*
> - Sometimes returns `\{1=2}` ✅
> - Sometimes returns `\{1=1}` ❌
> - Non-deterministic failures in CI environments
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
