Introducing StateSchemaV3 for the TransformWithState operator #8
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ericm-db
force-pushed
the
tws-state-schema-changes
branch
from
d6cd20f to
37526de
Compare
ericm-db
force-pushed
the
tws-state-schema-changes
branch
from
ccee64e to
67acea8
Compare
ericm-db
pushed a commit
that referenced
this pull request
Feb 5, 2025
This is a trivial change to replace the loop index from `int` to `long`. Surprisingly, microbenchmark shows more than double performance uplift.
Analysis
--------
The hot loop of `arrayEquals` method is simplifed as below. Loop index `i` is defined as `int`, it's compared with `length`, which is a `long`, to determine if the loop should end.
```
public static boolean arrayEquals(
Object leftBase, long leftOffset, Object rightBase, long rightOffset, final long length) {
......
int i = 0;
while (i <= length - 8) {
if (Platform.getLong(leftBase, leftOffset + i) !=
Platform.getLong(rightBase, rightOffset + i)) {
return false;
}
i += 8;
}
......
}
```
Strictly speaking, there's a code bug here. If `length` is greater than 2^31 + 8, this loop will never end because `i` as a 32 bit integer is at most 2^31 - 1. But compiler must consider this behaviour as intentional and generate code strictly match the logic. It prevents compiler from generating optimal code.
Defining loop index `i` as `long` corrects this issue. Besides more accurate code logic, JIT is able to optimize this code much more aggressively. From microbenchmark, this trivial change improves performance significantly on both Arm and x86 platforms.
Benchmark
---------
Source code:
https://gist.github.com/cyb70289/258e261f388e22f47e4d961431786d1a
Result on Arm Neoverse N2:
```
Benchmark Mode Cnt Score Error Units
ArrayEqualsBenchmark.arrayEqualsInt avgt 10 674.313 ± 0.213 ns/op
ArrayEqualsBenchmark.arrayEqualsLong avgt 10 313.563 ± 2.338 ns/op
```
Result on Intel Cascake Lake:
```
Benchmark Mode Cnt Score Error Units
ArrayEqualsBenchmark.arrayEqualsInt avgt 10 1130.695 ± 0.168 ns/op
ArrayEqualsBenchmark.arrayEqualsLong avgt 10 461.979 ± 0.097 ns/op
```
Deep dive
---------
Dive deep to the machine code level, we can see why the big gap. Listed below are arm64 assembly generated by Openjdk-17 C2 compiler.
For `int i`, the machine code is similar to source code, no deep optimization. Safepoint polling is expensive in this short loop.
```
// jit c2 machine code snippet
0x0000ffff81ba8904: mov w15, wzr // int i = 0
0x0000ffff81ba8908: nop
0x0000ffff81ba890c: nop
loop:
0x0000ffff81ba8910: ldr x10, [x13, w15, sxtw] // Platform.getLong(leftBase, leftOffset + i)
0x0000ffff81ba8914: ldr x14, [x12, w15, sxtw] // Platform.getLong(rightBase, rightOffset + i)
0x0000ffff81ba8918: cmp x10, x14
0x0000ffff81ba891c: b.ne 0x0000ffff81ba899c // return false if not equal
0x0000ffff81ba8920: ldr x14, [x28, apache#848] // x14 -> safepoint
0x0000ffff81ba8924: add w15, w15, #0x8 // i += 8
0x0000ffff81ba8928: ldr wzr, [x14] // safepoint polling
0x0000ffff81ba892c: sxtw x10, w15 // extend i to long
0x0000ffff81ba8930: cmp x10, x11
0x0000ffff81ba8934: b.le 0x0000ffff81ba8910 // if (i <= length - 8) goto loop
```
For `long i`, JIT is able to do much more aggressive optimization. E.g, below code snippet unrolls the loop by four.
```
// jit c2 machine code snippet
unrolled_loop:
0x0000ffff91de6fe0: sxtw x10, w7
0x0000ffff91de6fe4: add x23, x22, x10
0x0000ffff91de6fe8: add x24, x21, x10
0x0000ffff91de6fec: ldr x13, [x23] // unroll-1
0x0000ffff91de6ff0: ldr x14, [x24]
0x0000ffff91de6ff4: cmp x13, x14
0x0000ffff91de6ff8: b.ne 0x0000ffff91de70a8
0x0000ffff91de6ffc: ldr x13, [x23, #8] // unroll-2
0x0000ffff91de7000: ldr x14, [x24, #8]
0x0000ffff91de7004: cmp x13, x14
0x0000ffff91de7008: b.ne 0x0000ffff91de70b4
0x0000ffff91de700c: ldr x13, [x23, #16] // unroll-3
0x0000ffff91de7010: ldr x14, [x24, #16]
0x0000ffff91de7014: cmp x13, x14
0x0000ffff91de7018: b.ne 0x0000ffff91de70a4
0x0000ffff91de701c: ldr x13, [x23, #24] // unroll-4
0x0000ffff91de7020: ldr x14, [x24, #24]
0x0000ffff91de7024: cmp x13, x14
0x0000ffff91de7028: b.ne 0x0000ffff91de70b0
0x0000ffff91de702c: add w7, w7, #0x20
0x0000ffff91de7030: cmp w7, w11
0x0000ffff91de7034: b.lt 0x0000ffff91de6fe0
```
### What changes were proposed in this pull request?
A trivial change to replace loop index `i` of method `arrayEquals` from `int` to `long`.
### Why are the changes needed?
To improve performance and fix a possible bug.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Existing unit tests.
### Was this patch authored or co-authored using generative AI tooling?
No.
Closes apache#49568 from cyb70289/arrayEquals.
Authored-by: Yibo Cai <[email protected]>
Signed-off-by: Sean Owen <[email protected]>
ericm-db
pushed a commit
that referenced
this pull request
Aug 26, 2025
…onicalized expressions
### What changes were proposed in this pull request?
Make PullOutNonDeterministic use canonicalized expressions to dedup group and aggregate expressions. This affects pyspark udfs in particular. Example:
```
from pyspark.sql.functions import col, avg, udf
pythonUDF = udf(lambda x: x).asNondeterministic()
spark.range(10)\
.selectExpr("id", "id % 3 as value")\
.groupBy(pythonUDF(col("value")))\
.agg(avg("id"), pythonUDF(col("value")))\
.explain(extended=True)
```
Currently results in a plan like this:
```
Aggregate [_nondeterministic#15](#15), [_nondeterministic#15 AS dummyNondeterministicUDF(value)#12, avg(id#0L) AS avg(id)#13, dummyNondeterministicUDF(value#6L)#8 AS dummyNondeterministicUDF(value)#14](#15%20AS%20dummyNondeterministicUDF(value)#12,%20avg(id#0L)%20AS%20avg(id)#13,%20dummyNondeterministicUDF(value#6L)#8%20AS%20dummyNondeterministicUDF(value)#14)
+- Project [id#0L, value#6L, dummyNondeterministicUDF(value#6L)#7 AS _nondeterministic#15](#0L,%20value#6L,%20dummyNondeterministicUDF(value#6L)#7%20AS%20_nondeterministic#15)
+- Project [id#0L, (id#0L % cast(3 as bigint)) AS value#6L](#0L,%20(id#0L%20%%20cast(3%20as%20bigint))%20AS%20value#6L)
+- Range (0, 10, step=1, splits=Some(2))
```
and then it throws:
```
[[MISSING_AGGREGATION] The non-aggregating expression "value" is based on columns which are not participating in the GROUP BY clause. Add the columns or the expression to the GROUP BY, aggregate the expression, or use "any_value(value)" if you do not care which of the values within a group is returned. SQLSTATE: 42803
```
- how canonicalized fixes this:
- nondeterministic PythonUDF expressions always have distinct resultIds per udf
- The fix is to canonicalize the expressions when matching. Canonicalized means that we're setting the resultIds to -1, allowing us to dedup the PythonUDF expressions.
- for deterministic UDFs, this rule does not apply and "Post Analysis" batch extracts and deduplicates the expressions, as expected
### Why are the changes needed?
- the output of the query with the fix applied still makes sense - the nondeterministic UDF is invoked only once, in the project.
### Does this PR introduce _any_ user-facing change?
Yes, it's additive, it enables queries to run that previously threw errors.
### How was this patch tested?
- added unit test
### Was this patch authored or co-authored using generative AI tooling?
No
Closes apache#52061 from benrobby/adhoc-fix-pull-out-nondeterministic.
Authored-by: Ben Hurdelhey <[email protected]>
Signed-off-by: Wenchen Fan <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
Why are the changes needed?
Does this PR introduce any user-facing change?
How was this patch tested?
Was this patch authored or co-authored using generative AI tooling?