Andy Grove
33b86fe02e
## Which issue does this PR close?
N/A - performance optimization
## Rationale for this change
In the SMJ tight loop (`join_partial`), `num_unfrozen_pairs()` was
called **twice per iteration**: once in the loop guard and once inside
`append_output_pair`. This method iterates all chunks in
`output_indices` and sums their lengths — O(num_chunks). Over a full
batch of `batch_size` iterations, this makes the inner loop O(batch_size
* num_chunks) instead of O(batch_size).
## What changes are included in this PR?
Add a `num_output_rows` field to `StreamedBatch` that is incremented on
each append and reset on freeze, replacing the O(n) summation with an
O(1) field read.
- Added `num_output_rows: usize` field to `StreamedBatch`, initialized
to `0`
- Increment `num_output_rows` in `append_output_pair()` after each
append
- `num_output_rows()` now returns the cached field directly
- Reset to `0` in `freeze_streamed()` when `output_indices` is cleared
- Removed the `num_unfrozen_pairs` parameter from `append_output_pair()`
since it can now read `self.num_output_rows` directly
## Are these changes tested?
Yes — all 48 existing `sort_merge_join` tests pass. This is a pure
refactor of an internal counter with no behavioral change.
## Performance
Very minor improvement.
### Before
```
sort_merge_join/inner_1to1/100000
time: [3.8146 ms 3.8229 ms 3.8314 ms]
sort_merge_join/inner_1to10/100000
time: [16.094 ms 16.125 ms 16.161 ms]
Found 7 outliers among 100 measurements (7.00%)
6 (6.00%) high mild
1 (1.00%) high severe
sort_merge_join/left_1to1_unmatched/100000
time: [3.7823 ms 3.7861 ms 3.7902 ms]
Found 4 outliers among 100 measurements (4.00%)
4 (4.00%) high mild
sort_merge_join/left_semi_1to10/100000
time: [3.0523 ms 3.0755 ms 3.1023 ms]
Found 14 outliers among 100 measurements (14.00%)
3 (3.00%) high mild
11 (11.00%) high severe
sort_merge_join/left_anti_partial/100000
time: [3.3458 ms 3.3498 ms 3.3542 ms]
Found 12 outliers among 100 measurements (12.00%)
8 (8.00%) high mild
4 (4.00%) high severe
```
### After
```
sort_merge_join/inner_1to1/100000
time: [3.7162 ms 3.7207 ms 3.7254 ms]
change: [−4.2320% −3.9309% −3.6431%] (p = 0.00 < 0.05)
Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
4 (4.00%) high mild
sort_merge_join/inner_1to10/100000
time: [15.556 ms 15.589 ms 15.626 ms]
change: [−5.2786% −4.8329% −4.4351%] (p = 0.00 < 0.05)
Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
1 (1.00%) high mild
3 (3.00%) high severe
sort_merge_join/left_1to1_unmatched/100000
time: [3.7059 ms 3.7101 ms 3.7146 ms]
change: [−4.4526% −4.1565% −3.8660%] (p = 0.00 < 0.05)
Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
2 (2.00%) high mild
sort_merge_join/left_semi_1to10/100000
time: [3.0832 ms 3.0899 ms 3.0981 ms]
change: [−4.0965% −3.4158% −2.7657%] (p = 0.00 < 0.05)
Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
1 (1.00%) high mild
2 (2.00%) high severe
sort_merge_join/left_anti_partial/100000
time: [3.2963 ms 3.3048 ms 3.3153 ms]
change: [−3.9413% −3.5316% −3.0884%] (p = 0.00 < 0.05)
Performance has improved.
Found 8 outliers among 100 measurements (8.00%)
3 (3.00%) high mild
5 (5.00%) high severe
```
## Are there any user-facing changes?
No.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Apache DataFusion
DataFusion is an extensible query engine written in Rust that uses Apache Arrow as its in-memory format.
This crate provides libraries and binaries for developers building fast and feature-rich database and analytic systems, customized for particular workloads. See use cases for examples. The following related subprojects target end users:
- DataFusion Python offers a Python interface for SQL and DataFrame queries.
- DataFusion Comet is an accelerator for Apache Spark based on DataFusion.
"Out of the box," DataFusion offers SQL and DataFrame APIs, excellent performance, built-in support for CSV, Parquet, JSON, and Avro, extensive customization, and a great community.
DataFusion features a full query planner, a columnar, streaming, multi-threaded, vectorized execution engine, and partitioned data sources. You can customize DataFusion at almost all points including additional data sources, query languages, functions, custom operators and more. See the Architecture section for more details.
Here are links to important resources:
- Project Site
- Installation
- Rust Getting Started
- Rust DataFrame API
- Rust API docs
- Rust Examples
- Python DataFrame API
- Architecture
What can you do with this crate?
DataFusion is great for building projects such as domain-specific query engines, new database platforms and data pipelines, query languages and more. It lets you start quickly from a fully working engine, and then customize those features specific to your needs. See the list of known users.
Contributing to DataFusion
Please see the contributor guide and communication pages for more information.
Crate features
This crate has several features which can be specified in your Cargo.toml.
Default features:
nested_expressions: functions for working with nested types such asarray_to_stringcompression: reading files compressed withxz2,bzip2,flate2, andzstdcrypto_expressions: cryptographic functions such asmd5andsha256datetime_expressions: date and time functions such asto_timestampencoding_expressions:encodeanddecodefunctionsparquet: support for reading the Apache Parquet formatsql: support for SQL parsing and planningregex_expressions: regular expression functions, such asregexp_matchunicode_expressions: include Unicode-aware functions such ascharacter_lengthunparser: enables support to reverse LogicalPlans back into SQLrecursive_protection: uses recursive for stack overflow protection.
Optional features:
avro: support for reading the Apache Avro formatbacktrace: include backtrace information in error messagesparquet_encryption: support for using Parquet Modular Encryptionserde: enable arrow-schema'sserdefeature
DataFusion API Evolution and Deprecation Guidelines
Public methods in Apache DataFusion evolve over time: while we try to maintain a stable API, we also improve the API over time. As a result, we typically deprecate methods before removing them, according to the deprecation guidelines.
Dependencies and Cargo.lock
Following the guidance on committing Cargo.lock files, this project commits
its Cargo.lock file.
CI uses the committed Cargo.lock file, and dependencies are updated regularly
using Dependabot PRs.