feat: Add microbenchmark for string functions #26
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
viirya
approved these changes
Dec 30, 2025
viirya
reviewed
Dec 30, 2025
| def format_results_markdown(results: list[BenchmarkResult]) -> str: | ||
| """Format benchmark results as a markdown table.""" | ||
| lines = [ | ||
| "# String Function Microbenchmarks: DataFusion vs DuckDB", |
Member
There was a problem hiding this comment.
Can we get versions from DataFusion and DuckDB automatically and put into this title?
Member
Author
There was a problem hiding this comment.
Thanks, I updated this. Header now shows:
| Function | DataFusion 50.0.0 (ms) | DuckDB 1.4.3 (ms) | Speedup | Faster |
This was referenced Dec 30, 2025
paleolimbot
reviewed
Dec 31, 2025
Comment on lines
113
to
124
| def setup_datafusion(parquet_path: str) -> datafusion.SessionContext: | ||
| """Create and configure DataFusion context.""" | ||
| ctx = datafusion.SessionContext() | ||
| ctx.register_parquet('test_data', parquet_path) | ||
| return ctx | ||
|
|
||
|
|
||
| def setup_duckdb(parquet_path: str) -> duckdb.DuckDBPyConnection: | ||
| """Create and configure DuckDB connection.""" | ||
| conn = duckdb.connect(':memory:') | ||
| conn.execute(f"CREATE VIEW test_data AS SELECT * FROM read_parquet('{parquet_path}')") | ||
| return conn |
Member
There was a problem hiding this comment.
In SedonaDB we found that there was wildly differing concurrency that resulted from the default settings for DataFusion and DuckDB for our micro-ish benchmarks. For these types of benchmarks we set DataFusion to use one partition and DuckDB to specifically use a single thread (we don't do this for more macro-scale benchmarks where we really do want to know what happens when a user sits down a types something against all the defaults):
I might also suggest trying the config for DuckDB to return StringViews to see if there's any Arrow conversion overhead getting in the way (I think the config is SET produce_arrow_string_view = true;).
Another thing to try is having both DuckDB and DataFusion operate on Arrow data from memory instead of Parquet (to make sure we're not just measuring the speed of the Parquet read). For DuckDB that's SELECT ... FROM the_name_of_a_python_variable_that_is_a_pyarrow_table.
microbenchmarks/microbenchmarks.py
Outdated
| strings.append(s) | ||
|
|
||
| table = pa.table({ | ||
| 'str_col': pa.array(strings, type=pa.string()) |
Member
There was a problem hiding this comment.
It may be worth trying both string and string_view. In theory string_view -> DuckDB has less overhead because the string view is closer to its internal representation. It also might be that DataFusion performs differently when one or the other is used as input.
github-merge-queue bot
pushed a commit
to apache/datafusion
that referenced
this pull request
Dec 31, 2025
## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes #123` indicates that this PR will close issue #123. --> - Part of #19569 ## Rationale for this change <!-- Why are you proposing this change? If this is already explained clearly in the issue then this section is not needed. Explaining clearly why changes are proposed helps reviewers understand your changes and offer better suggestions for fixes. --> I ran microbenchmarks comparing DataFusion with DuckDB for string functions (see apache/datafusion-benchmarks#26) and noticed that DF was very slow for `md5`. This PR improves performance: | Benchmark | Before | After | Speedup | |----------------------------|--------|--------|-------------| | md5_array (1024 strings) | 206 µs | 100 µs | 2.1x faster | | md5_scalar (single string) | 337 ns | 221 ns | 1.5x faster | ## What changes are included in this PR? <!-- There is no need to duplicate the description in the issue here but it is sometimes worth providing a summary of the individual changes in this PR. --> Avoid using `write!` with a format string and use a more efficient approach ## Are these changes tested? <!-- We typically require tests for all PRs in order to: 1. Prevent the code from being accidentally broken by subsequent changes 2. Serve as another way to document the expected behavior of the code If tests are not included in your PR, please explain why (for example, are they covered by existing tests)? --> ## Are there any user-facing changes? <!-- If there are user-facing changes then we may require documentation to be updated before approving the PR. --> <!-- If there are any breaking changes to public APIs, please add the `api change` label. -->
Member
Author
|
Thanks for the great feedback @paleolimbot! I have pushed commits to address those points |
github-merge-queue bot
pushed a commit
to apache/datafusion
that referenced
this pull request
Jan 6, 2026
## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes #123` indicates that this PR will close issue #123. --> - Closes #. ## Rationale for this change <!-- Why are you proposing this change? If this is already explained clearly in the issue then this section is not needed. Explaining clearly why changes are proposed helps reviewers understand your changes and offer better suggestions for fixes. --> I ran microbenchmarks comparing DataFusion with DuckDB for string functions (see apache/datafusion-benchmarks#26) and noticed that DF was very slow for `split_part`. This PR fixes some obvious performance issues. Speedups are: | Benchmark | Before | After | Speedup | |-----------------------------------|--------|-------|--------------| | single_char_delim/pos_first | 1.27ms | 140µs | 9.1x faster | | single_char_delim/pos_middle | 1.39ms | 396µs | 3.5x faster | | single_char_delim/pos_last | 1.47ms | 738µs | 2.0x faster | | single_char_delim/pos_negative | 1.35ms | 148µs | 9.1x faster | | multi_char_delim/pos_first | 1.22ms | 174µs | 7.0x faster | | multi_char_delim/pos_middle | 1.22ms | 407µs | 3.0x faster | | string_view_single_char/pos_first | 1.42ms | 139µs | 10.2x faster | | many_parts_20/pos_second | 2.48ms | 201µs | 12.3x faster | | long_strings_50_parts/pos_first | 8.18ms | 178µs | 46x faster | ## What changes are included in this PR? <!-- There is no need to duplicate the description in the issue here but it is sometimes worth providing a summary of the individual changes in this PR. --> ## Are these changes tested? <!-- We typically require tests for all PRs in order to: 1. Prevent the code from being accidentally broken by subsequent changes 2. Serve as another way to document the expected behavior of the code If tests are not included in your PR, please explain why (for example, are they covered by existing tests)? --> ## Are there any user-facing changes? <!-- If there are user-facing changes then we may require documentation to be updated before approving the PR. --> <!-- If there are any breaking changes to public APIs, please add the `api change` label. --> --------- Co-authored-by: Martin Grigorov <[email protected]> Co-authored-by: Andrew Lamb <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
This PR adds microbenchmarks for scanning a Parquet file and evaluating a single string expression per row.
The benchmark runs against DuckDB and DataFusion and compares the results.
Assuming that this PR can be merged, I will follow up by covering other expressions.