perf: improve performance of levenshtein by reusing cache buffer (#19532)

## 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.
-->

## 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.
-->

Add levenshtein_with_buffer() function that accepts a reusable cache
buffer to avoid allocating a new Vec<usize> for each distance
calculation.

Changes:
- Added levenshtein_with_buffer() in datafusion-common that takes a
mutable
      Vec<usize> buffer parameter
- Updated levenshtein function to use the optimized version with a
reusable
      buffer across all rows
  - Added benchmark to measure performance improvements

Optimization:
  - Before: Allocated new Vec<usize> cache for every row
  - After: Single Vec<usize> buffer reused across all rows

Benchmark Results:
  - size=1024, str_len=8:  60.6 µs → 45.9 µs (24% faster)
  - size=1024, str_len=32: 615.5 µs → 598.5 µs (3% faster)
  - size=4096, str_len=8:  234.7 µs → 180.5 µs (23% faster)
  - size=4096, str_len=32: 2.46 ms → 2.38 ms (3% faster)

The optimization shows significant improvements for shorter strings
(23-24%) where allocation overhead is more prominent relative to
algorithm cost. For longer strings, the O(m×n) algorithm complexity
dominates, but still shows measurable 3% improvement from eliminating
per-row allocations.

## 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.
-->

Author: viirya

datafusion/common/src/utils/mod.rs

@@ -693,10 +693,14 @@ pub mod datafusion_strsim {
     }
 
     /// Calculates the minimum number of insertions, deletions, and substitutions
-    /// required to change one sequence into the other.
-    fn generic_levenshtein<'a, 'b, Iter1, Iter2, Elem1, Elem2>(
+    /// required to change one sequence into the other, using a reusable cache buffer.
+    ///
+    /// This is the generic implementation that works with any iterator types.
+    /// The `cache` buffer will be resized as needed and reused across calls.
+    fn generic_levenshtein_with_buffer<'a, 'b, Iter1, Iter2, Elem1, Elem2>(
         a: &'a Iter1,
         b: &'b Iter2,
+        cache: &mut Vec<usize>,
     ) -> usize
     where
         &'a Iter1: IntoIterator<Item = Elem1>,
@@ -709,7 +713,9 @@ pub mod datafusion_strsim {
             return b_len;
         }
 
-        let mut cache: Vec<usize> = (1..b_len + 1).collect();
+        // Resize cache to fit b_len elements
+        cache.clear();
+        cache.extend(1..=b_len);
 
         let mut result = 0;
 
@@ -729,6 +735,21 @@ pub mod datafusion_strsim {
         result
     }
 
+    /// Calculates the minimum number of insertions, deletions, and substitutions
+    /// required to change one sequence into the other.
+    fn generic_levenshtein<'a, 'b, Iter1, Iter2, Elem1, Elem2>(
+        a: &'a Iter1,
+        b: &'b Iter2,
+    ) -> usize
+    where
+        &'a Iter1: IntoIterator<Item = Elem1>,
+        &'b Iter2: IntoIterator<Item = Elem2>,
+        Elem1: PartialEq<Elem2>,
+    {
+        let mut cache = Vec::new();
+        generic_levenshtein_with_buffer(a, b, &mut cache)
+    }
+
     /// Calculates the minimum number of insertions, deletions, and substitutions
     /// required to change one string into the other.
     ///
@@ -741,6 +762,15 @@ pub mod datafusion_strsim {
         generic_levenshtein(&StringWrapper(a), &StringWrapper(b))
     }
 
+    /// Calculates the Levenshtein distance using a reusable cache buffer.
+    /// This avoids allocating a new Vec for each call, improving performance
+    /// when computing many distances.
+    ///
+    /// The `cache` buffer will be resized as needed and reused across calls.
+    pub fn levenshtein_with_buffer(a: &str, b: &str, cache: &mut Vec<usize>) -> usize {
+        generic_levenshtein_with_buffer(&StringWrapper(a), &StringWrapper(b), cache)
+    }
+
     /// Calculates the normalized Levenshtein distance between two strings.
     /// The normalized distance is a value between 0.0 and 1.0, where 1.0 indicates
     /// that the strings are identical and 0.0 indicates no similarity.

datafusion/functions/Cargo.toml

@@ -269,3 +269,8 @@ required-features = ["string_expressions"]
 harness = false
 name = "ends_with"
 required-features = ["string_expressions"]
+
+[[bench]]
+harness = false
+name = "levenshtein"
+required-features = ["unicode_expressions"]

datafusion/functions/benches/levenshtein.rs

@@ -0,0 +1,87 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+extern crate criterion;
+
+use arrow::array::OffsetSizeTrait;
+use arrow::datatypes::{DataType, Field};
+use arrow::util::bench_util::create_string_array_with_len;
+use criterion::{Criterion, SamplingMode, criterion_group, criterion_main};
+use datafusion_common::DataFusionError;
+use datafusion_common::config::ConfigOptions;
+use datafusion_expr::{ColumnarValue, ScalarFunctionArgs};
+use datafusion_functions::string;
+use std::hint::black_box;
+use std::sync::Arc;
+use std::time::Duration;
+
+fn create_args<O: OffsetSizeTrait>(size: usize, str_len: usize) -> Vec<ColumnarValue> {
+    let string1_array = Arc::new(create_string_array_with_len::<O>(size, 0.1, str_len));
+    let string2_array = Arc::new(create_string_array_with_len::<O>(size, 0.1, str_len));
+
+    vec![
+        ColumnarValue::Array(string1_array),
+        ColumnarValue::Array(string2_array),
+    ]
+}
+
+fn invoke_levenshtein_with_args(
+    args: Vec<ColumnarValue>,
+    number_rows: usize,
+) -> Result<ColumnarValue, DataFusionError> {
+    let arg_fields = args
+        .iter()
+        .enumerate()
+        .map(|(idx, arg)| Field::new(format!("arg_{idx}"), arg.data_type(), true).into())
+        .collect::<Vec<_>>();
+    let config_options = Arc::new(ConfigOptions::default());
+
+    string::levenshtein().invoke_with_args(ScalarFunctionArgs {
+        args,
+        arg_fields,
+        number_rows,
+        return_field: Field::new("f", DataType::Int32, true).into(),
+        config_options: Arc::clone(&config_options),
+    })
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    for size in [1024, 4096] {
+        let mut group = c.benchmark_group(format!("levenshtein size={size}"));
+        group.sampling_mode(SamplingMode::Flat);
+        group.sample_size(10);
+        group.measurement_time(Duration::from_secs(10));
+
+        for str_len in [8, 32] {
+            let args = create_args::<i32>(size, str_len);
+            group.bench_function(
+                format!("levenshtein_string [size={size}, str_len={str_len}]"),
+                |b| {
+                    b.iter(|| {
+                        let args_cloned = args.clone();
+                        black_box(invoke_levenshtein_with_args(args_cloned, size))
+                    })
+                },
+            );
+        }
+
+        group.finish();
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);

datafusion/functions/src/string/levenshtein.rs

@@ -151,12 +151,18 @@ fn levenshtein<T: OffsetSizeTrait>(args: &[ArrayRef]) -> Result<ArrayRef> {
             DataType::Utf8View => {
                 let str1_array = as_string_view_array(&str1)?;
                 let str2_array = as_string_view_array(&str2)?;
+
+                // Reusable buffer to avoid allocating for each row
+                let mut cache = Vec::new();
+
                 let result = str1_array
                     .iter()
                     .zip(str2_array.iter())
                     .map(|(string1, string2)| match (string1, string2) {
                         (Some(string1), Some(string2)) => {
-                            Some(datafusion_strsim::levenshtein(string1, string2) as i32)
+                            Some(datafusion_strsim::levenshtein_with_buffer(
+                                string1, string2, &mut cache,
+                            ) as i32)
                         }
                         _ => None,
                     })
@@ -166,12 +172,18 @@ fn levenshtein<T: OffsetSizeTrait>(args: &[ArrayRef]) -> Result<ArrayRef> {
             DataType::Utf8 => {
                 let str1_array = as_generic_string_array::<T>(&str1)?;
                 let str2_array = as_generic_string_array::<T>(&str2)?;
+
+                // Reusable buffer to avoid allocating for each row
+                let mut cache = Vec::new();
+
                 let result = str1_array
                     .iter()
                     .zip(str2_array.iter())
                     .map(|(string1, string2)| match (string1, string2) {
                         (Some(string1), Some(string2)) => {
-                            Some(datafusion_strsim::levenshtein(string1, string2) as i32)
+                            Some(datafusion_strsim::levenshtein_with_buffer(
+                                string1, string2, &mut cache,
+                            ) as i32)
                         }
                         _ => None,
                     })
@@ -181,12 +193,18 @@ fn levenshtein<T: OffsetSizeTrait>(args: &[ArrayRef]) -> Result<ArrayRef> {
             DataType::LargeUtf8 => {
                 let str1_array = as_generic_string_array::<T>(&str1)?;
                 let str2_array = as_generic_string_array::<T>(&str2)?;
+
+                // Reusable buffer to avoid allocating for each row
+                let mut cache = Vec::new();
+
                 let result = str1_array
                     .iter()
                     .zip(str2_array.iter())
                     .map(|(string1, string2)| match (string1, string2) {
                         (Some(string1), Some(string2)) => {
-                            Some(datafusion_strsim::levenshtein(string1, string2) as i64)
+                            Some(datafusion_strsim::levenshtein_with_buffer(
+                                string1, string2, &mut cache,
+                            ) as i64)
                         }
                         _ => None,
                     })