This directory contains microbenchmarks for PySpark using ASV (Airspeed Velocity).

Install ASV:

pip install asv

For running benchmarks with isolated environments (without --python=same), you need an environment manager. The default configuration uses virtualenv, but ASV also supports conda, mamba, uv, and some others. See the official docs for details.

Run benchmarks using your current Python environment (fastest for development):

cd python/benchmarks
asv run --python=same --quick

You can also specify the test class to run:

cd python/benchmarks
asv run --python=same --quick -b 'bench_arrow.LongArrowToPandasBenchmark'

Run benchmarks in an isolated virtualenv (builds pyspark from source):

cd python/benchmarks
asv run master^!          # Run on latest master commit
asv run v3.5.0^!          # Run on a specific tag
asv run abc123^!          # Run on a specific commit

Compare current branch against upstream/main with 10% threshold:

asv continuous -f 1.1 upstream/main HEAD

asv check          # Validate benchmark syntax

Benchmarks are Python classes with methods prefixed by:

• time_* - Measure execution time
• peakmem_* - Measure peak memory usage
• mem_* - Measure memory usage of returned object

Example:

class MyBenchmark:
    params = [[1000, 10000], ["option1", "option2"]]
    param_names = ["n_rows", "option"]

    def setup(self, n_rows, option):
        # Called before each benchmark method
        self.data = create_test_data(n_rows, option)

    def time_my_operation(self, n_rows, option):
        # Benchmark timing
        process(self.data)

    def peakmem_my_operation(self, n_rows, option):
        # Benchmark peak memory
        process(self.data)

See ASV documentation for more details.