Snapshot Testing Tool for ASV Benchmarks

A tool for capturing and comparing function outputs from ASV benchmarks to verify correctness after optimizations.

Overview

This tool integrates with ASV (airspeed-velocity) benchmarks to:

• Capture function return values before optimization
• Compare outputs after optimization to ensure correctness
• Use execution tracing to capture the deepest function call's return value
• Support both function-level and class-based parameterized benchmarks

Installation

Using uv (recommended)

# Install with uv
uv pip install -e .

# Or install directly from repository
uv pip install git+https://github.com/formula-code/snapshot-tester.git

Using pip

# Install in development mode
pip install -e .

# Or install directly
pip install git+https://github.com/formula-code/snapshot-tester.git

Usage

Command Line Interface

After installation, use the snapshot-tool command:

List Benchmarks

snapshot-tool list <benchmark_dir> [--filter PATTERN]

Example:

snapshot-tool list examples/astropy-benchmarks/benchmarks --filter time_angle

Capture Snapshots

snapshot-tool capture <benchmark_dir> [--filter PATTERN] [--snapshot-dir DIR]

This captures baseline snapshots of all benchmark outputs.

Verify Against Snapshots

snapshot-tool verify <benchmark_dir> [--filter PATTERN] [--snapshot-dir DIR] [--tolerance RTOL ATOL]

This runs benchmarks and compares outputs against stored snapshots.

Configuration

snapshot-tool config --init    # Create default config
snapshot-tool config --show    # Show current config

Running with uv

You can also run the tool without installation using uv run:

uv run snapshot-tool list <benchmark_dir>
uv run snapshot-tool capture <benchmark_dir>
uv run snapshot-tool verify <benchmark_dir>

Programmatic Usage

from snapshot_tool import BenchmarkDiscovery, BenchmarkRunner, SnapshotManager, Comparator

# Discover benchmarks
discovery = BenchmarkDiscovery("benchmarks/")
benchmarks = discovery.discover_all()

# Run benchmarks with tracing
runner = BenchmarkRunner("benchmarks/")
result = runner.run_benchmark(benchmark)

# Store snapshots
storage = SnapshotManager(".snapshots/")
storage.store_snapshot(
    benchmark_name="my_benchmark",
    module_path="my_module", 
    parameters=(),
    param_names=None,
    return_value=result.return_value
)

# Compare outputs
comparator = Comparator()
comparison = comparator.compare(actual_value, expected_value)

Architecture

Core Components

1. BenchmarkDiscovery: Parses ASV benchmark files to find benchmark classes and functions
2. ExecutionTracer: Uses sys.settrace() to capture the deepest function call's return value
3. BenchmarkRunner: Executes benchmarks with tracing enabled
4. SnapshotManager: Stores and retrieves snapshots using pickle files
5. Comparator: Compares outputs using numpy.allclose for numerical data
6. CLI: Command-line interface for easy usage

Key Features

• Hierarchical Discovery: Supports nested benchmark directories
• Parameter Handling: Automatically generates parameter combinations for parameterized benchmarks
• Global Variables: Handles benchmarks that use global variables
• Setup Methods: Supports class-based benchmarks with setup methods
• Flexible Comparison: Uses numpy.allclose for numerical data with configurable tolerances
• Metadata Tracking: Stores git commit, timestamp, and other metadata with snapshots

Configuration

Create a snapshot_config.json file to customize behavior:

{
  "benchmark_dir": "benchmarks/",
  "snapshot_dir": ".snapshots/",
  "tolerance": {
    "rtol": 1e-5,
    "atol": 1e-8,
    "equal_nan": false
  },
  "exclude_benchmarks": [],
  "trace_depth_limit": 100,
  "verbose": false,
  "quiet": false
}

Example Workflow

1. Initial Setup: Capture baseline snapshots

   python snapshot_test.py capture benchmarks/

2. Make Optimizations: Modify your code to improve performance

3. Verify Correctness: Compare outputs against snapshots

   python snapshot_test.py verify benchmarks/

4. Review Results: The tool will report which benchmarks passed/failed

Limitations

• Requires the benchmarked project to be importable
• Some benchmarks may not return meaningful values to capture
• Execution tracing adds overhead to benchmark runs
• Pickle format may not be compatible across Python versions

Testing

Run the test script to verify the tool works:

python test_snapshot_tool.py

This will test discovery, storage, comparison, and basic benchmark execution.

Contributing

The tool is designed to be extensible. Key areas for enhancement:

• Support for more data types in comparison
• Integration with CI/CD systems
• Web-based reporting interface
• Support for distributed snapshot storage