🐻‍❄️💿 polars-diskcache

A diskcache decorator for Polars DataFrames and LazyFrames that saves results as Parquet files.

What it does

Caches results of Polars operations to disk using Parquet format. When you call a decorated function with the same arguments, it loads the cached Parquet file instead of recomputing.

It aims to provide a low-effort outlet for the DataFrames you store in memory without reinventing the wheel. Feel free to use it as a convenient way to dump DataFrames to disk even when not caching!

Installation

uv pip install polars-diskcache

Requirements

• Python 3.13+
• polars
• diskcache

Features

• Automatic type detection: Caches and restores DataFrames/LazyFrames with their original types
• Parquet storage: Preserves column datatypes and metadata in the Parquet format
• Human-readable cache structure: Symlinked directory structure organised by module, function, and arguments for easy browsing
• Flexible organisation: Choose between nested module/function directories or flat structure
• Filesystem-safe encoding: Automatically handles special characters in module/function names
• Configurable symlink names: Customize output filenames for different use cases
• Argument handling: Supports complex argument types with configurable truncation
• SQLite-backed tracking: Uses diskcache with SQLite to track Parquet blob files
• Type-safe: Full type hints and ty type checker compatibility

Quick Start

import polars as pl
from plcache import cache

# Simple caching - just add the decorator
@cache()
def expensive_computation(n: int) -> pl.DataFrame:
    return pl.DataFrame({
        "values": range(n),
        "squared": [i**2 for i in range(n)]
    })

# First call: executes function and caches result
df1 = expensive_computation(1000)

# Second call: loads from cache (much faster!)
df2 = expensive_computation(1000)

assert df1.equals(df2)  # Identical results

How it works

We hash the function name and arguments (with sorted **kwargs) to create a unique cache key:

call_str = f"{func_name}({bound_args})"
cache_key = hashlib.sha256(call_str.encode()).hexdigest()

The Parquet file is saved to {cache_dir}/blobs/{hash}.parquet and the cache key plus file path are stored in a SQLite database at {cache_dir}/metadata/.

Human-readable symlinks are created at {cache_dir}/functions/module/function/args/ that point back to the blob files, so you can browse your cached results easily. If no args are passed, the directory is given the name no_args/ rather than the empty string.

Cache Structure

plcache creates an organised, browsable cache structure with two layout options:

Nested module path (default: nested=True) organises cache by separate module and function directories:

.polars_cache/
├── metadata/            # diskcache SQLite database
├── blobs/               # actual parquet files (by hash)
│   ├── a1b2c3d4.parquet
│   └── e5f6g7h8.parquet
└── functions/           # human-readable symlinks
    └── __main__/        # module name
        └── expensive_computation/  # function name
            ├── arg0=1000/
            │   └── output.parquet -> ../../../blobs/a1b2c3d4.parquet
            └── arg0=5000/
                └── output.parquet -> ../../../blobs/e5f6g7h8.parquet

Flat module path (nested=False) uses encoded full module.function names in a single directory level:

.polars_cache/
├── metadata/
├── blobs/
│   ├── a1b2c3d4.parquet
│   └── e5f6g7h8.parquet
└── functions/
    └── __main__.expensive_computation/  # encoded module.function
        ├── arg0=1000/
        │   └── output.parquet -> ../../blobs/a1b2c3d4.parquet
        └── arg0=5000/
            └── output.parquet -> ../../blobs/e5f6g7h8.parquet

Configurable Options

Cache Directory Location

By default, plcache creates a hidden .polars_cache directory:

# Default: creates .polars_cache in current working directory
@cache()
def my_function(): ...

# Custom location
@cache(cache_dir="/path/to/my/cache")
def my_function(): ...

# Use system temp directory
@cache(use_tmp=True)
def my_function(): ...

# Non-hidden directory name
@cache(hidden=False)  # creates "polars_cache" instead of ".polars_cache"
def my_function(): ...

Directory Structure Options

# Custom readable directory name (great for organising different cache types)
@cache(symlinks_dir="analytics")  # creates cache/analytics/ instead of cache/functions/
def analytics_function(): ...

@cache(symlinks_dir="data_loading")
def load_data(): ...

# Choose layout style
@cache(nested=True)   # module/function/ (default)
def split_example(): ...

@cache(nested=False)  # module.function/ (flat)
def flat_example(): ...

Argument Handling

# Control argument length in directory names
@cache(trim_arg=20)  # truncate long argument values
def function_with_long_args(very_long_argument_name: str): ...

# Custom symlink filename
@cache(symlink_name="results.parquet")
def custom_output(): ...

@cache(symlink_name="processed_data.parquet")
def data_processor(): ...

All Parameters

@cache(
    cache_dir="/custom/path",           # Cache directory location
    use_tmp=False,                      # Use system temp directory
    hidden=True,                        # Prefix with dot (default)
    size_limit=2**30,                   # Max cache size (1GB default)
    symlinks_dir="functions",           # Readable directory name
    nested=True,                        # Module/function vs flat layout
    trim_arg=50,                        # Max argument length in paths
    symlink_name="output.parquet"       # Custom symlink filename (or a callback)
    cache_key=my_cache_function,        # Custom cache key generation
    entry_dir=my_dir_function,          # Custom directory naming
)
def fully_configured(): ...

Custom Cache Keys

Control what gets cached together by customizing the cache key generation:

def preprocessing_cache_key(func, bound_args):
    """Cache key that ignores debug flags but includes data params."""
    # Remove debug/logging flags from cache consideration
    cache_params = {k: v for k, v in bound_args.items()
                   if k not in ['debug', 'verbose', 'log_level']}
    return f"{func.__name__}({cache_params})"

cache = PolarsCache(
    cache_dir="./preprocessing_cache",
    cache_key=preprocessing_cache_key
)

@cache.cache_polars()
def preprocess_data(raw_data, normalize=True, remove_outliers=False, debug=False):
    # Expensive data preprocessing that shouldn't re-run for debug flag changes
    return expensive_preprocessing(raw_data, normalize, remove_outliers)

# These calls will hit the same cache entry:
clean_data1 = preprocess_data(df, normalize=True, debug=False)
clean_data2 = preprocess_data(df, normalize=True, debug=True)  # Cache hit!

The debug flag doesn't affect the actual processing, so you want the same cached result whether the debug parameter is on or off.

Custom Directory Names

Organize your cache directories with meaningful names:

def experiment_dir_name(func, bound_args):
    """Create experiment-specific directory names."""
    model = bound_args['model_type']
    dataset_size = len(bound_args['data'])
    return f"{model}_experiment_{dataset_size}samples"

cache = PolarsCache(
    cache_dir="./experiments",
    entry_dir=experiment_dir_name,
    symlink_name="model_output.parquet"
)

@cache.cache_polars()
def run_experiment(data, model_type, learning_rate=0.01):
    return train_and_evaluate(data, model_type, learning_rate)

# Creates: experiments/functions/.../xgboost_experiment_1000samples/model_output.parquet
result = run_experiment(large_dataset, "xgboost", 0.001)

Callback Signatures

All callbacks receive normalized arguments where positional and keyword arguments are unified:

# These calls produce identical bound_args:
func(10, name="test")           # bound_args = {'value': 10, 'name': 'test'}
func(value=10, name="test")     # bound_args = {'value': 10, 'name': 'test'}
func(name="test", value=10)     # bound_args = {'value': 10, 'name': 'test'} (sorted)

Cache Key Callback: (func, bound_args) -> str

• Controls which function calls share cache entries
• Return same string to share cache, different strings for separate entries

Entry Directory Callback: (func, bound_args) -> str

• Controls symlink directory names for organization
• Should return filesystem-safe directory names
• Used for browsing, not for cache hits/misses

Symlink Name Callback: (func, bound_args, result, cache_key) -> str

• Controls the symlink filename within each directory
• Receives the actual result and cache key for context

Advanced Usage

Using the PolarsCache Class

For more control, use the PolarsCache class directly:

from plcache import PolarsCache

# Create custom cache instance
my_cache = PolarsCache(
    cache_dir="./analysis_cache",
    symlinks_dir="experiments",
    nested=True,
    symlink_name="experiment_result.parquet"
)

@my_cache.cache_polars()
def run_experiment(params: dict) -> pl.DataFrame:
    # Expensive experiment
    return pl.DataFrame({"result": [1, 2, 3]})

Complex Arguments

plcache handles various argument types intelligently:

@cache(trim_arg=20)
def complex_function(
    data_list: list[int],
    config: dict,
    enabled: bool = True,
    mode: str = "advanced_processing"
) -> pl.DataFrame:
    # Arguments are safely encoded in directory structure
    # Long values are truncated to trim_arg
    return pl.DataFrame({"processed": [len(data_list)]})

# Creates: functions/__main__/complex_function/arg0=[1, 2, 3]_config={'key': 'val'}_enabled=True_mode=super_long_mode_name/
result = complex_function([1, 2, 3], {"key": "val"}, False, "super_long_mode_name_that_gets_truncated")

Lazy vs Eager Handling

plcache automatically preserves the return type:

@cache()
def get_lazy_data(n: int) -> pl.LazyFrame:
    return pl.LazyFrame({"x": range(n)})

@cache()
def get_eager_data(n: int) -> pl.DataFrame:
    return pl.DataFrame({"x": range(n)})

# Returns LazyFrame (cached with lazy semantics)
lazy_result = get_lazy_data(100)

# Returns DataFrame (cached as computed result)
eager_result = get_eager_data(100)

Cache Management

# Clear all cached data
from plcache import PolarsCache

cache_instance = PolarsCache(cache_dir="./my_cache")
cache_instance.clear()

# Or clear default cache
default_cache = PolarsCache()  # Uses default location
default_cache.clear()

Real-World Example

import polars as pl
from plcache import cache

@cache(
    cache_dir="./data_cache",
    symlinks_dir="datasets",
    symlink_name="raw_data.parquet"
)
def load_stock_data(symbol: str, start_date: str, end_date: str) -> pl.LazyFrame:
    """Load stock data - expensive API call, perfect for caching."""
    # Expensive API call or file I/O
    return pl.scan_csv(f"data/{symbol}.csv").filter(
        pl.col("date").is_between(start_date, end_date)
    )

@cache(
    cache_dir="./analysis_cache",
    symlinks_dir="technical_analysis",
    symlink_name="indicators.parquet"
)
def technical_analysis(symbol: str, window: int = 20) -> pl.DataFrame:
    """Compute technical indicators - expensive computation."""
    stock_data = load_stock_data(symbol, "2024-01-01", "2024-12-31")

    return (
        stock_data
        .with_columns([
            pl.col("close").rolling_mean(window).alias("sma"),
            pl.col("close").rolling_std(window).alias("volatility")
        ])
        .collect()
    )

# Usage - only computes once per unique combination
aapl_analysis = technical_analysis("AAPL", window=20)

# Cache structure created:
# ./data_cache/datasets/__main__/load_stock_data/arg0=AAPL_arg1=2024-01-01_arg2=2024-12-31/raw_data.parquet
# ./analysis_cache/technical_analysis/__main__/technical_analysis/arg0=AAPL_arg1=20/indicators.parquet

Examples

See the examples/ directory for comprehensive usage examples:

• examples/basic/ - Simple usage patterns and getting started
• examples/advanced/ - Configuration options and advanced features
• examples/perf/ - Performance comparisons and benchmarks

Usage Tips

1. Use appropriate return types: Return LazyFrame for large datasets you'll filter later
2. Cache at the right level: Cache expensive I/O operations, not cheap transformations
3. Monitor cache size: Set reasonable size_limit to avoid disk space issues
4. organise with symlinks_dir: Use descriptive names like "experiments", "datasets", "analysis" for different cache types
5. Custom symlink names: Use descriptive filenames like "raw_data.parquet", "results.parquet" to identify cache contents

License

MIT License - see LICENSE file for details.

Contributing

Contributions welcome! Please see CONTRIBUTING.md for guidelines.