richfile

A more natural approach to saving hierarchical data structures.

richfile saves any Python object to disk and loads it back into the same Python objects.

Four backends are available:

• backend="directory": classic richfile directory trees (default).
• backend="sqlar": single-file SQLite archive (.sqlar) with no compression.
• backend="zip": single-file ZIP archive (.zip) in stored mode (no compression).
• backend="tar": single-file plain TAR archive (.tar) with no compression.

richfile can save any atomic Python object, including custom classes, so long as you can write a function to save and load it. It is intended as a replacement for things like: pickle, json, yaml, HDF5, Parquet, netCDF, zarr, numpy, etc. when you want to save a complex data structure in a human-readable and editable format. We find the richfile format ideal to use when you are building a data processing pipeline and you want to contain intermediate results in a format that allows for custom data types, is insensitive to version changes (pickling issues), allows for easy debugging, and is human readable.

It is easy to use, the code is simple and pure python, and the operations follow ACID principles.

Installation

pip install richfile

Examples

Try out the examples in the demo_notebook.ipynb file.

Usage

Saving and loading data is simple:

## Given some complex data structure
data = {
    "name": "John Doe",
    "age": 25,
    "address": {
        "street": "1234 Elm St",
        "zip": None
    },
    "siblings": [
        "Jane",
        "Jim"
    ],
    "data": [1,2,3],
    (1,2,3): "complex key",
}

## Save it
import richfile as rf
r = rf.RichFile("path/to/data.richfile").save(data)

## Load it back
data = rf.RichFile("path/to/data.richfile").load()

Backends

By default, richfile will use the 'directory' backend. However, you can use other backends:

• 'directory': places the contents into a directory. Slow saving, fast loading. Unwieldy when there are many leaf elements.
• 'sqlar': single-file SQLite archive (.sqlar). Best general use choice. Fast and allows for random access, but does not allow easy navigating in a file browser.
• 'zip': single-file ZIP archive (.zip) in stored mode (no compression). Slower than sqlar, but still performant, and easy to handle.
• 'tar': single-file plain TAR archive (.tar) with no compression. Slower than sqlar, but still performant, and fairly easy to handle.

Save and load using the SQLAR backend:

import richfile as rf

rf.RichFile("path/to/data.sqlar", backend="sqlar").save(data)
data = rf.RichFile("path/to/data.sqlar", backend="sqlar").load()

Convert between backends (raw byte-preserving conversion):

import richfile as rf

## Archive -> directory-style richfile
rf.extract_backend_to_directory(
    path_source="path/to/data.zip",
    backend_source="zip",
    path_directory_out="path/to/data.richfile",
    overwrite=True,
)

## Directory-style richfile -> archive backend
rf.pack_directory_to_backend(
    path_directory_in="path/to/data.richfile",
    backend_target="sqlar",
    path_target="path/to/data.sqlar",
    overwrite=True,
)

## Generic backend -> backend conversion
rf.convert_backend(
    path_source="path/to/data.sqlar",
    backend_source="sqlar",
    path_target="path/to/data.tar",
    backend_target="tar",
    mode="raw",        ## "raw" (byte-preserving) or "semantic" (load/save)
    overwrite=True,
)

You can also load just a part of the data:

r = rf.RichFile("path/to/data.richfile")  ## Path to an existing richfile
first_sibling = r["siblings"][0].load()  ## Lazily load a single item using pythonic indexing
print(f"First sibling: {first_sibling}")

>>> First sibling: Jane

View the contents of a richfile directory without loading it:

r.view_directory_tree()

Output:

Directory structure:
Viewing tree structure of richfile at path: ~/path/data.richfile (dict)
├── name.dict_item (dict_item)
|   ├── key.json (str)
|   ├── value.json (str)
|   
├── age.dict_item (dict_item)
|   ├── key.json (str)
|   ├── value.json (int)
|   
├── address.dict_item (dict_item)
|   ├── key.json (str)
|   ├── value.dict (dict)
|   |   ├── street.dict_item (dict_item)
|   |   |   ├── key.json (str)
|   |   |   ├── value.json (str)
|   |   |   
|   |   ├── zip.dict_item (dict_item)
|   |   |   ├── key.json (str)
|   |   |   ├── value.json (None)
|   |   |   
|   |   
|   
├── siblings.dict_item (dict_item)
|   ├── key.json (str)
|   ├── value.list (list)
|   |   ├── 0.json (str)
|   |   ├── 1.json (str)
|   |   
|   
├── data.dict_item (dict_item)
|   ├── key.json (str)
|   ├── value.list (list)
|   |   ├── 0.json (int)
|   |   ├── 1.json (int)
|   |   ├── 2.json (int)
|   |   
|   
├── 5.dict_item (dict_item)
|   ├── key.tuple (tuple)
|   |   ├── 0.json (int)
|   |   ├── 1.json (int)
|   |   ├── 2.json (int)
|   |   
|   ├── value.json (str)
|   

You can also add new data types easily:

## Add type to a RichFile object
r = rf.RichFile("path/to/data.richfile")
r.register_type(
    type_name='numpy_array',
    function_load=lambda path: np.load(path),
    function_save=lambda path, obj: np.save(path, obj),
    object_class=np.ndarray,
    library='numpy',
    suffix='npy',
)

## OR
## Add type to the global workspace / kernel so that all new RichFile objects can use it
rf.functions.register_type(
    type_name='numpy_array',
    function_load=lambda path: np.load(path),
    function_save=lambda path, obj: np.save(path, obj),
    object_class=np.ndarray,
    library='numpy',
    suffix='npy',
)

Installation from source

git clone https://github.com/RichieHakim/richfile
cd richfile
pip install -e .

Considerations and Limitations

• Inversibility: When creating custom data types, it is important to consider whether the saving and loading operations are exactly reversible.
• ACID principles are reasonably followed via the use of temporary files, file locks, and atomic operations. However, the library is not a database, and therefore cannot guarantee the same level of ACID compliance as a database. In addition, atomic replacements of existing non-empty directories require two operations, which reduces atomicity.
• Backend selection: You can pass backend explicitly ("directory", "sqlar", "zip", "tar"), or rely on backend="auto" for loading from existing paths. Path suffixes remain informational only.
• Archive performance tradeoff: SQLAR/ZIP/TAR store raw bytes without compression in v1 for faster save behavior. This can increase on-disk size compared with compressed formats.
• Archive scope in v1: SQLAR/ZIP/TAR currently support root-object save and lazy/query load behavior. Nested path mutation/append APIs are intentionally deferred.
• TAR scope in v1: TAR backend writes plain .tar only (no .tar.gz, .tgz, .tar.bz2, or .tar.xz output modes).
• Custom type compatibility in archive backends: Custom function_save(path, ...) and function_load(path, ...) callbacks are supported via a selective temporary-path bridge when needed.
• Backend conversion: convert_backend(..., mode="raw") performs byte-preserving layout conversion and does not deserialize objects. mode="semantic" performs load() + save() and requires matching type registrations.

TODO:

• Tests
• Documentation
• Examples
• Readme
• License
• PyPi
• Hashing
• Item assignment (safely)
• Custom saving/loading functions
• Put the library imports in the function calls
• Add handling for data without a known type
• Change name of library to something more descriptive
• Test out memmap stuff
• Make it a .zip type
• Add mutability
• Archive packing