df_file_interchange 0.0.14a0

Ensures exact round-trip saves with Pandas (CSV/Parquet) and storage of validatable metadata

df_file_interchange

For data pipelines, it's often advantageous to be able to "round trip" a dataframe to disc: the results of what is proccessed at one stage being required by independent stages afterwards.

Pandas can store and manipulate various structures on top of a basic table including indexes and dtype specifications. These indexes can be generated from a small number of parameters that cannot naturally be encoded in a column, e.g. a pd.RangeIndex. When saving to a CSV file, these indexes are typically enumerated. Therefore, when reading the dataframe back, it is not actually the same as the original. Although this is handled better when writing to Parquet format, there are still a few issues such as all entries in a column, and so an index, being required to have the same dtype. Similarly, it's often desirable to store some custom structured metadata along with the dataframe, e.g. the author or columnwise metadata such as what unit each column is expressed in.

df_file_interchange is a wrapper around Pandas that tries to address these requirements. It stores additional metadata in an accompanying YAML file that (mostly) ensures the indexes are recreated properly, column dtypes are specified properly, and that the dataframe read from disc is actually the same as what was written. It also manages the storage of additional (structured) custom metadata.

In the future, the intention is also to opportunistically serialise columns when required since there are some dtypes that are not supported by Parquet.

This is not trying to reinvent-the-wheel, only supplement existing functionality found in Pandas.

Documentation

The documentation is here.

Summary of Usage

Basic Usage for Writing+Reading CSV/Parquet

The import should be done as,

import df_file_interchange as fi

Next, lets write a dataframe to file. In simple form, this can be done as,

metafile = fi.write_df_to_file(df, datafile)

where datafile should be a Path object (str is converted to Path) and the extension of datafile determines whether a CSV or Parquet file is created. To specify the datafile format explicitly, there are convenience functions,

metafile = fi.write_df_to_csv(df, datafile)

metafile = fi.write_df_to_parquet(df, datafile)

The returned metafile is a Path object pointing to the YAML file containing additional metadata.

Additional parameters can be used, e.g. metafile to specify the YAML filename explicitly, or custom_info. metafile must be in the same directory as datafile so can be specified either as only the filename or with the directory part matching datafile. The file format can also be specified in the parameters.

metafile = fi.write_df_to_file(
    df, datafile, metafile, file_format="csv"
)

Additional encoding options can be specified using the encoding argument (as a FIEncoding object) but this is unnecessary and usually unwise.

To read:

(df, metainfo) = fi.read_df(metafile)

where metafile is the location of the YAML metafile that accompanies the datafile (returned from the write functions). The returned df is of course the dataframe and metainfo is a Pydantic object containing all the metainfo associated with the file encoding, indexes, dtypes, etc, and also the user-supplied custom info.

Structured Metadata

There's support for storing custom metadata, in the sense that it both can be validated using Pydantic models and is extensible.

NOTE: for development purposes, we currently use a "loose" validation, which may result in 'missing' info/units, and a warning; later, this will be made strict and so would raise an exception.

In principle, we can store any custom information that is derived from the FIBaseCustomInfo class. However, it's strongly advised to use FIStructuredCustomInfo. At time of writing, this supports general 'table-wide' information (as FIBaseExtraInfo or FIStdExtraInfo), and columnwise units (derived from FIBaseUnit).

To allow both validation using Pydantic and extensibility is slightly fiddly: when reading, we have to know which classes to instantiate. By default, the code at the moment checks globals() and checks that the supplied class is a subclass of the appropriate base class. However, one can manually define a context to enforce which classes can be instantiated, which might be necessary when extending functionality.

Anyway, a simple example (same example used in tests).

# Create simple dataframe
df = pd.DataFrame(np.random.randn(3, 4), columns=["a", "b", "c", "d"])
df["pop"] = pd.array([1234, 5678, 101010])

# Create units
unit_cur_a = fi.ci.unit.currency.FICurrencyUnit(unit_desc="USD", unit_multiplier=1000)
unit_cur_b = fi.ci.unit.currency.FICurrencyUnit(unit_desc="EUR", unit_multiplier=1000)
unit_cur_c = fi.ci.unit.currency.FICurrencyUnit(unit_desc="JPY", unit_multiplier=1000000)
unit_cur_d = fi.ci.unit.currency.FICurrencyUnit(unit_desc="USD", unit_multiplier=1000)
unit_pop = fi.ci.unit.population.FIPopulationUnit(unit_desc="people", unit_multiplier=1)

# Create extra info
extra_info = fi.ci.structured.extra.FIStdExtraInfo(author="Spud", source="Potato")

# Create custom info with these
custom_info = fi.ci.structured.FIStructuredCustomInfo(
    extra_info=extra_info,
    col_units={
        "a": unit_cur_a,
        "b": unit_cur_b,
        "c": unit_cur_c,
        "d": unit_cur_d,
        "pop": unit_pop,
    },
)

# Write to file
metafile = fi.write_df_to_csv(df, Path("./test_with_structured_custom_info.csv"), custom_info=custom_info)

and to read (we use the default context)

(df_reload, metainfo_reload) = fi.read_df(metafile)

If you were extending the extra info or units, you'll probably have to include a context. Please see the documentation for an explanation of how to do this.

Known Problems

• Pyarrow won't encode numpy's complex64. So, we've disabled this in the tests for now although the functionality will work in CSV. Solution would be to serialize the actual data column when necessary but that'd be new functionality.

• Float16s won't work properly and will never work, due to Pandas decision.

• Some of the newer or more esoteric features of Pandas are not yet accounted for.