The interface between Avro and pandas DataFrame

apache-avro, pandas, python
pip install pandavro==1.7.0



Build Status

The interface between Apache Avro and pandas DataFrame.


pandavro is available to install from PyPI.

$ pip install pandavro


It prepares like pandas APIs:

  • read_avro
    • Read the records from Avro file and fit them into pandas DataFrame using fastavro.
  • to_avro
    • Write the rows of pandas DataFrame to Avro file with the original schema infer.

What can and can't pandavro do?

Avro can represent the following kinds of types:

  • Primitive types (null, bool, int etc.)
  • Complex types (records, arrays, maps etc.)
  • Logical types (annotated primitive/complex type to represent e.g. datetime)

When converting to Avro, pandavro will try to infer the schema. It will output a non-nested schema without any indexes set on the dataframe and it will also not try to infer if any column can be nullable so all columns are set as nullable, i.e. a boolean will be encoded in Avro schema as ['null', 'bool'].

Pandavro can handle these primitive types:

Numpy/pandas type Avro primitive type
np.bool_ boolean
np.float32 float
np.float64 double
np.unicode_ string
np.object_ string
np.int8, np.int16, np.int32 int
np.uint8, np.uint16, np.uint32 "unsigned" int*
np.uint64 "unsigned" long*
np.int64, pd.Int64Dtype long
pd.Int8Dtype, pd.Int16Dtype, pd.Int32Dtype int
pd.UInt8Dtype, pd.UInt16Dtype, pd.UInt32Dtype "unsigned" int*
pd.StringDtype** string
pd.BooleanDtype** boolean

* We represent the unsigned versions of these integers by adding the non-standard "unsigned" flag as such: {'type': 'int', 'unsigned': True}. Pandas 0.24 added support for nullable integers. Writing pd.UInt64Dtype is not supported by fastavro.

** Pandas 1.0.0 added support for nullable string and boolean datatypes.

Pandavro also supports these logical types:

Numpy/pandas type Avro logical type
np.datetime64, pd.DatetimeTZDtype, pd.Timestamp timestamp-micros*
If a boolean column includes empty values, pandas classifies the column as having a dtype of object - this is accounted for in complex column handling.

And these complex types - all complex types other than 'fixed' will be classified by pandas as having a dtype of object, so their underlying python types are used to determine the Avro type:

Numpy/Python type Avro complex type
dict, collections.OrderedDict record
list array
np.void fixed

Record and array types can be arbitrarily nested within each other.

The schema definition of a record requires a unique name for the record separate from the column itself. This does not map to any concept in pandas, so for this we just append '_record' to the original column name and a number to ensure that there are zero duplicate 'name' values.

The remaining Avro complex types are not currently supported for the following reasons:

  1. Enum: The closest pandas type to Avro's enum type is pd.Categorical, but it still is not a complete match. Possible values of the enum type can only be alphanumeric strings, whereas pd.Categorical values have no such limitation.
  2. Map: No strictly matching concept in Python/pandas - Python dictionaries can have arbitrarily typed keys. Functionality can be essentially be achieved with the record type.
  3. Union: Any column with mixed types (other than empty values/NoneType) are treated by pandas as having a dtype of object, and will be written as strings. It would be difficult to deterministically infer multiple allowed data types based solely on a column's contents.

And these logical types:

Numpy/pandas type Avro logical type
np.datetime64, pd.DatetimeTZDtype, pd.Timestamp timestamp-micros/timezone-millis

Note that the timestamp must not contain any timezone (it must be naive) because Avro does not support timezones. Timestamps are encoded as microseconds by default, but can be encoded in milliseconds by using times_as_micros=False

* If passed to_avro(..., times_as_micros=False), this has a millisecond resolution.

Due to an inherent design choice in fastavro, it interprets a naive datetime in the system's timezone before serializing it. This has the consequence that your naive datetime will not correctly roundtrip to and from an Avro file. Always indicate a timezone to avoid the system timezone introducing problems.

If you don't want pandavro to infer the schema but instead define it yourself, pass it using the schema kwarg to to_avro.

Loading Pandas nullable datatypes

The nullable datatypes indicated in the table above are easily written to Avro, but loading them introduces ambiguity as we can use either the old, default or these new datatypes. We solve this by using a special keyword when loading to force conversion to these new NA-supporting datatypes:

import pandavro as pdx

# Load datatypes as NA-compatible datatypes where possible
pdx.read_avro(path, na_dtypes=True)

This is different from convert_dtypes as it does not infer the datatype based on the actual values, but it looks at the Avro schema so is deterministic and not dependent on the actual values.

Also note that, in "normal" mode, numpy int/uint dtypes are all read back as np.int64 due to how fastavro reads them. (This could be worked around by converting type after loading, PRs welcome.) In na_dtypes=True mode they are loaded correctly as Pandas NA-dtypes, but with no less than 32 bits of resolution (less is not supported by Avro so we can not infer it from the schema).


See tests/ for more examples.

import os
import numpy as np
import pandas as pd
import pandavro as pdx


def main():
    df = pd.DataFrame({
        "Boolean": [True, False, True, False],
        "pdBoolean": pd.Series([True, None, True, False], dtype=pd.BooleanDtype()),
        "Float64": np.random.randn(4),
        "Int64": np.random.randint(0, 10, 4),
        "pdInt64":  pd.Series(list(np.random.randint(0, 10, 3)) + [None], dtype=pd.Int64Dtype()),
        "String": ['foo', 'bar', 'foo', 'bar'],
        "pdString": pd.Series(['foo', 'bar', 'foo', None], dtype=pd.StringDtype()),
        "DateTime64": [pd.Timestamp('20190101'), pd.Timestamp('20190102'),
                       pd.Timestamp('20190103'), pd.Timestamp('20190104')]

    pdx.to_avro(OUTPUT_PATH, df)
    saved = pdx.read_avro(OUTPUT_PATH)

if __name__ == '__main__':