Avro record class and reader generator
Clone or download
rbystrit Merge pull request #3 from crashfrog/master
Supporting module-as-command syntax ("python -m avrogen")
Latest commit bce34b6 Sep 25, 2018



Build Status codecov

Avro record class and specific record reader generator.

Current Avro implementation in Python is completely typelss and operates on dicts. While in many cases this is convenient and pythonic, not being able to discover the schema by looking at the code, not enforcing schema during record constructions, and not having any context help from the IDE could hamper developer performance and introduce bugs.

This project aims to rectify this situation by providing a generator for constructing concrete record classes and constructing a reader which wraps Avro DatumReader and returns concrete classes instead of dicts. In order not to violate Avro internals, this functionality is built strictly on top of the DatumReader and all the specific record classes dict wrappers which define accessor properties with proper type hints for each field in the schema. For this exact reason the generator does not provide an overloaded DictWriter; each specific record appears just to be a regular dictionary.

schema_json = "....."
output_directory = "....."
from avrogen import write_schema_files

write_schema_files(schema_json, output_directory)

The generator will create output directory if it does not exist and put generated files there. The generated files will be:


  • __init__.py
  • schema_classes.py
  • submodules*

In order to deal with Avro namespaces, since python doesn't support circular imports, the generator will emit all records into schema_classes.py as nested classes. The top level class there will be SchemaClasses, whose children will be classes representing namespaces. Each namespace class will in turn contain classes for records belonging to that namespace.

Consider following schema:

 {"type": "record", "name": "tweet", "namespace": "com.twitter.avro", "fields": [{"name": "ID", "type": "long" }

Then schema_classes.py would contain:

class SchemaClasses(object):
    class com(object):
        class twitter(object):
            class acro(object):
                class tweetClass(DictWrapper):
                    def __init__(self, inner_dict=None):
                    def ID(self):
                        :rtype: long
                        return self._inner_dict.get('ID', None)
                    def ID(self, value):
                        #:param long value:
                        self._inner_dict['ID'] = value                        

In order to map specific record types and namespaces to modules, so that proper importing can be supported, there generator will create a sub-module under the output directory for each namespace which will export names of all types contained in that namespace. Types declared with empty namespace will be exported from the root module.

So for the example above, output directory will look as follows:


  • __init__.py
  • schema_classes.py
  • com
  • twitter
    • avro
      • __init__.py

The contents of OUTPUT_DIR/com/twitter/avro/__init__.py will be:

from ....schema_classes import SchemaClasses
tweet = SchemaClasses.com.twitter.avro.tweet

So in your code you will be able to say:

from OUTPUT_DIR.com.twitter.avro import tweet
from OUTPUT_DIR import SpecificDatumReader as TweetReader, SCHEMA as your_schema
from avro import datafile, io
my_tweet = tweet()

my_tweet.ID = 1
with open('somefile', 'w+b') as f:
    writer = datafile.DataFileWriter(f,io.DatumWriter(), your_schema)

with open('somefile', 'rb') as f:
    reader = datafile.DataFileReader(f,TweetReader(readers_schema=your_schema))
    my_tweet1 = reader.next()

Avro protocol support

Avro protocol support is implemented the same way as schema support. To generate classes for a protocol:

protocol_json = "....."
output_directory = "....."
from avrogen import write_protocol_files

write_protocol_files(protocol_json, output_directory)

The structure of the generated code will be exactly same as for schema, but in addition to regular types, *Request types will be generated in the root namespace of the protocol for each each message defined.

Logical types support

Avrogen implements logical types on top of standard avro package and supports generation of classes thus typed. To enable logical types support, pass use_logical_types=True to schema and protocol generators. If custom logical types are implemented and such types map to types other than simple types or datetime.* or decimal.* then pass custom_imports parameter to generator functions so that your types are imported. Types implemented out of the box are:

  • decimal (using string representation only)
  • date
  • time-millis
  • time-micros
  • timestamp-millis
  • timestamp-micros

To register your custom logical type, inherit from avrogen.logical.LogicalTypeProcessor, implement abstract methods, and add an instance to avrogen.logical.DEFAULT_LOGICAL_TYPES dictionary under the name of your logical type. A sample implementation looks as follows:

class DateLogicalTypeProcessor(LogicalTypeProcessor):
    _matching_types = {'int', 'long', 'float', 'double'}

    def can_convert(self, writers_schema):
        return isinstance(writers_schema, schema.PrimitiveSchema) and writers_schema.type == 'int'

    def validate(self, expected_schema, datum):
        return isinstance(datum, datetime.date)

    def convert(self, writers_schema, value):
        if not isinstance(value, datetime.date):
            raise Exception("Wrong type for date conversion")
        return (value - EPOCH_DATE).total_seconds() // SECONDS_IN_DAY

    def convert_back(self, writers_schema, readers_schema, value):
        return EPOCH_DATE + datetime.timedelta(days=int(value))

    def does_match(self, writers_schema, readers_schema):
        if isinstance(writers_schema, schema.PrimitiveSchema):
            if writers_schema.type in DateLogicalTypeProcessor._matching_types:
                return True
        return False

    def typename(self):
        return 'datetime.date'

    def initializer(self, value=None):
        return ((
                    'logical.DateLogicalTypeProcessor().convert_back(None, None, %s)' % value) if value is not None
                else 'datetime.datetime.today().date()')

To read/write data with logical type support, use generated SpecificDatumReader and a LogicalDatumWriter from avro.logical.