Using PyTransform from YAML

Beam YAML provides the ability to easily invoke Python transforms via the PyTransform type, simply referencing them by fully qualified name. For example,

- type: PyTransform
  config:
    constructor: apache_beam.pkg.module.SomeTransform
    args: [1, 'foo']
    kwargs:
       baz: 3

will invoke the transform apache_beam.pkg.mod.SomeTransform(1, 'foo', baz=3). This fully qualified name can be any PTransform class or other callable that returns a PTransform. Note, however, that PTransforms that do not accept or return schema’d data may not be as useable to use from YAML. Restoring the schema-ness after a non-schema returning transform can be done by using the callable option on MapToFields which takes the entire element as an input, e.g.

- type: PyTransform
  config:
    constructor: apache_beam.pkg.module.SomeTransform
    args: [1, 'foo']
    kwargs:
       baz: 3
- type: MapToFields
  config:
    language: python
    fields:
      col1:
        callable: 'lambda element: element.col1'
        output_type: string
      col2:
        callable: 'lambda element: element.col2'
        output_type: integer

This can be used to call arbitrary transforms in the Beam SDK, e.g.

pipeline:
  transforms:
    - type: PyTransform
      name: ReadFromTsv
      input: {}
      config:
        constructor: apache_beam.io.ReadFromCsv
        kwargs:
           path: '/path/to/*.tsv'
           sep: '\t'
           skip_blank_lines: True
           true_values: ['yes']
           false_values: ['no']
           comment: '#'
           on_bad_lines: 'skip'
           binary: False
           splittable: False

Defining a transform inline using __constructor__

If the desired transform does not exist, one can define it inline as well. This is done with the special __constructor__ keywords, similar to how cross-language transforms are done.

With the __constuctor__ keyword, one defines a Python callable that, on invocation, returns the desired transform. The first argument (or source keyword argument, if there are no positional arguments) is interpreted as the Python code. For example

- type: PyTransform
  config:
    constructor: __constructor__
    kwargs:
      source: |
        def create_my_transform(inc):
          return beam.Map(lambda x: beam.Row(a=x.col2 + inc))

      inc: 10

will apply beam.Map(lambda x: beam.Row(a=x.col2 + 10)) to the incoming PCollection.

As a class object can be invoked as its own constructor, this allows one to define a beam.PTransform inline, e.g.

- type: PyTransform
  config:
    constructor: __constructor__
    kwargs:
      source: |
        class MyPTransform(beam.PTransform):
          def __init__(self, inc):
            self._inc = inc
          def expand(self, pcoll):
            return pcoll | beam.Map(lambda x: beam.Row(a=x.col2 + self._inc))

      inc: 10

which works exactly as one would expect.

Defining a transform inline using __callable__

The __callable__ keyword works similarly, but instead of defining a callable that returns an applicable PTransform one simply defines the expansion to be performed as a callable. This is analogous to BeamPython’s ptransform.ptransform_fn decorator.

In this case one can simply write

- type: PyTransform
  config:
    constructor: __callable__
    kwargs:
      source: |
        def my_ptransform(pcoll, inc):
          return pcoll | beam.Map(lambda x: beam.Row(a=x.col2 + inc))

      inc: 10

External transforms

One can also invoke PTransforms define elsewhere via a python provider, for example

pipeline:
  transforms:
    - ...
    - type: MyTransform
      config:
        kwarg: whatever

providers:
  - ...
  - type: python
    input: ...
    config:
      packages:
        - 'some_pypi_package>=version'
    transforms:
      MyTransform: 'pkg.module.MyTransform'

These can be defined inline as well, with or without dependencies, e.g.

pipeline:
  transforms:
    - ...
    - type: ToCase
      input: ...
      config:
        upper: True

providers:
  - type: python
    config: {}
    transforms:
      'ToCase': |
        @beam.ptransform_fn
        def ToCase(pcoll, upper):
          if upper:
            return pcoll | beam.Map(lambda x: str(x).upper())
          else:
            return pcoll | beam.Map(lambda x: str(x).lower())