Solid input requirements about dagster

[schrockn]

    sum_sq_solid = dagster_pd.dataframe_solid(
        name='sum_sq',
        inputs=[
            dagster_pd.depends_on(sum_solid),
        ],
        transform_fn=sum_sq_transform_fn
    )

[schrockn]

sum:
   source:CSVORPARQUET
   args:
      format: CSV
      path: whatever

[freiksenet]

Here is summary of my proposal:

• Get rid of typed solid inputs. Inputs are either other solids or named inputs
• Named inputs have an expectation of OUTPUT format, for example dataframe or raw, but they don't have expectation of how exactly to get there

def define_pipeline():
    def sum_transform_fn(num_csv):
        sum_df = num_csv.copy()
        sum_df['sum'] = sum_df['num1'] + sum_df['num2']
        return sum_df

    sum_solid = dagster_pd.dataframe_solid(
        inputs=[dagster_pd.dataframe_input('num')
        name='sum')

    def sum_sq_transform_fn(sum):
        sum_sq = sum.copy()
        sum_sq['sum_sq'] = sum['sum']**2
        return sum_sq

    sum_sq_solid = dagster_pd.dataframe_solid(
        name='sum_sq', input=[dagster_pd.depends_on(sum_solid)], transform_fn=sum_sq_transform_fn
    )

    def always_fails_transform_fn(*_args, **_kwargs):
        raise Exception('I am a programmer and I make error')

    always_fails_solid = dagster_pd.dataframe_solid(
        name='always_fails',
        input=[dagster_pd.depends_on(sum_solid)],
        transform_fn=always_fails_transform_fn
    )

    return dagster.core.pipeline(name='pandas_hello_world', solids=[sum_solid, sum_sq_solid])

Env

environment:
  inputs:
    # either name dependencie name or name of output solid
    - input_name: num
     # Source defines how the input is going to be read into
      source: CSV
      args:
        path: "pandas_hello_world/num.csv"
    - input_name: sum
      # No CSVORPARQUET source cause the actual sournce type is only defined in env
      source: CSV
      args:
        path: "sum.csv"
        format: 'CSV'

This allows changing the source type

environment:
  inputs:
    # either name dependencie name or name of output solid
    - input_name: num
     # Source defines how the input is going to be read into
      source: SQL
      args:
        sql_query: "SELECT * FROM NUM"
    - input_name: sum
      # No CSVORPARQUET source cause the actual sournce type is only defined in env
      source: CSV
      args:
        path: "sum.csv"
        format: 'CSV'

[freiksenet]

This keeps the ability to have some expectations of what the source should return, while not binding pipelines to concrete inputs.

[freiksenet]

def process_raw():
   # Some code to get raw binary data from file and return dataframe
   pass

# We can still typecheck inside the solid that output is valid
transforming_solid = dagster_pd.dataframe_solid(
        inputs=[
          dagster.raw_input('num')
        ],
        name='preprocessing'
        transform_fn=process_raw
)

[freiksenet]

Ok, I'll just do it in even shorter way.

Currently you need to specify input=[dagster_pd.csv_input('num_csv')]. This binds pipeline and solid to concrete input_source.

In addition, when you specify input=[dagster_pd.depends_on(solid)], it gets weird CSVORPARQUET input source type, exactly because input has some opinions on what it's sources should be.

I propose that you can use any input source for any input, without specifying it from pipeline, as long as the result of that source is compatible. So the first thing becomes input=[dagster_pd.dataframe_input(name='num')]. Then you can specify CSV source for it. Or SQL source. Or parquet source. As long as all those sources return dataframe.

For raw data (like getting json file), we use dagster.raw_input or eg dagster.json_input. Again, actual source is specified in env.

[schrockn]

1) CSVORPARQUET is super dumb. This is a holdover from before I created the separate concept of Sources. Hence why I created this issue. #1 Pandas dataframes should a bunch of different sources by default. I think you are proposing exactly what I ended up getting to pretty much. In my view the dataframe solid type should evolve to have a *bunch* of different SourceDefinition's defined by default that cover all of the use cases you mention.

…

On Wed, Jun 6, 2018 at 7:38 AM, Mikhail Novikov ***@***.***> wrote: Ok, I'll just do it in even shorter way. Currently you need to specify input=[dagster_pd.csv_input('num_csv')]. This binds pipeline and solid to concrete input_source. In addition, when you specify input=[dagster_pd.depends_on(solid)], it gets weird CSVORPARQUET input source type, exactly because input has some opinions on what it's sources should be. I propose that you can use any input source for any input, without specifying it from pipeline, as long as the result of that source is compatible. So the first thing becomes input=[dagster_pd.dataframe_ input(name='num')]. Then you can specify CSV source for it. Or SQL source. Or parquet source. As long as all those sources return dataframe. For raw data (like getting json file), we use dagster.raw_input or eg dagster.json_input. Again, actual source is specified in env. — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AbaFeVokVqVa7HoUdUNubJkl9NfOpIGuks5t5-l0gaJpZM4Ua1Ss> .