6

So dask.dataframe.map_partitions() takes a func argument and the meta kwarg. How exactly does it decide its return type? As an example:

Lots of csv's in ...\some_folder. 

ddf = dd.read_csv(r"...\some_folder\*", usecols=['ColA', 'ColB'], 
                                        blocksize=None, 
                                        dtype={'ColA': np.float32, 'ColB': np.float32})
example_func = lambda x: x.iloc[-1] / len(x)
metaResult = pd.Series({'ColA': .1234, 'ColB': .1234})
result = ddf.map_partitions(example_func, meta=metaResult).compute()

I'm pretty new to "distributed" computing, but I would intuitively expect this to return a collection (a list or dict, most likely) of Series objects, yet the result is a Series object that could be considered a concatenation of the results of example_func on each partition. This in and of itself would also suffice, if this series had a MultiIndex to indicate the partition label.

From what I can tell from this question, the docs, and the source code itself, this is because ddf.divisions will return a (None, None, ..., None) as a result of reading csv's? Is there a dask-native way to do this, or do I need to manually go in and break the returned Series (a concatenation of the Series that were returned by example_func on each partition) myself?

Also, feel free to correct my assumptions/practices here, as I'm new to dask.

Community
  • 1
  • 1
StarFox
  • 529
  • 3
  • 10

1 Answers1

5

So dask.dataframe.map_partitions() takes a func argument and the meta kwarg. How exactly does it decide its return type?

map_partition tries to concatenate the results returned by func to either a dask DataFrame or a dask Series object in an 'intelligent' way. This decision is based on the return value of func:

  • If func returns a scalar, map_partitions returns a dask Series object.
  • If func returns a pd.Series object, map_partition returns a dask Series object, in which all pd.Series objects returned by func are concatenated.
  • If func returns a pd.DataFrame, map_partitions returns a dask Dataframe object, in which these pd.DataFrame obejcts are concatenated along the first axis.

If you are interested in the result of a special partition, you could use get_partition(). If the partition label is in general an important information for you, I would consider to assign a separate column of your ddf directly after reading in the data from csv, which contains all the information you need. Afterwards, you could construct func in a way, that it returns a pd.DataFrame containing the result of your calculation in one column and the information you need to identify the result in another.

Arco Bast
  • 3,595
  • 2
  • 26
  • 53