Question

从CSV加载数据时，无法加载某些CSV，导致分区为空。我想删除所有空分区，因为有些方法似乎不适用于空分区。我试图重新分区，例如repartition(npartitions=10)工作，但是大于此值的值仍然可能导致空分区。

实现这一目标的最佳方法是什么？感谢。

Answer 1

我发现过滤Dask数据帧（例如按日期）通常会导致空分区。如果你在使用带有空分区的数据帧时遇到问题，这里有一个基于MRocklin指导的函数来剔除它们：

def cull_empty_partitions(df):
    ll = list(df.map_partitions(len).compute())
    df_delayed = df.to_delayed()
    df_delayed_new = list()
    pempty = None
    for ix, n in enumerate(ll):
        if 0 == n:
            pempty = df.get_partition(ix)
        else:
            df_delayed_new.append(df_delayed[ix])
    if pempty is not None:
        df = dd.from_delayed(df_delayed_new, meta=pempty)
    return df

Answer 2

对于使用Bags（而不是DataFrames）的任何人，此功能都能解决问题：

def cull_empty_partitions(bag):
    """
    When bags are created by filtering or grouping from a different bag,
    it retains the original bag's partition count, even if a lot of the
    partitions become empty.
    Those extra partitions add overhead, so it's nice to discard them.
    This function drops the empty partitions.
    """
    bag = bag.persist()
    def get_len(partition):
        # If the bag is the result of bag.filter(),
        # then each partition is actually a 'filter' object,
        # which has no __len__.
        # In that case, we must convert it to a list first.
        if hasattr(partition, '__len__'):
            return len(partition)
        return len(list(partition))
    partition_lengths = bag.map_partitions(get_len).compute()

    # Convert bag partitions into a list of 'delayed' objects
    lengths_and_partitions = zip(partition_lengths, bag.to_delayed())

    # Drop the ones with empty partitions
    partitions = (p for l,p in lengths_and_partitions if l > 0)

    # Convert from list of delayed objects back into a Bag.
    return dask.bag.from_delayed(partitions)

Answer 3

这是我删除空分区的尝试：

import numpy as np

def remove_empty_partitions(ddf):
    """ remove empty partitions """
    partition_lens = ddf.map_partitions(len).compute()
    ids_of_empty_partitions = np.where(partition_lens==0)
    if len(ids_of_empty_partitions) == len(partition_lens):
        # all partitions are empty
        ddf_nonzero = ddf.partitions[0]
    elif len(ids_of_empty_partitions)>0:
        ddf_nonzero = dd.concat([
            ddf.get_partition(num_partition) for num_partition, partition in enumerate(ddf.partitions)
            if num_partition not in ids_of_empty_partitions
        ])
    return ddf_nonzero

FWIW，@tpegbert 的答案在获取过滤数据帧所需的任务数量方面似乎更有效。

Answer 4

没有简单的API可以做到这一点。您可以致电df.map_partitions(len)以确定哪些分区为空，然后明确删除它们，可能是使用df.to_delayed()和dask.dataframe.from_delayed(...)。

将来，如果您发现一个功能不能很好地处理空分区，那么您将非常感激。 https://github.com/dask/dask/issues/new

删除Dask中的空分区

4 个答案: