1
我想知道为什么dd.from_bcolz（）在调用时立即开始做一些处理（当N列上升并且存在字符串类型列时，它会很快增长）。dask.dataframe.from_bcolz（）开始立即处理
而且dd.read_hdf（）调用时没有做太多的处理，但只有当dask.dataframe使用 - 那么read_hdf（）读取和处理HDF5由大块大块......
我怎么样read_hdf现在工作，唯一的问题是hdf5表不能有多于1200列，而dataframe不支持列数组。并且hdf5格式毕竟不是列的...
In [1]: import dask.dataframe as dd
In [2]: import pandas as pd
In [3]: import bcolz, random
In [4]: import numpy as np
In [5]: N = int(1e7)
In [6]: int_col = np.linspace(0, 1, N)
In [7]: ct_disk = bcolz.fromiter(((i,i) for i in range(N)), dtype="i8,i8",\
...: count=N, rootdir=r'/mnt/nfs/ct_.bcolz')
In [8]: for i in range(10): ct_disk.addcol(int_col)
In [9]: import dask.dataframe as dd
In [10]: %time dd.from_bcolz(r'/mnt/nfs/ct_.bcolz', chunksize=1000000, lock=False)
CPU times: user 8 ms, sys: 16 ms, total: 24 ms
Wall time: 32.6 ms
Out[10]: dd.DataFrame<from_bc..., npartitions=10, divisions=(0, 1000000, 2000000, ..., 9000000, 9999999)>
In [11]: str_col= [''.join(random.choice('ABCD1234') for _ in range(5)) for i in range(int(N/10))]*10
In [12]: ct_disk.addcol(str_col, dtype='S5')
In [13]: %time dd.from_bcolz(r'/mnt/nfs/ct_.bcolz', chunksize=1000000, lock=False)
CPU times: user 2.36 s, sys: 56 ms, total: 2.42 s
Wall time: 2.44 s
Out[13]: dd.DataFrame<from_bc..., npartitions=10, divisions=(0, 1000000, 2000000, ..., 9000000, 9999999)>
In [14]: for i in range(10): ct_disk.addcol(str_col, dtype='S5')
In [15]: %time dd.from_bcolz(r'/mnt/nfs/ct_.bcolz', chunksize=1000000, lock=False)
CPU times: user 25.3 s, sys: 511 ms, total: 25.8 s
Wall time: 25.9 s
Out[15]: dd.DataFrame<from_bc..., npartitions=10, divisions=(0, 1000000, 2000000, ..., 9000000, 9999999)>
而且，当N（nrows）长大时它会变得更糟。