我的数据是一个10GB的文件,格式如下。
[ 1234567890 ][ 2020052701020201 ][ value1 ][ value2 ][ key3 = value3 ]...[ keyn = valuen ]
注:
[
和 ]
是在值本身,例如 [ hello = wo[rld] ]
[
和 ]
值中。在我简单的 for line in f:
函数,我可以通过 ' ][ '
然而鉴于文件的大小,dask是非常有利可图的。
我知道,与 engine='c'
我不能使用多字符分隔符,但切换到了 engine='python'
导致了不可预知的结果,下面是一个例子。
def init_ddf(filename):
return ddf.read_csv(
filename,
blocksize="1GB",
sep="]",
usecols=[1, 8],
na_filter=False,
names=["hello", World" ],
engine="c",
)
上面的代码按照预期的结果是 ParserError: Too many columns specified: expected 25 and found 24
. 这个错误很难重现,因为它只发生在一些特定的行,我很难识别。它不会发生在每次有更多列的时候。所以在上面的函数中,我修改了 engine="python"
和 sep=" \]\[ "
. 但在10G文件中,我得到以下不可预测的行为。
def init_pyddf(filename, usecols, names):
return ddf.read_csv(
filename,
blocksize="1GB",
sep=" \]\[ ",
usecols=usecols,
na_filter=False,
names=names,
engine="python",
)
In [50]: !head /tmp/foo /tmp/bar
==> /tmp/foo <==
[ 1234567890 ][ 2020052701020201 ][ value1 ][ value2 ][ key3 = value3 ][ keyn = valuen ]
[ 1590471107 ][ 20200526T0731460 ][ THEOQQ ][ e = CL ][ Even = 175134 ][ rded = a12344 ][ blah = INVALID ][ N = T ][ ED = 13606 ]
==> /tmp/bar <==
[ 1234567890 ][ 2020052701020201 ][ value1 ][ value2 ][ key3 = value3 ][ keyn = valuen ]
[ 1590471107 ][ 20200526T0731460 ][ THEOQQ ][ e = CL ][ Even = 175134 ][ rded = a12344 ]
In [51]: init_pyddf("/tmp/foo", [1,2], ["time", "name"]).compute()
Out[51]:
time name
[ 1234567890 2020052701020201 value1 key3 = value3 keyn = valuen ]
[ 1590471107 20200526T0731460 THEOQQ Even = 175134 rded = a12344
In [52]: init_pyddf("/tmp/bar", [1,2], ["time", "name"]).compute()
Out[52]:
time name
0 2020052701020201 value1
1 20200526T0731460 THEOQQ
一些更多的例子:
In [110]: !cat /tmp/dummy
[ 0 ][ 000000000000000000000000000 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
[ 1 ][ 20200526T073146.901861+0200 ][ T ][ E ][ E ][ F ][ W ][ N ][ E ][ E ][ 5 ]
In [111]: init_pyddf("/tmp/dummy", [1,7], ["time", "name"]).compute().head()
Out[111]:
time name
[ 0 0 0
[ 1 T E
In [112]: !cat /tmp/dummy
[ 0 ][ 000000000000000000000000000 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
[ 1 ][ 20200526T073146.901861+0200 ][ T ][ E ][ E ][ F ][ W ][ N ][ E ][ E ]
In [113]: init_pyddf("/tmp/dummy", [1,7], ["time", "name"]).compute().head()
Out[113]:
time name
0 000000000000000000000000000 0
1 20200526T073146.901861+0200 N
In [119]: !cat /tmp/dummy
[ 0 ][ 000000000000000 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
[ 1 ][ 20200526T073146 ][ T ][ D ][ F ][ W ][ e ][ E ][ E ][ I ][ T ][ T ][ S ][ S ][ B ][ A ][ E ][ F ][ S ][ P][ T = Y ][ 0 ]
In [120]: init_pyddf("/tmp/dummy", [1,7], ["time", "name"]).compute()
Out[120]:
time name
[ 0 000000000000000 0 0 0 0 0 0 0 0 ] NaN None None
[ 1 20200526T073146 T D F W e E E I T S S
鉴于您有一个更复杂的基于文本的文件格式,您可以先从 Dask Bag 开始,使用普通的 Python 函数来生成 python 字典,然后用以下方法将 Bag 转换为 Dask Dataframe to_dataframe
方法。
import dask.bag
b = dask.bag.read_text("my-files.*.txt")
def parse(line: str) -> dict:
...
records = b.map(parse)
df = b.to_dataframe()