如何使用Dask读取HTTP(s) octet-stream返回的Parquet文件

部分后端接口会以octet-stream格式返回Parquet文件。在Pandas中可以通过如下代码完成读取：

import io
import requests
import pandas as pd

result = requests.get("https://..../file.parquet")
df = pd.read_parquet(io.BytesIO(result.content))

我能否在Dask中实现相同的能力？我尝试直接运行以下代码：

import dask.dataframe as dd

dd.read_parquet("https://..../file.parquet")

程序抛出了异常（返回结果为类字节对象导致）：

File "to_parquet_dask.py", line 153, in <module>
    main(*parser.parse_args())
  File "to_parquet_dask.py", line 137, in main
    download_parquet(
  File "to_parquet_dask.py", line 121, in download_parquet
    dd.read_parquet(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/dask/dataframe/io/parquet/core.py", line 313, in read_parquet
    read_metadata_result = engine.read_metadata(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/dask/dataframe/io/parquet/fastparquet.py", line 733, in read_metadata
    parts, pf, gather_statistics, base_path = _determine_pf_parts(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/dask/dataframe/io/parquet/fastparquet.py", line 148, in _determine_pf_parts
    elif fs.isdir(paths[0]):
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fsspec/asyn.py", line 88, in wrapper
    return sync(self.loop, func, *args, **kwargs)
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fsspec/asyn.py", line 69, in sync
    raise result[0]
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fsspec/asyn.py", line 25, in _runner
    result[0] = await coro
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fsspec/implementations/http.py", line 418, in _isdir
    return bool(await self._ls(path))
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fsspec/implementations/http.py", line 195, in _ls
    out = await self._ls_real(url, detail=detail, **kwargs)
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fsspec/implementations/http.py", line 150, in _ls_real
    text = await r.text()
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/aiohttp/client_reqrep.py", line 1082, in text
    return self._body.decode(encoding, errors=errors)  # type: ignore
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x90 in position 7: invalid start byte

更新 #

参照@mdurant的回答对fsspec进行调整后，遇到了如下报错：

ValueError: Cannot seek streaming HTTP file

在URL前添加了"simplecache::"前缀后，又出现了新的报错：

Traceback (most recent call last):
  File "to_parquet_dask.py", line 161, in <module>
    main(*parser.parse_args())
  File "to_parquet_dask.py", line 145, in main
    download_parquet(
  File "to_parquet_dask.py", line 128, in download_parquet
    dd.read_parquet(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/dask/dataframe/io/parquet/core.py", line 313, in read_parquet
    read_metadata_result = engine.read_metadata(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/dask/dataframe/io/parquet/fastparquet.py", line 733, in read_metadata
    parts, pf, gather_statistics, base_path = _determine_pf_parts(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/dask/dataframe/io/parquet/fastparquet.py", line 185, in _determine_pf_parts
    pf = ParquetFile(
  File "/home/bc30138/Documents/CODE/flexydrive/driver_style/.venv/lib/python3.8/site-packages/fastparquet/api.py", line 127, in __init__
    raise ValueError("Opening directories without a _metadata requires"
ValueError: Opening directories without a _metadata requiresa filesystem compatible with fsspec

临时解决方案 #

该方式不够规范也不是最优实现，但目前可以正常运行：

import io
import requests
import pandas as pd
import dask
import dask.dataframe as dd

@dask.delayed
def parquet_from_http(url, token):
    result = requests.get(
        url,
        headers={'Authorization': token}
    )
    return pd.read_parquet(io.BytesIO(result.content))
    
delayed_download = parquet_from_http(url, token)
df = dd.from_delayed(delayed_download, meta=meta)

备注：该方案中必须传入meta参数，否则Dask会两次调用该函数：第一次用于获取元数据，第二次用于执行计算，会发起两次重复请求。

内容的提问来源于stack exchange，提问作者bc30138