[
https://issues.apache.org/jira/browse/ARROW-7885?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17040383#comment-17040383
]
Antoine Pitrou commented on ARROW-7885:
---------------------------------------
Perhaps we should update the PyArrow documentation. Nowadays you can do
efficient zero-copy serialization using pickle protocol 5 and out-of-band
buffers: https://docs.python.org/3/library/pickle.html#out-of-band-buffers
Even without using out-of-band buffers, pickle should make fewer copies than it
used to. In any case, before believing PyArrow serialization is "more
efficient", I would suggest you run benchmarks on your own data (for example a
Pandas dataframe).
As for Parquet, you should be able to write it in memory if you call
{{pyarrow.parquet.write_table}} with a {{pyarrow.BufferOutputStream}}. For
example:
{code:python}
>>> tab = pa.Table.from_pydict({'a': list(range(10000))})
>>>
>>>
>>> stream = pa.BufferOutputStream()
>>>
>>>
>>> pq.write_table(tab, stream)
>>>
>>>
>>> buf = stream.getvalue()
>>>
>>>
>>> buf.size
>>>
>>>
58126
{code}
> [Python] pyarrow.serialize does not support dask dataframe
> ----------------------------------------------------------
>
> Key: ARROW-7885
> URL: https://issues.apache.org/jira/browse/ARROW-7885
> Project: Apache Arrow
> Issue Type: Wish
> Components: Python
> Reporter: Benjamin
> Priority: Minor
>
> Currently pyarrow knows how to serialize pandas dataframes but not dask
> dataframes.
> {code:java}
> SerializationCallbackError: pyarrow does not know how to serialize objects of
> type <class 'dask.dataframe.core.DataFrame'>. {code}
> Pickling the dask dataframe foregoes the benefits of using pyarrow for the
> sub dataframes.
> Pyarrow support for serializing dask dataframes would allow storing
> dataframes efficiently in a database instead of a file system (e.g. parquet).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
