Read / Write Parquet files without reading into memory (using Python)

---

I looked at the standard documentation that I would expect to capture my need (Apache Arrow and Pandas), and I could not seem to figure it out.

I know Python best, so I would like to use Python, but it is not a strict requirement.

Problem

I need to move Parquet files from one location (a URL) to another (an Azure storage account, in this case using the Azure machine learning platform, but this is irrelevant to my problem).

These files are too large to simply perform pd.read_parquet("https://my-file-location.parquet"), since this reads the whole thing into an object.

Expectation

I thought that there must be a simple way to create a file object and stream that object line by line -- or maybe column chunk by column chunk. Something like

import pyarrow.parquet as pq

with pq.open("https://my-file-location.parquet") as read_file_handle:
    with pq.open("https://my-azure-storage-account/my-file.parquet", "write") as write_filehandle:
        for next_line in read_file_handle{
            write_file_handle.append(next_line)

I understand it will be a little different because Parquet is primarily meant to be accessed in a columnar fashion. Maybe there is some sort of config object that I would pass which specifies which columns of interest, or maybe how many lines can be grabbed in a chunk or something similar.

But the key expectation is that there is a means to access a parquet file without loading it all into memory. How can I do this?

FWIW, I did try to just use Python's standard open function, but I was not sure how to use open with a URL location and a byte stream. If it is possible to do this via just open and skip anything Parquet-specific, that is also fine.

Update

Some of the comments have suggested using bash-like scripts, such as here. I can use this if there is nothing else, but it is not ideal because:

• I would rather keep this all in a full language SDK, whether Python, Go, or whatever. If the solution moves into a bash script with pipes, it requires an external call since the final solution will not be written entirely bash, Powershell, or any scripting language.
• I really want to leverage some of the benefits of Parquet itself. As I mentioned in the comment below, Parquet is columnar storage. So if I have a "data frame" that is 1.1 billion rows and 100 columns, but I only care about 3 columns, I would love to be able to only download those 3 columns, saving a bunch of time and some money, too.

from Read / Write Parquet files without reading into memory (using Python)