Uniform file IO API and consolidated codebase

[dhimmel]

There are at least three things that many of the IO methods must deal with: reading from URL, reading/writing to a compressed format, and different text encodings. It would be great if all io functions where these factors were relevant could use the same code (consolidated codebase) and expose the same options (uniform API).

In #14576, we consolidated the codebase but more consolidation is possible. In io.common.py, there are three functions that must be sequentially called to get a file-like object: get_filepath_or_buffer, _infer_compression, and _get_handle. This should be consolidated into a single function, which can then delegate to sub functions.

Currently, pandas supports the following io methods. First for reading:

• read_csv
• read_excel
• read_hdf
• read_feather
• read_sql
• read_json
• read_msgpack (experimental)
• read_html
• read_gbq (experimental)
• read_stata
• read_sas
• read_clipboard
• read_pickle

And then for writing:

• to_csv
• to_excel
• to_hdf
• to_feather
• to_sql
• to_json
• to_msgpack (experimental)
• to_html
• to_gbq (experimental)
• to_stata
• to_clipboard
• to_pickle

Some of these should definitely use the consilidated/uniform API, such as read_csv, read_html, read_pickle, read_excel.

Some functions perhaps should be kept separate, such as read_feather or read_clipboard.

[dhimmel]

Here are my thoughts on the API.

• Read methods should support the following compression methods: None, 'infer', 'gzip', 'bz2', 'xz', 'zip'. Xref ENH: add gzip/bz2 compression to read_pickle() (and perhaps other read_*() methods) #11666

• Write methods should support the following compression methods: None, 'infer', 'gzip', 'bz2', 'xz' (no zip since it's perhaps bad practice).

• We may want to support both long and short compression names. Currently, you specify gzip not gz, but bz2 not bzip2.

• Read methods should support reading from a path, buffer, or URL.

• Write methods should support writing to a path or buffer.

• Textual payloads should support the encoding argument

• Iterator interface should be consistent (support chunksize)

Regarding the consolidated codebase:

• I'd favor greater separation of the code for 2 and 3. This way when pandas becomes 3-only, the entire 2 sections can be deleted.

[jorisvandenbossche]

That sounds great!
If you would like to work towards this, that would be very welcome.

Regarding the py2/py3 separation, I think we should just do what is most practical here (having a certain separation makes the code more clear, too much separation can make it more complex again. In any case, having a few but scattered if PY2 statements are also rather easy to delete). But if all related code is contained in io/common.py, it should not be too difficult to find a good balance in that one file here.

One more consolidation that would be possible for read_csv is between the python and c engine. I think the c engine still has its own logic for handling compression, while I do not think this is needed to be in the cython/c code (I don't think this is the performance sensitive part?)

[dhimmel]

If you would like to work towards this, that would be very welcome.

Let's wait for #13317 and any other IO PRs that I don't know about to be merged. I'm hesitant to commit since I know it will cut into my other obligations. But if no one else is interested in implementing, I'll consider.

I think we should just do what is most practical here

Totally agree. There are still a few things I need to understand before I can make that call. One issue is mode in _get_handle, which currently is poorly documented. Presumably this could include t for text or b for bytes, which will have some interactions with Py 2 or 3.

I think the c engine still has its own logic for handling compression, while I do not think this is needed to be in the cython/c code

Agree the c engine implementation should be consolidated, unless there is a major performance issue. But the duplicated functionality with _get_handle appears not to be c optimized (I'm not sure as I don't know cython).

[jreback]

@dhimmel can you annotate the above (or maybe make it a table)

add an x/check if supports pathlib like things / compression / url

[goldenbull]

agree! 👍
I'm now working on #13317 and found _get_handle a bit complex to understand.
_get_handle needs to deal with varies situations:

• py2 or py3
• binary (pickle, msgpack) or text (csv)
• if text, what's the encoding
• compression
• memory map
• open for read or write

It seems to be better to spilt _get_handle into two or more functions to make each single function simpler