Support inserting data into BigQuery directly from a Polars DataFrame

[henryharbeck]

Is your feature request related to a problem? Please describe.
The Polars DataFrame library has been gaining a lot of traction and many are writing new pipelines in Polars and/or moving from pandas to Polars. It would be great to add native support between the BigQuery client library and Polars.

Describe the solution you'd like
This request is to allow inserting data directly from a Polars DataFrame into a BigQuery table.
An additional bonus would be not requiring PyArrow to be installed.
I would be open to expanding client.load_table_from_dataframe to also accept Polars DataFrames, or new dedicated method(s) being created.

Describe alternatives you've considered

1. Convert to pandas at the end of the pipeline and use client.load_table_from_dataframe to insert the data. Not ideal to require an additional dependency just to insert data. Furthermore, I don't believe that pandas supports complex types available in both BigQuery and Polars, such as structs and arrays.

2. Write the DataFrame to a bytes stream as a parquet file and insert the data with client.load_table_from_file. This intent of this code is a lot less obvious, and it would be much nicer to have more native support. Note that this is also the suggested approach in the Polars user guide (rightfully so IMO as it does not require any additional dependencies).

3. Do not support Polars directly, but instead support inserting data from a PyArrow table. This is not currently feasible, but would be an alternative feature request. This is not preferable as the option above already allows inserting data without a PyArrow dependency. From looking at the docs (haven't checked the source), this potentially looks to have some overlap with what client.load_table_from_dataframe already does.

[tswast]

I have some thoughts on this: I think a separate package for BigQuery + Polars (similar to pandas-gbq) would be most appropriate here, especially given the desire to avoid pyarrow as a dependency.

[henryharbeck]

Thanks for the response @tswast

It would be great if the BigQuery client accepted Polars DataFrames in addition to pandas DataFrames in client.load_table_from_dataframe. It would currently be possible without a pyarrow dependency as outlined in point 2 in the issue description. I would be happy to provide an implementation if accepted. I do acknowledge that this would be an additional optional dependency.

I can also appreciate a separate package would provide a consolidated reading and writing interface, which I am definitely in support of. Do you have in mind that this package would be developed / owned by Google? If yes, would be it be possible / in-scope to support reading from BigQuery into Polars without a pyarrow dependency? At the moment it seems that polars.from_arrow(client.query(...).to_arrow()) is (from my understanding) the most performant option.

As a heads up, I think I will also request that Polars provide some support for the BigQuery client, and implement the approaches mentioned in their user guide.

EDIT: here is that request - pola-rs/polars#18547

[tswast]

Do you have in mind that this package would be developed / owned by Google?

Yes, that's my thought, though if the community were to create one before I make it through all the red tape needed to make such a thing happen, I'd gladly contribute there, instead. ;-)

If yes, would be it be possible / in-scope to support reading from BigQuery into Polars without a pyarrow dependency?

That's my thought with regards to a separate package. No pyarrow necessary if the focus is just polars. The test suite for google-cloud-bigquery is complicated enough as it is without adding a test environment where polars is installed but pyarrow is not.

[tswast]

I know you asked for writes, but I figured I'd try the read path today, and I was able to get BigQuery table -> polars DataFrame without pyarrow. Checkout this gist: https://gist.github.com/tswast/99b017b20386e324f5c7d2bd49f21b5f#file-bigquery-to-polars-no-pyarrow-ipynb

Obviously, it's single threaded, missing a lot of boilerplate, and doesn't support query inputs, but as a proof of concept, I was happy to see it's possible.