Materializer classes should be able to specify multiple formats/keys about hamilton HOT 1 CLOSED

Isn't this solved by the MatplotlibWriter? First, we could have the sklearn extension built on top of the matplotlib extension. When multiple file types are supported, I think we should infer the type from the path and accept an explicit argument.

For the broader point, I'm starting to wonder if materializers are closer to "a file type" (e.g., parquet, json) or a "render engines" (e.g., matplotlib, s3, mlflow).

On one hand, the "file type" approach relies on a coupling between the Python object class and the output (e.g., saving an XGBoost as JSON). Also, it's convenient when the file type is set, but the "renderer" changes. For example, you specify "parquet" and have the ability to change "where/how" that parquet is handled.

On the other hand, the "renderer" approach is valuable when the Python object -> artifact is looser or that you'd want many formats of the same artifact. One should be able to produce multiple formats without much code duplication. For example, visualizations are the same matplotlit.figure.Figure object, but you could want PNG, SVG or a table to materialize CSV, parquet. What about to.snowflake and having the "renderer" support many operations "create table", "create view", "insert", etc. even though it's essentially "parquet" data. This seems to be a rarer but valuable case.

I'm thinking there is room for both types and they both should live under the name "materializer"

from hamilton.