chore: add ROADMAP document describing the purpose of the package

ROADMAP.md

@@ -0,0 +1,57 @@
+# pandas-gbq Roadmap
+
+The purpose of this package is to provide a small subset of BigQuery
+functionality that maps well to
+[pandas.read_gbq](https://pandas.pydata.org/docs/reference/api/pandas.read_gbq.html#pandas.read_gbq)
+and
+[pandas.DataFrame.to_gbq](https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.to_gbq.html#pandas.DataFrame.to_gbq).
+Those methods in the pandas library are a thin wrapper to the equivalent
+methods in this package.
+
+## Adding features to pandas-gbq
+
+Considerations when adding new features to pandas-gbq:
+
+* New method? Consider an alternative, as the core focus of this library is
+  `read_gbq` and `to_gbq`.
+* Breaking change to an existing parameter? Consider an alternative, as folks
+  could be using an older version of `pandas` that doesn't account for the
+  change when a newer version of `pandas-gbq` is installed. If you must, please
+  follow a 1+ year deprecation timeline.
+* New parameter? Go for it! Be sure to also send a PR to `pandas` after the
+  feature is released so that folks using the `pandas` wrapper can take
+  advantage of it.
+* New data type? OK. If there's not a good mapping to an existing `pandas`
+  dtype, consider adding one to the `db-dtypes` package.
+
+## Vision
+
+The `pandas-gbq` package should do the "right thing" by default. This means you
+should carefully choose dtypes for maximum compatibility with BigQuery and
+avoid data loss. As new data types are added to BigQuery that don't have good
+equivalents yet in the `pandas` ecosystem, equivalent dtypes should be added to
+the `db-dtypes` package.
+
+As new features are added that might improve performance, `pandas-gbq` should
+offer easy ways to use them without sacrificing usability. For example, one
+might consider using the `api_method` parameter of `to_gbq` to support the
+BigQuery Storage Write API.
+
+A note on `pandas.read_sql`: we'd like to be compatible with this too, for folks
+that need better performance compared to the SQLAlchemy connector.
+
+## Usability
+
+Unlike the more object-oriented client-libraries, it's natural to have a method
+with many parameters in the Python data science ecosystem. That said, the
+`configuration` argument is provided, which takes the REST representation of
+the job configuration so that power users can use new features without the need
+for an explicit parameter being added.
+
+## Conclusion
+
+Keep it simple.
+
+Don't break existing users.
+
+Do the right thing by default.