Yikun commented on a change in pull request #32926: URL: https://github.com/apache/spark/pull/32926#discussion_r652403201
########## File path: python/docs/source/development/contributing.rst ########## @@ -72,17 +72,86 @@ Preparing to Contribute Code Changes ------------------------------------ Before starting to work on codes in PySpark, it is recommended to read `the general guidelines <https://spark.apache.org/contributing.html>`_. -There are a couple of additional notes to keep in mind when contributing to codes in PySpark: +Additionally, there are a couple of additional notes to keep in mind when contributing to codes in PySpark: + +* **Be Pythonic.** +* **APIs are matched with Scala and Java sides in general.** +* **PySpark specific APIs can still be considered as long as they are Pythonic and do not conflict with other existent APIs, for example, decorator usage of UDFs.** +* **If you extend or modify public API, please adjust corresponding type hints. See `Contributing and Maintaining Type Hints`_ for details.** + +If you are fixing pandas API on Spark (``pyspark.pandas``) package, please consider the design principles below: + +* **Return pandas-on-Spark data structure for big data, and pandas data structure for small data** + Often developers face the question whether a particular function should return a pandas-on-Spark DataFrame/Series, or a pandas DataFrame/Series. The principle is: if the returned object can be large, use a pandas-on-Spark DataFrame/Series. If the data is bound to be small, use a pandas DataFrame/Series. For example, `DataFrame.dtypes` return a pandas Series, because the number of columns in a DataFrame is bounded and small, whereas ``DataFrame.head()`` or ``Series.unique()`` returns a pandas-on-Spark DataFrame/Series, because the resulting object can be large. + +* **Provide discoverable APIs for common data science tasks** + At the risk of overgeneralization, there are two API design approaches: the first focuses on providing APIs for common tasks; the second starts with abstractions, and enables users to accomplish their tasks by composing primitives. While the world is not black and white, pandas takes more of the former approach, while Spark has taken more of the latter. + + One example is value count (count by some key column), one of the most common operations in data science. pandas ``DataFrame.value_count`` returns the result in sorted order, which in 90% of the cases is what users prefer when exploring data, whereas Spark's does not sort, which is more desirable when building data pipelines, as users can accomplish the pandas behavior by adding an explicit ``orderBy``. + + Similar to pandas, pandas API on Spark should also lean more towards the former, providing discoverable APIs for common data science tasks. In most cases, this principle is well taken care of by simply implementing pandas' APIs. However, there will be circumstances in which pandas' APIs don't address a specific need, e.g. plotting for big data. + +* **Guardrails to prevent users from shooting themselves in the foot** + Certain operations in pandas are prohibitively expensive as data scales, and we don't want to give users the illusion that they can rely on such operations in pandas API on Spark. That is to say, methods implemented in pandas API on Spark should be safe to perform by default on large datasets. As a result, the following capabilities are not implemented in pandas API on Spark: + + 1. Capabilities that are fundamentally not parallelizable: e.g. imperatively looping over each element + 2. Capabilities that require materializing the entire working set in a single node's memory. This is why we do not implement `pandas.DataFrame.to_xarray <https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.to_xarray.html>`_. Another example is the ``_repr_html_`` call caps the total number of records shown to a maximum of 1000, to prevent users from blowing up their driver node simply by typing the name of the DataFrame in a notebook. + + A few exceptions, however, exist. One common pattern with "big data science" is that while the initial dataset is large, the working set becomes smaller as the analysis goes deeper. For example, data scientists often perform aggregation on datasets and want to then convert the aggregated dataset to some local data structure. To help data scientists, we offer the following: + + * :func:`DataFrame.to_pandas` that returns a pandas DataFrame, koalas only Review comment: koalas only --> pandas API on Spark only? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
