No surprise I use python, but I’ve recently started experimenting with polars instead of pandas. I’ve enjoyed it so far, but Im not sure if the benefits for my team’s work will be enough to outweigh the cost of moving from our existing pandas/numpy code over to polars.
I’ve also started playing with grafana, as a quick dashboarding utility to make some basic visualizations on some live production databases.
Its a paradigm shift from pandas. In polars, you define a pipeline, or a set of instructions, to perform on a dataframe, and only execute them all at once at the end of your transformation. In other words, its lazy. Pandas is eager, which every part of the transformation happens sequentially and in isolation. Polars also has an eager API, but you likely want to use the lazy API in a production script.
Because its lazy, Polars performs query optimization, like a database does with a SQL query. At the end of the day, if you’re using polars for data engineering or in a pipeline, it’ll likely work much faster and more memory efficient. Polars also executes operations in parallel, as well.
What kind of query optimization can it for scanning data that’s already in memory?
A big feature of polars is only loading applicable data from disk. But during exporatory data analysis (EDA) you often have the whole dataset in memory. In this case, filters wont help much there. Polars has a good page in their docs about all the possible optimizations it is capable of. https://docs.pola.rs/user-guide/lazy/optimizations/
One I see off the top is projection pushdown, which only selects relevant columns for a final transformations. In pandas, if you perform a group by with aggregation, then only look at a few columns, you still perform aggregation across all the data. In polars lazy API, you would define the entire process upfront, and it would know not to aggregate certain columns, for instance.
Hm, that’s kind of interesting
But my first reaction is that optimizations only at the “Python processing level” are going to be pretty limited since it’s not going to have metadata/statistics, and it’d depend heavily on the source data layout, e.g. CSV vs parquet