perf(python): Improve performance of expr_to_lit_or_expr for arguments of type Expr by ~80%

[jakob-keller]

expr_to_lit_or_expr is used extensively throughout the codebase. When providing arguments of type Expr in an attempt to improve performance, the current implementation can do better IMO.

Benchmark

python3 -m timeit -s "import polars; expr = polars.col('test')" "polars.internals.expr_to_lit_or_expr(expr)"

0.16.6: 351 nsec per loop
After PR: 74.1 nsec per loop

[stinodego]

Fair addition. But please change the final elif not isinstance(expr, Expr) line to else. Then this can be merged.

I will probably revise this parsing logic in the near future, but this is a good thing to consider.

[stinodego]

Thanks! Good addition. Expression inputs should be the default, anything else could be considered an ergonomic shortcut. So I think it's good to give 'preferential treatment' to expressions here.

[jakob-keller]

Thanks! Good addition. Expression inputs should be the default, anything else could be considered an ergnomic shortcut. So I think it's good to give 'preferential treatment' to expressions here.

Thanks for the speedy review!

If the polars community is interested, I might suggest a number of similar improvements. At the moment, I am mostly using the low-level Python bindings (in polars.polars), since the high-level Python wrappers are not performant enough for my particular use case (high-frequency, real-time feature building for use in ML inference). If I could remove the main performance issues, I would use the regular Python API and be happy.

[stinodego]

I'd be very interested in your input. Feel free to create an issue for this to start a discussion and track progress on similar improvements, or just send some more PRs our way!

[jakob-keller]

I'd be very interested in your input. Feel free to create an issue for this to start a discussion and track progress on similar improvements, or just send some more PRs our way!

There you go: #7019, #7020