How do I ensure that a Polars expression plugin properly uses multiple CPUs?

I'm writing a polars plugin that works, but never seems to use more than one CPU. The plugin's function is element-wise, and is marked as such in register_plugin_function. What might I need to do to get this plugin to properly use parallelism?

It's a bit of a complex plugin, and I can't share the code. Some relevant information:

• The plugin function acts on a String column, and takes a string argument. Operations should be very similar speed on any inputs, so there shouldn't be a thread imbalance. There's no file I/O in the function.
• The plugin function is element-wise, and is marked is_elementwise=True in the register_plugin_function call.
• pl.threadpool_size() (Python-side) gives the right number, as does printing the pool size on the Rust side with polars_core::POOL.
• I know the plugin isn't using all CPUs because I can see that a long call on a 32-core machine uses about 3% of total processing power. I've line-profiled the Python code and verified that the vast majority (80%+) of the code's time is being spent on one line that calls the plugin function.
• I'm using Polars 0.46.0 on the Rust side, 1.29.0 on Python side.

Any thoughts?

Thanks to @jqurious for getting me to an answer!

I was able to get the plugin running in parallel by forcing it through a LazyFrame and collecting with .collect(engine="streaming"). Instead of doing

df = df.with_columns(my_plugin(colname, arg))

I did

df = df.lazy().with_columns(my_plugin(colname, arg)).collect(engine="streaming")

and this worked as expected, giving me a ~30x speedup on a 32-core machine. I'm not sure if this is the way Polars intends plugins to work, but it did work.