PERF: load plotting entrypoint only when necessary

[TLouf]

• closes ENH: load plotting entry point only when necessary #41492
• tests added / passed (I only ran plotting/test_backend.py, which seemed appropriate)
• Ensure all linting tests pass, see here for how to run them
• whatsnew entry

[jreback]

is there some way to actually test this?

[TLouf]

is there some way to actually test this?

A first thing I see in test_backend is that test_setting_backend_without_plot_raises does not test _get_plot_backend directly but via set_option, and does not test with an entry point without a .plot method, only for unregistered modules, because it wasn't checked in the code either.

A second is to test whether _get_plot_backend("matplotlib") returns pandas.plotting._matplotlib.

[jreback]

how do we know that this is better perf wise? we don't have any registered backends in pandas tests (I think). I am very leary of changing this as last time this broke things for a while as its pretty fragile. If we have some testsing & perf benchmarks would feel better here.

[TLouf]

I've added a benchmark that runs _get_plot_backend with many (1000) dummy backends added in the setup. That's the way I found to have some loading time, another would be to add a dummy but slow to load backend, which I don't know how to make.

Here's the output from asv continuous -f 1.1 master entry-pt-lazy-load -b plotting.BackendLoading:

       before           after         ratio
     [b2a36bdf]       [76a912ec]
     <master>         <entry-pt-lazy-load>
-      2.20±0.6μs        159±0.5ns     0.07  plotting.BackendLoading.time_get_plot_backend`

I'm surprised by how low the measured times are, when I run the same thing outside of asv I get times of the order of ~100ms. I'm very new to benchmarking so let me know if you can explain this, especially if it's because I did something wrong.

Edit: had already tried setting repeat and number to 1 and saw no change, but I wasn't aware of warmup_time! Here are the new results in 6105c6f, which make much more sense:

       before           after         ratio
     [b2a36bdf]       [47130fad]
     <master>         <entry-pt-lazy-load>
-       232±0.2ms       4.00±0.3ms     0.02  plotting.BackendLoading.time_get_plot_backend

[jreback]

@TLouf 1000 is not realistic, let's see 10

[TLouf]

It's not realistic in terms of number of backends, but these are also extremely fast to load, as opposed to some real plotting backends. As I mentioned in the issue, hvPlot takes 2s to load on my machine, and in the current implementation, if it is present in the environment, it will be loaded, even if the user plots with matplotlib. So the performance improvement highly depends on what backends have been installed.

Anyway, here's the result with 10:

       before           after         ratio
     [b2a36bdf]       [47130fad]
     <master>         <entry-pt-lazy-load>
-       10.9±0.1ms       3.29±0.5ms     0.30  plotting.BackendLoading.time_get_plot_backend

[jreback]

a few comments, pls rebase as well

[TLouf]

Done. I had to add a boolean found_backend in _load_backend to make mypy happy, because initialising module to None was incompatible with the return type, which is necessarily a module, as the function raises when no backend is found.

[jreback]

what i meant is that you can move all of this code here to L1766 (e.g. just load and import there) right?

[TLouf]

I don't think so, I first have to iterate through all entry points before falling back to importing a module. That's because for instance, the hvPlot backend is an entry point whose name is "holoviews" (see https://github.com/holoviz/hvplot/blob/master/setup.py#L134). So if I move up the unregistered entry point approach, in that case I might import holoviews, which has no top level plot method, thus raising when the entry point is actually available.

[jreback]

thanks @TLouf