Default styling options on ScatterplotLayer

[zacdezgeo]

Hey @kylebarron! I'm not sure if this is the right spot for this discussion. Please point me to the right path as needed.

So, I've been using lonboard with a hexagon dataset. I wanted to do more advanced styling, but it took me a moment to debug how to use the Scatterplotlayer. I thought I was encountering some issue because the points were not showing up on the map:

But the viz() function was rendering my gdf without any issues:

Turns out that the default radius was set to None:

It seems viz() has some defaults where it sets the radius and the random color, while the ScatterplotLayer does not.
Is this the expected behavior? There may be some reason I'm unaware of, but I would've expected the ScatterplotLayer to have the same default settings as the viz() function.

[kylebarron]

The question of what to use for defaults is indeed tricky.

Upstream, deck.gl has default values for virtually all parameters. So for example, if you look at the JS ScatterplotLayer docs, you'll see all the render options and data accessors. And the default for get_radius is 1. That, combined with the default for radiusUnits being "meters" means that by default, deck.gl renders a ScatterplotLayer with points that are 1 meter wide. Of course, this default means that points are almost never visible without the user passing explicit parameters. (Note that in Lonboard None does not mean "invisible". Rather it means "defer to deck.gl's default value in JavaScript")

This also speaks to differences between the audience of the JS library and the Python library. Users of the JS library are usually building an application, where they use a single dataset or class of data, and spend a lot of time making it look good. Lonboard, on the contrary, is more often used for exploratory data analysis, where you might have lots of different datasets you want to inspect, and where just seeing your data is more important than exactly how it looks.

This is the rationale between having lonboard.viz and lonboard.Layer have different defaults: use viz for exploratory data analysis where you just want to see your data, and use the latter for customizing exactly what you want to see.

Maybe Lonboard's layers should have defaults? But choosing a good default is really hard, and choosing when to apply those defaults is even harder. If a user passes in get_radius, should we rescind any defaults for radius_min_pixels and radius_max_pixels?

[zacdezgeo]

I'm not sure whether the minimum or maximum radius should be used; either would probably work. Still, for users onboarding a lonboard, seeing their data without immediately diving into the intricacies of styling would make a difference. The benefit to user onboarding outweighs the tradeoff of potentially displaying the data in a style that is not quite proper.

[kylebarron]

But the question of when to apply defaults is still really hard. If you pass

ScatterplotLayer.from_geopandas(gdf, get_radius=...)

should any defaults be set? Because the default unit for radius is meters, so if you pass get_radius=1, you won't see anything on the map. But alternatively, if we set a default of radius_min_pixels=5 or something, then the user would have to turn off the default when they actually want small points.

[kylebarron]

Maybe this is well-served by having a default_styling: bool parameter, where the user can turn on/off whether defaults should be applied. But then the question of: what should be the default value of default_styling?

[zacdezgeo]

Not sure there's a right answer. IMHO I feel that a user who wants to customize their styling will be less burdened by setting these parameters than users onboarding and not seeing their points.

[kylebarron]

In #620 we added an add_layer method to Map, so now it's easier to add another point layer with default styling settings with map.add_layer(viz(gdf_with_point_data)). This makes me more inclined to point people who want default styling settings to use viz, and potentially change the styling attributes after the fact:

m = viz(point_data)
layer = m.layers[0]
layer.get_radius = 5