By default no layout mechanism is active, which means Axes are placed in their initial position and not moved afterwards.

You either have to set the suplot spec anew ax2.set_subplotspec(gs[0, 1]), or you can use a layout manager to control Axes positioning in the background (plt.figure(layout="constrained") or plt.figure(layout="tight")).

Overall, I don't think we can do anything here as long as there is no default layout mechanism in place.

Is this issue still open?

IMHO there is nothing to do here. Just waiting for reply before closing.

IMHO there is nothing to do here. Just waiting for reply before closing.

Thanks Tim for letting me know.

Thanks @timhoffm That clears things up.

The issue isn't really about the layout engine. set_width_ratios() is a setter that updates internal state but has no visible effect, with no warning, no error, and no mention in the docstring. From a user's perspective, calling a setter and seeing nothing change is unexpected behavior regardless of the layout engine.

From a user's perspective, calling a setter and seeing nothing change is unexpected behavior regardless of the layout engine.

Yes this is not ideal, but it's a setter on another object, so the expectation is not clear. fig.add_subplot(gs[0, 0]) could be regarded as "create the sublot at the given position" - This by itself does not indicate whether the Axes maintains a connection to the gridspec (It does technically, internally, but that is not used unless you do a repositioning through layout). So the logical default behavior is that the gridspec is just a parameter evalulated at creation time.

Technical: By default, there is no mechanism to react later changes. GridSpec/SubplotSpec are pure data objects. They don't know about Axes that use them, so can't push any changes. The only way to update is for Axes to re-evalutate their position. And that's the purpose of the layout mechanism.

So this is not ideal but bearable. You could write a minimal layout engine that updates the Axes from the specs, and make it the default. However, that's a specific solution for a very particular problem. On a larger, scale I'd rather discuss whether we can have a reasonable layout manager activated by default. We haven't done this for backward-compatibility (exact positioning reproducibility), but for most use cases having a default layout manger would be the sane and user-friendly choice nowadays.

a setter that updates internal state but has no visible effect, with no warning, no error, and no mention in the docstring

I guess there is no harm in updating the docstrings to clarify that these methods won't by themselves update existing axes.

I guess there is no harm in updating the docstrings to clarify that these methods won't by themselves update existing axes.

I agree with you @rcomer