I do not understand this. It works on my local machine (although I get an image comparison error). Trying some variants...

Edit: Seems OK now.

I share @timhoffm 's concerns.

It may be worth looking at the code that extracts the properties to mirror and see if we can stash some of the style on the container and retrieve it from there.

I haven't had time to look into it in detail yet, but I've been thinking about it and I agree that it probably is better to add the properties to the container in case it is empty.

Please check the following: I think the container does not matter for the legend, what counts is the objects that are added to the Axes (e.g. self.add_patch(r)). One can still create the Artists for an empty dataset just don't add it to the returned container. - Basically what you're doing already but returning an empty container. I don't think we have to somehow artificially add properties to container.

We do walk the container list (that is how error bars get in), I may be miss-remembering the details. If Tim is right about how we discover the patches for the legend, lets not teach more things about the containers.

Just remembered this one.

It seems like creating the rectangle but not adding it doesn't help. I've pushed the latest version, but I guess the added test will fail now.

flake8:

./lib/matplotlib/axes/_axes.py:2460:80: E501 line too long (82 > 79 characters)

Was this working after changing to an empty list in BarContainer? I tried rebasing, but it wasn't working AFAICT.

The legend actually uses a custom handler for BarContainer:

So my suggestion (on top of this PR) would be something like (plus/minus decisions like naming or putting it in __init__, etc.):

diff --git a/lib/matplotlib/axes/_axes.py b/lib/matplotlib/axes/_axes.py
index 47a76e1f40..ff17ce5443 100644
--- a/lib/matplotlib/axes/_axes.py
+++ b/lib/matplotlib/axes/_axes.py
@@ -2598,8 +2598,9 @@ class Axes(_AxesBase):
                 r.sticky_edges.y.append(b)
             else:  # horizontal
                 r.sticky_edges.x.append(l)
-            self.add_patch(r)
-            patches.append(r)
+            if not data_empty:
+                self.add_patch(r)
+                patches.append(r)
 
         if xerr is not None or yerr is not None:
             if orientation == 'vertical':
@@ -2626,9 +2627,11 @@ class Axes(_AxesBase):
         else:  # horizontal
             datavalues = width
 
         bar_container = BarContainer(patches, errorbar, datavalues=datavalues,
                                      orientation=orientation,
                                      label=bar_container_label)
+        if data_empty:
+            bar_container._style_child = r
         self.add_container(bar_container)
 
         if tick_labels is not None:
diff --git a/lib/matplotlib/legend_handler.py b/lib/matplotlib/legend_handler.py
index 263945b050..84fdbeec58 100644
--- a/lib/matplotlib/legend_handler.py
+++ b/lib/matplotlib/legend_handler.py
@@ -41,6 +41,8 @@ def update_from_first_child(tgt, src):
     first_child = next(iter(src.get_children()), None)
     if first_child is not None:
         tgt.update_from(first_child)
+    elif hasattr(src, '_style_child'):
+        tgt.update_from(src._style_child)
 
 
 class HandlerBase:

Was this working after changing to an empty list in BarContainer? I tried rebasing, but it wasn't working AFAICT.

I do not recall. I think it did, at least in the sense that it produced the plot above. Trying to think back, I think at least it advances the color cycle? But something else might have changed?

But this was one of my first PRs to Matplotlib and apparently it wasn't the right approach then, so it's been in the back of my head to one day try to do it the right way.