Apply the default x category order per axis to facets, marginals and 3D

Address review on #5776: rather than excluding faceted and marginal
figures, each x axis now gets the whole-data-frame order restricted to
the categories carried by the traces placed on it, so every facet gets
the same order as its single-subplot counterpart and matched facet axes
agree on the relative positions of shared categories. The right-hand
marginal of marginal_y spans the y dimension and keeps no x order, and
px.ecdf is excluded because it sorts the x values inside each trace.
3D figures get the data order on the scene x axis. Adds tests for the
facet repro from review, per-panel categories, marginals, 3D and the
ecdf guard, and updates the changelog entry.

Author: vineoy

CHANGELOG.md

@@ -23,7 +23,7 @@ This project adheres to [Semantic Versioning](http://semver.org/).
 - Fix `mpl_to_plotly` conversion of matplotlib contour lines and line collections: close contour rings ending with `Z` codes, support dash styles, convert date x-axes, separate disjoint subpaths with `None` separators, hide line collection traces from the legend by default, and group consecutive same-style lines into single traces [[#5770](https://github.com/plotly/plotly.py/pull/5770)], with thanks to @robertoffmoura for the contribution!
 - Fix serialization of box/lasso select data for figures with multiple subplots [[#5710](https://github.com/plotly/plotly.py/pull/5710)], with thanks to @flpi-sun for the contribution!
 - Fix `plotly.io.write_images` ignoring `layout.width`/`layout.height` (and the template's) and exporting every figure at the default 700x500 size; it now uses the same width/height priority as `write_image` and `to_image` [[#5784](https://github.com/plotly/plotly.py/pull/5784)], with thanks to @breken-ai for the contribution!
-- Fix Plotly Express scrambling the order of the categorical x-axis in single-subplot figures when the data is split into several traces (e.g. by `color`); the x-axis now follows the order of first appearance in the underlying data, while faceted and marginal figures are unchanged [[#3198](https://github.com/plotly/plotly.py/issues/3198), [#3291](https://github.com/plotly/plotly.py/issues/3291)]
+- Fix Plotly Express scrambling the order of the categorical x-axis when the data is split into several traces (e.g. by `color`); the x-axis now follows the order of first appearance in the underlying data, on single-subplot figures, on each facet of faceted figures and in 3D figures, while marginal axes spanning the y dimension are left alone [[#3198](https://github.com/plotly/plotly.py/issues/3198), [#3291](https://github.com/plotly/plotly.py/issues/3291)]
 
 ## [7.1.0] - 2026-09-15
 

plotly/express/_core.py

@@ -2530,58 +2530,64 @@ def get_groups_and_orders(args, grouper):
     return groups, orders
 
 
-def _default_category_orders_for_axes(args, groups, orders, constructor):
+def _default_x_category_order(args, groups, orders, constructor):
     """
-    Add data-appearance order for the categorical x axis.
+    Return the data-appearance order of the categorical x column, or None.
 
     When the data is split into several traces (e.g. by `color`), plotly.js
     orders categories by first appearance across traces, which can differ from
     the order in the underlying data. Recording the data-appearance order here
-    lets `set_cartesian_axis_opts` emit an explicit `categoryarray`, matching
-    the documented behavior that Plotly Express lays out categorical data in
-    the order in which it appears in the data. See #3198.
-
-    The default is only emitted for figures with a single subplot, so the one
-    x axis it touches spans the whole data-frame:
-
-    - With facets each subplot only carries its own categories and the
-      documented independent-axes technique (disabling `matches`, see
-      `doc/python/facet-plots.md`) must keep listing each axis only its own
-      categories, which an explicit data-wide `categoryarray` would break. See
-      https://github.com/plotly/plotly.py/pull/5776#pullrequestreview-5701715902
-
-    - With marginals the extra subplots span the y dimension, so an explicit
-      x `categoryarray` on their axes would be wrong.
-
-    Figures with subplots could get per-axis data orders in a follow up.
+    lets the axis configurators emit an explicit `categoryarray`, matching the
+    documented behavior that Plotly Express lays out categorical data in the
+    order in which it appears in the data. See #3198.
+
+    The default only applies to string categoricals the user did not order
+    explicitly, to figures split into more than one trace, and to the x axis:
+
+    - Cartesian figures with facets or marginals apply the default per axis
+      in `make_figure`: each x axis gets the whole-data-frame order
+      restricted to the categories carried by the traces on its subplot,
+      exactly the order its single-subplot counterpart would use with the
+      categories absent from the subplot dropped. Shared categories keep the
+      same relative positions on matched facet axes, and the documented
+      independent-axes technique (disabling `matches`, see
+      `doc/python/facet-plots.md`) still lists each axis only its own
+      categories. The right-hand marginal of `marginal_y` spans the y
+      dimension and carries the binned y values, so it never receives the x
+      order.
+
+    - `px.ecdf` sorts the x values inside each trace, so its categorical
+      axis must keep the trace-derived order for the step function to read
+      in sequence.
+
+    - `px.scatter_3d` has a single scene, so the whole-data-frame order is
+      applied through the per-column mechanism in `configure_3d_axes`.
 
     Only the x axis is handled: y categoricals are intentionally left alone
     because `set_cartesian_axis_opts` reverses y orders for top-down display
     and figures such as `px.timeline` / horizontal `px.bar` pair that with
     `autorange="reversed"` (see `doc/python/gantt.md`), so an explicit y
     `categoryarray` would double-reverse and flip the documented order.
     """
-    if constructor not in cartesians:
-        return orders
+    if constructor not in cartesians and constructor is not go.Scatter3d:
+        return None
     if len(groups) <= 1:
-        return orders
-    if args.get("facet_row") or args.get("facet_col"):
-        return orders
-    if args.get("marginal_x") or args.get("marginal_y"):
-        return orders
+        return None
+    if args.get("ecdfmode") is not None:
+        return None
     df = args["data_frame"]
     col = args.get("x")
     if not isinstance(col, str) or col in orders or col not in df.columns:
-        return orders
+        return None
     uniques = df.get_column(col).unique(maintain_order=True).to_list()
     categories = [
         value
         for value in uniques
         if value is not None and not (isinstance(value, float) and math.isnan(value))
     ]
     if categories and all(isinstance(value, str) for value in categories):
-        orders[col] = categories
-    return orders
+        return categories
+    return None
 
 
 def make_figure(args, constructor, trace_patch=None, layout_patch=None):
@@ -2610,8 +2616,25 @@ def make_figure(args, constructor, trace_patch=None, layout_patch=None):
     )
     grouper = [x.grouper or one_group for x in grouped_mappings] or [one_group]
     groups, orders = get_groups_and_orders(args, grouper)
-    orders = _default_category_orders_for_axes(args, groups, orders, constructor)
-
+    default_x_categories = _default_x_category_order(args, groups, orders, constructor)
+    per_axis_order = (
+        default_x_categories is not None
+        and constructor in cartesians
+        and (
+            args.get("facet_row") is not None
+            or args.get("facet_col") is not None
+            or args.get("marginal_x") is not None
+            or args.get("marginal_y") is not None
+        )
+    )
+    if default_x_categories is not None and not per_axis_order:
+        # The figure has a single x axis spanning the whole data frame (a
+        # plain cartesian subplot or a 3D scene), so the per-column
+        # mechanism in `set_cartesian_axis_opts` / `configure_3d_axes`
+        # applies the data order directly.
+        orders[args["x"]] = default_x_categories
+
+    trace_axes_rows = []
     col_labels = []
     row_labels = []
     nrows = ncols = 1
@@ -2660,6 +2683,11 @@ def make_figure(args, constructor, trace_patch=None, layout_patch=None):
         for trace_spec in trace_specs:
             # Create the trace
             trace = trace_spec.constructor(name=trace_name)
+            if per_axis_order:
+                # Remember the rows feeding this trace so the data order can
+                # be computed per x axis once the traces are placed on their
+                # subplots.
+                trace_axes_rows.append((trace, group, trace_spec.marginal))
             if trace_spec.constructor not in [
                 go.Parcats,
                 go.Parcoords,
@@ -2908,6 +2936,27 @@ def make_figure(args, constructor, trace_patch=None, layout_patch=None):
         fig._px_trendlines = []
 
     configure_axes(args, constructor, fig, orders)
+    if per_axis_order:
+        # Each x axis gets the whole-data-frame order restricted to the
+        # categories carried by the traces placed on it: the same order its
+        # single-subplot counterpart would use, with the categories absent
+        # from the subplot dropped, so matched facet axes agree on the
+        # relative positions of shared categories. Marginal axes spanning
+        # the y dimension do not carry the x column and get no default.
+        axis_x_values = {}
+        for trace, group, marginal in trace_axes_rows:
+            xaxis_id = getattr(trace, "xaxis", None)
+            if xaxis_id is None or marginal == "y":
+                continue
+            axis_x_values.setdefault(xaxis_id, set()).update(
+                group.get_column(args["x"]).to_list()
+            )
+        for xaxis_id, values in axis_x_values.items():
+            categories = [value for value in default_x_categories if value in values]
+            if categories:
+                axis = fig.layout["xaxis" + xaxis_id[1:]]
+                axis["categoryorder"] = "array"
+                axis["categoryarray"] = categories
     configure_animation_controls(args, constructor, fig)
     return fig
 

tests/test_optional/test_px/test_px.py

@@ -431,35 +431,130 @@ def test_unlinked_facet_axes_only_list_their_own_categories():
         assert axis.categoryarray is None or set(axis.categoryarray) <= set(trace.x)
 
 
-def test_no_default_category_array_with_subplots():
-    # https://github.com/plotly/plotly.py/pull/5776#pullrequestreview-5701715902
-    # Facet axes list their own categories (the `matches=None` technique in
-    # doc/python/facet-plots.md), and marginal subplots span the y dimension,
-    # so no axis of a faceted or marginal figure may get a data-wide default.
+def test_default_category_order_applies_to_facets():
+    # https://github.com/plotly/plotly.py/pull/5776#discussion_r4429048272
+    # A facet must not change the x order relative to the same chart without
+    # facets: every facet axis follows the data order of its own rows, and
+    # the y axes stay untouched.
     df = pd.DataFrame(
         {
-            "panel": ["left", "right", "left", "right", "left"],
-            "fruit": ["apple", "banana", "cherry", "date", "elderberry"],
-            "sold": [5, 3, 4, 6, 2],
+            "month": ["Jan", "Feb", "Mar", "Apr", "Jan", "Feb", "Mar", "Apr"],
+            "product": [
+                "Basic",
+                "Basic",
+                "Pro",
+                "Basic",
+                "Basic",
+                "Basic",
+                "Pro",
+                "Basic",
+            ],
+            "region": ["East", "East", "East", "East", "West", "West", "West", "West"],
+            "units": [10, 12, 8, 15, 11, 13, 9, 16],
         }
     )
-    fig = px.bar(df, x="fruit", y="sold", facet_col="panel")
-    for axname in fig.layout:
-        if axname.startswith("xaxis"):
-            assert fig.layout[axname].categoryorder is None
-            assert fig.layout[axname].categoryarray is None
-    fig = px.scatter(
-        df,
-        x="fruit",
-        y="sold",
-        color="panel",
-        marginal_y="rug",
-        marginal_x="rug",
+    fig = px.bar(df, x="month", y="units", color="product")
+    assert fig["layout"]["xaxis"]["categoryorder"] == "array"
+    assert fig["layout"]["xaxis"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    fig = px.bar(df, x="month", y="units", color="product", facet_col="region")
+    assert fig["layout"]["xaxis"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    assert fig["layout"]["xaxis2"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    fig = px.bar(df, x="month", y="units", color="product", facet_row="region")
+    assert fig["layout"]["xaxis"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    assert fig["layout"]["xaxis2"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    assert fig["layout"]["yaxis"]["categoryorder"] is None
+    assert fig["layout"]["yaxis2"]["categoryorder"] is None
+    # numeric axes never get a category array, faceted or not
+    fig = px.bar(df, x="units", y="units", color="product", facet_col="region")
+    assert fig["layout"]["xaxis"]["categoryorder"] is None
+    assert fig["layout"]["xaxis2"]["categoryorder"] is None
+
+
+def test_facet_axes_list_their_own_categories_in_data_order():
+    df = pd.DataFrame(
+        {
+            "region": ["E", "E", "W", "E", "W"],
+            "cat": ["a", "b", "a", "c", "b"],
+            "kind": ["p", "q", "q", "q", "p"],
+            "val": [1, 2, 3, 4, 5],
+        }
     )
-    for axname in fig.layout:
-        if axname.startswith("xaxis"):
-            assert fig.layout[axname].categoryorder is None
-            assert fig.layout[axname].categoryarray is None
+    fig = px.bar(df, x="cat", y="val", color="kind", facet_col="region")
+    # each facet lists only its own categories, in the order of the whole
+    # data frame: the trace-derived order scrambles the second facet because
+    # its "kind=p" trace carries "b" and is created before its "kind=q" trace
+    assert fig["layout"]["xaxis"]["categoryarray"] == ("a", "b", "c")
+    assert fig["layout"]["xaxis2"]["categoryarray"] == ("a", "b")
+
+
+def test_default_category_order_with_marginals():
+    # The top marginal of `marginal_x` shares the x dimension with the main
+    # subplot and gets the same data order. The right marginal of
+    # `marginal_y` spans the y dimension and carries the binned y values, so
+    # it must not receive the x order.
+    df = pd.DataFrame(
+        {
+            "month": ["Jan", "Feb", "Mar", "Apr", "Jan", "Feb", "Mar", "Apr"],
+            "product": [
+                "Basic",
+                "Basic",
+                "Pro",
+                "Basic",
+                "Basic",
+                "Basic",
+                "Pro",
+                "Basic",
+            ],
+            "units": [10, 12, 8, 15, 11, 13, 9, 16],
+        }
+    )
+    fig = px.scatter(df, x="month", y="units", color="product", marginal_x="rug")
+    assert fig["layout"]["xaxis"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    assert fig["layout"]["xaxis2"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    fig = px.scatter(df, x="month", y="units", color="product", marginal_y="rug")
+    assert fig["layout"]["xaxis"]["categoryarray"] == ("Jan", "Feb", "Mar", "Apr")
+    assert fig["layout"]["xaxis2"]["categoryarray"] is None
+
+
+def test_default_category_order_for_3d():
+    df = pd.DataFrame(
+        {
+            "month": ["Jan", "Feb", "Mar", "Apr", "Jan", "Feb", "Mar", "Apr"],
+            "product": [
+                "Basic",
+                "Basic",
+                "Pro",
+                "Basic",
+                "Basic",
+                "Basic",
+                "Pro",
+                "Basic",
+            ],
+            "units": [10, 12, 8, 15, 11, 13, 9, 16],
+        }
+    )
+    fig = px.scatter_3d(df, x="month", y="units", z="units", color="product")
+    assert fig["layout"]["scene"]["xaxis"]["categoryorder"] == "array"
+    assert fig["layout"]["scene"]["xaxis"]["categoryarray"] == (
+        "Jan",
+        "Feb",
+        "Mar",
+        "Apr",
+    )
+    assert fig["layout"]["scene"]["yaxis"]["categoryorder"] is None
+    assert fig["layout"]["scene"]["zaxis"]["categoryorder"] is None
+
+
+def test_no_default_category_array_for_ecdf():
+    # px.ecdf sorts the x values inside each trace, so its categorical axis
+    # must keep the trace-derived order for the step function to read in
+    # sequence.
+    df = pd.DataFrame(
+        {"x": ["b", "a", "c", "a", "b", "c"], "g": ["p", "p", "p", "q", "q", "q"]}
+    )
+    fig = px.ecdf(df, x="x", color="g")
+    assert fig["layout"]["xaxis"]["categoryorder"] is None
+    assert fig["layout"]["xaxis"]["categoryarray"] is None
 
 
 def test_permissive_defaults():