[BUG] Manual-zoom resampling snaps back on dash ≥ 4.2.0 when a data Patch must clip an x-referenced layout shape (dash PR #3785)`

[AlexisJanin]

Describe the bug 🖍️

With dash >= 4.2.0 (incl. 4.3.0), a FigureResampler in a Dash app no longer resamples on manual zoom: the axes snap straight back to the original range immediately after dragging a zoom box. The same app works on dash == 4.1.0.

I bisected it to a narrow, specific trigger:

The snap-back happens only when the new zoom range must clip an
x-axis-referenced layout.shape (a shape whose x-extent is not fully
contained in the zoomed window). If the zoom fully contains every shape, the
zoom holds and resampling works — even on dash 4.2+.

This points at dash PR #3785 ("Fix patch with dcc.Graph figure", first released in 4.2.0), which deep-clones and re-syncs layout (including shape coordinates) whenever a Patch is applied to a dcc.Graph figure. plotly-resampler's zoom path returns a data-only Patch (construct_update_data_patch → assigns only data[i].x / y / name); the re-sync it now triggers rebuilds layout from the (zoom-less) figure prop and discards the user's live drag-zoom.

Reproducing the bug 🔍

Minimal trigger (bisected — all four required; dropping any one makes dash 4.2
keep the zoom):

1. a FigureResampler whose zoom callback returns a data-only Patch
   (construct_update_data_patch);
2. the dcc.Graph is injected by a callback, not present in app.layout at
   startup (a statically-declared Graph keeps its zoom on dash 4.2);
3. one x-axis-referenced layout shape that spans an x-range (a rect with
   xref="x", x0 != x1). A zero-width vertical line (x0 == x1) is not
   enough;
4. dash >= 4.2.0.

Things that turned out not to matter (removed, bug persists): make_subplots
/ multiple axes, matched x-axes (matches=), annotations, hovermode,
pattern-matching callback ids, allow_duplicate, and any server-side resampler
cache. A single trace on a single axis is enough.

It fires only when the zoom clips the shape — rect spans x ∈ [0.60·n, 0.70·n]:

Zoom box vs. rect [0.60, 0.70]	Result
entirely left of rect (e.g. 0–30%)	snaps back
entirely right of rect (e.g. 80–100%)	snaps back
straddles rect's left edge (50–65%)	snaps back
straddles rect's right edge (65–80%)	snaps back
strictly inside rect (62–68%)	snaps back
fully contains rect (e.g. 50–80%)	holds — no clip, no bug

Steps:

python3 -m venv .venv && .venv/bin/pip install "dash==4.2.0" "plotly==5.24.1" \
    "plotly-resampler==0.11.0" numpy pandas pytz
.venv/bin/python mwe.py     # http://127.0.0.1:8050

Click "Load plot", then drag a zoom box on the left half of the plot (so
the salmon rectangle stays off-screen — first row of the table). On
dash==4.2.0 the axes snap back; pin dash==4.1.0 instead and the zoom holds
and resamples. (Same plotly / plotly-resampler in both → dash is the only
variable.)

mwe.py (~80 lines, self-contained)

import numpy as np
import plotly.graph_objects as go
from dash import Dash, Input, Output, callback, dcc, html, no_update
from dash.exceptions import PreventUpdate
from plotly_resampler import FigureResampler

USE_UIREVISION_WORKAROUND = False  # set True to test the layout.uirevision fix

n = 1_500_000
x = np.arange(n)
y = np.sin(x / 300) + np.random.default_rng(0).standard_normal(n) / 5


def build_figure() -> FigureResampler:
    base = go.Figure()
    base.add_trace(go.Scattergl(name="signal", x=x, y=y))
    # REQUIRED: one x-referenced shape spanning an x-range. Removing it (or making
    # it a zero-width line) makes dash 4.2 KEEP the zoom -- this x-extent is what
    # PR #3785 re-syncs against the axis range and discards the live zoom from.
    base.add_shape(type="rect", x0=n * 0.6, x1=n * 0.7, xref="x",
                   y0=0, y1=1, yref="y domain",
                   fillcolor="LightSalmon", opacity=0.3, line_width=0)
    fig = FigureResampler(base, default_n_shown_samples=2_000)
    if USE_UIREVISION_WORKAROUND:
        fig.update_layout(uirevision="constant-token")
    return fig


fig = build_figure()

# suppress_callback_exceptions: the graph id only exists after the load callback fires.
app = Dash(__name__, suppress_callback_exceptions=True)
app.layout = html.Div(
    [
        html.H3("plotly-resampler manual-zoom repro"),
        html.P("Click 'Load plot', then drag a zoom box on the LEFT half (so the "
               "salmon rectangle stays off-screen). On dash 4.1.0 the zoom holds "
               "and resamples; on dash 4.2.0 it snaps back."),
        html.Button("Load plot", id="load-btn"),
        html.Div(id="container"),  # empty at startup -- Graph injected below
    ]
)


@callback(
    Output("container", "children"),
    Input("load-btn", "n_clicks"),
    prevent_initial_call=True,
)
def load_plot(_n_clicks):
    return dcc.Graph(id="graph", figure=fig, config={"displayModeBar": True},
                     style={"height": "800px"})


@callback(
    Output("graph", "figure"),
    Input("graph", "relayoutData"),
    prevent_initial_call=True,
)
def resample_on_zoom(relayout):
    if not relayout:
        raise PreventUpdate
    patch = fig.construct_update_data_patch(relayout)
    if patch is no_update:
        raise PreventUpdate
    return patch


if __name__ == "__main__":
    app.run(debug=True, port=8050)

Expected behavior 🔧

Dragging a zoom box should keep the zoomed view and resample the trace to higher resolution within that window — exactly as it does on dash == 4.1.0 — instead of the axes reverting to the full range. The presence of an x-referenced layout.shape should not cause the live zoom to be discarded when a data-only Patch is applied.

Screenshots 📸

Environment information: (please complete the following information)

• OS: macOS (also reproducible on Linux)
• Python environment:
  • Python version: 3.12
  • plotly-resampler environment: Dash web app (Safari / Firefox)
• plotly-resampler version: 0.11.0
• dash version: broken on 4.2.0 and 4.3.0, works on 4.1.0 (only variable changed)
• plotly version: 5.24.1 (held constant across broken/working — also seen on 6.x)

Additional context

Root cause (bisected). dash PR #3785 — "Fix patch with dcc.Graph figure", first released in dash 4.2.0 (plotly/dash#3785) — makes dcc.Graph deep-clone and re-sync layout whenever a Patch is applied to its figure (reviewer notes: "Clone layout to avoid mutating the original (important for Patch)" and "Sync shapes from gd.layout to figure ... because getLayout() clones layout"). The Python-side Patch is byte-identical across dash 4.1 / 4.2 / 4.3 (same Assign ops on data[0].x/y/name), so the regression is entirely in dash's client-side patch application, not in plotly-resampler's output. The clipping observation above pinpoints the shape-coordinate sync path as where the live zoom is lost.

Workaround. Setting a constant per-figure layout.uirevision makes plotly.js preserve the drag-zoom across the data Patch, restoring manual-zoom resampling on dash 4.2+ (USE_UIREVISION_WORKAROUND = True in mwe.py). However it then runs into existing issue #252 ("Figures updated with layout.uirevision do not resample"): a programmatic figure update while zoomed preserves the view but fires no relayout, so the visible window isn't re-resampled.

Open questions
(1) Should plotly-resampler set uirevision
internally on dash >= 4.2.0 and address the #252 interaction so a programmatic
update while zoomed re-triggers resampling?
(2) Or is this squarely a dash
regression to escalate against PR #3785, with plotly-resampler documenting a
dash < 4.2 cap meanwhile?

Related (checked — not duplicates).

• [BUG] Figures updated with layout.uirevision do not resample #252 — uirevision ↔ resample (the workaround's side effect, not this regression).
• [BUG] Resample does not dynamically update on zoom with subplots #322 — resample-on-zoom with subplots (different root cause: _grid_ref/xaxis mapping).

Links.

• dash PR #3785 (root cause): Fix graph patch & duplicate clicks plotly/dash#3785
• Related issue [BUG] Figures updated with layout.uirevision do not resample #252: [BUG] Figures updated with layout.uirevision do not resample #252

Happy to help further — I have a ready-to-run repro repo with two pinned venvs
(dash==4.1.0 vs 4.2.0, everything else identical) and can record a short
screencast or test patches against my setup. Just ping me here.