Fix timeline bar lengths across daylight saving changes

I calculate timeline bar lengths in the same local wall-clock coordinates used by date axes, so timezone-aware endpoints remain aligned across daylight saving changes. I leave the original start column unchanged and cover spring, autumn, UTC, and naive timestamps across dataframe backends.

Fixes #4611

Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>

Author: mottopanikeiku

CHANGELOG.md

@@ -10,6 +10,7 @@ This project adheres to [Semantic Versioning](http://semver.org/).
 ### Fixed
 - Fix `mpl_to_plotly` tick marker mirroring, axis positioning, and trace visibility for twinned and overlaid subplots (such as `twinx` and `twiny`) [[#5310](https://github.com/plotly/plotly.py/pull/5310)], with thanks to @robertoffmoura for the contribution!
 - Fix `mpl_to_plotly` losing matplotlib axis line colors by exporting the color of each axis's displayed spine to the plotly x and y axis `linecolor` [[#5311](https://github.com/plotly/plotly.py/pull/5311)], with thanks to @robertoffmoura for the contribution!
+- Fix `px.timeline` bars ending an hour early or late when timezone-aware dates span a daylight saving time change [[#4611](https://github.com/plotly/plotly.py/issues/4611)]
 - Fix concurrent first access to lazily initialized graph object properties, which could raise `ValueError("Invalid value")` [[#5691](https://github.com/plotly/plotly.py/pull/5691)], with thanks to @hb1915 for the contribution!
 - Fix `mpl_to_plotly` crashing on touching bars (such as `plt.hist`) due to floating-point noise producing negative `bargap` values by clamping `bargap` to `[0, 1]` [[#5696](https://github.com/plotly/plotly.py/pull/5696)], with thanks to @robertoffmoura for the contribution!
 - Fix `mpl_to_plotly` drawing step plots with diagonal lines instead of vertical/horizontal step segments by mapping matplotlib step drawstyles (`steps-pre`, `steps-post`, `steps-mid`) to plotly line shapes [[#5697](https://github.com/plotly/plotly.py/pull/5697)], with thanks to @robertoffmoura for the contribution!

plotly/express/_core.py

@@ -2194,11 +2194,18 @@ def process_dataframe_timeline(args):
                 "Both x_start and x_end must refer to data convertible to datetimes."
             ) from exc
 
+    # Match the local times displayed on date axes when calculating bar lengths.
+    schema = df.schema
+    x_start, x_end = [
+        nw.col(col).dt.replace_time_zone(None)
+        if schema[col] == nw.Datetime and schema[col].time_zone is not None
+        else nw.col(col)
+        for col in (args["x_start"], args["x_end"])
+    ]
+
     # note that we are not adding any columns to the data frame here, so no risk of overwrite
     args["data_frame"] = df.with_columns(
-        (nw.col(args["x_end"]) - nw.col(args["x_start"]))
-        .dt.total_milliseconds()
-        .alias(args["x_end"])
+        (x_end - x_start).dt.total_milliseconds().alias(args["x_end"])
     )
     args["x"] = args["x_end"]
     args["base"] = args["x_start"]

tests/test_optional/test_px/test_px_functions.py

@@ -680,6 +680,32 @@ def test_timeline_cols_already_temporal(constructor, datetime_columns):
     assert fig.layout.xaxis.title.text is None
 
 
+@pytest.mark.parametrize("time_zone", [None, "UTC", "US/Pacific"])
+def test_timeline_daylight_saving(constructor, time_zone):
+    # Regression for https://github.com/plotly/plotly.py/issues/4611
+    starts = ["2024-03-09", "2024-03-10", "2024-03-11", "2023-11-05"]
+    finishes = ["2024-03-10", "2024-03-11", "2024-03-12", "2023-11-06"]
+    df = nw.from_native(
+        constructor({"Start": starts, "Finish": finishes, "Task": ["Job A"] * 4})
+    ).with_columns(
+        nw.col("Start", "Finish")
+        .str.to_datetime(format="%Y-%m-%d")
+        .dt.replace_time_zone(time_zone)
+    )
+
+    fig = px.timeline(df.to_native(), x_start="Start", x_end="Finish", y="Task")
+
+    # Date axes use local wall-clock coordinates: consecutive midnights are
+    # 24 hours apart even when the elapsed time is 23 or 25 hours.
+    assert_array_equal(fig.data[0].x, [24 * 60 * 60 * 1000] * 4)
+    base = np.asarray(fig.data[0].base, dtype="datetime64[ms]")
+    assert_array_equal(base, np.array(starts, dtype="datetime64[ms]"))
+    assert_array_equal(
+        base + fig.data[0].x.astype("timedelta64[ms]"),
+        np.array(finishes, dtype="datetime64[ms]"),
+    )
+
+
 def test_empty_histogram():
     """Empty px.histogram() should not raise, matching scatter/bar/pie behavior.