I'd love to see a test that fails before these changes! I'm a little concerned that the existing ones didn't catch this

Hey @marthacryan , thanks for getting back to me. Should the example in the issue enough as a test?

Here's a reproducing example that came out of the skimage CI:

import numpy as np

import plotly
import plotly.express as px
import plotly.graph_objects as go

plotly.io.renderers.default = "sphinx_gallery_png"

img = np.zeros((50, 50))
fig = px.imshow(img, binary_string=True)

fig.add_trace(
    go.Scatter(
        x=np.array([[1,2,3],[4,5,6]]),
        y=np.array([[1,2,3],[4,5,6]]),
        name="1"
    )
)

plotly.io.show(fig)

Which results in:

ValueError:
    Invalid value of type 'builtins.dict' received for the 'x' property of scatter
        Received value: {'dtype': 'i1', 'bdata': 'AQIDBAUG', 'shape': '2, 3'}

    The 'x' property is an array that may be specified as a tuple,
    list, numpy array, or pandas Series

If you toggle binary_string to False, it brings up another unexpected error:

ValueError:
    Invalid value of type 'builtins.dict' received for the 'z' property of heatmap
        Received value: {'dtype': 'f8', 'bdata': 'AAAAAAAAAAAAAAAAAA...AAA=', 'shape': '50, 50'}

    The 'z' property is an array that may be specified as a tuple,
    list, numpy array, or pandas Series

This is plotly 6.0.0rc0 and kaleido 1.0.0rc0.

I am not particularly confident with this part of the code. Minimal example failing in master:

import copy
import plotly.express as px

gapminder = px.data.gapminder()
fig = px.line(gapminder, x="year", y="gdpPercap", color="country")
fig_copy = copy.deepcopy(fig)

The entire error trace gives some more insights, here what I could figure out (maybe it is trivial, yet it wasn't for me)

• deepcopy calls BaseFigure.__reduce__, which calls BaseFigure.to_dict, which calls convert_to_base64.
• At this point every value is converted to the following form:

  'x' = {'dtype': 'i2', 'bdata': 'oAelB6oHrwe0B7kHvgfDB8gHzQfSB9cH'}
  'y' = {'dtype': 'f8', 'bdata': 'ESBU8...7M'}

• Then deepcopy will call _reconstruct, which tries to run Figure(*converted_values), hence BaseFigure(*converted_values). Finally here validate_coerce is hit and it will fail.

Hope it helps

Ah yeah so I can give a little context here. The approach we took for sending base64 encoded arrays was to add the encoding step to the to_json function. Our thinking here was that this would be basically the last step before sending data to the frontend, so it would have minimal effects in the rest of the codebase. The performance improvements from base64 encoding arrays are mostly in the step of sending data from python to the client. However, we didn't expect that the base64 encoded arrays (which are also called typed array specs in parts of the code) would be used as input for the Figure class. That's why we're skipping the array validation if we encounter this datatype.

I'm not sure why this isn't failing for other deepcopy tests, but one guess I have is that they aren't numpy/pandas arrays? The base64 encoding only happens for numpy/pandas.