'rounding' of x coordinates in plt.plot with large 64-bit numbers

[tunnell]

I've having a problem with 64 bit integers and pyplot.plot (or scatter). Each x value is unique, but they are plotted such that the (x1, y1) is directly above the (x0, x1). It seems to combine every 10 or so x coordinates into one or is doing some sort of rounding. Has anybody seen this before?

It looks like it is binning in x?

import matplotlib.pyplot as plt
import numpy as np

x = np.arange(136897843038780956, 136897843038780956 + 1000, dtype=np.int64)
#x = np.arange(0,1000, dtype=np.int64) # works                                                                                                                     
y = np.arange(0,1000, dtype=np.int64)
plt.plot(x, y)
plt.show()

[tunnell]

P.S. 136897843038780956 is representable by 64 bits... so it's not that.

[tunnell]

Update: is it casting to float?

[WeatherGod]

That is quite possible

[tacaswell]

Interesting.....

I have tagged this as 1.4.x as I suspect that is is something subtle/nasty to fix.

Naive guess is that there is a loss of precision (via a cast to a float) some place in the transform stack.

[efiring]

All plotting is done using floating point. This is not going to change. You need to subtract a sufficiently large offset manually, while you are still dealing with integers, before passing your array into mpl. This is not a bug, it is just an inherent limitation based on the way mpl works. I can't imagine that we would rework every nook and cranny of mpl to work with 64-bit integers.

[efiring]

Unfortunately, I think it would be very cumbersome to try to trap such cases and issue a warning. The case you describe is quite unusual. Notice that tick labeling also would require special care; I really think that recognizing and handling this type of situation has to be the responsibility of the user, who knows what the inputs are and what the desired plot should look like. The software can't anticipate and automatically deal with every possible extreme case.

[tunnell]

Thanks for feedback. I agree that for most cases, the user should know: it makes perfect sense if somebody starts thinking about how it would need to be coded on your end. However, at least in scientific settings, I think it's a somewhat less extreme case. People use matplotlib and numpy for 'data exploration' (i.e. making plots to understanding aspects of the data) that can involve 64 bit numbers since, for example, there are instruments with that level of information. However, now that this issue exists in the issue tracker, at least people can find this edge case. Thanks again for the great tool.