There are simply too many bars here for the resolution of your plot. each bar is less than one pixel wide, so some bars are just missing being seen in any pixel.

By default we produce a 640x480 100 dpi image. Even if your figure had no border, each bar would be less than half a pixel wide. You can improve things by increasing the dpi (or fig size... basically just need more pixels. However, you will still see effects of the discritization even up to quite high dpis (at 1000 dpi, obviously I didnt count to see whether or not bars are missing, though I think they are all present since each bar is 2-3 pixels, but there are still some bars that are directly adjacent and some that have a 1 pixel gap between them)

One solution would be to use a vector format (e.g. svg, pdf), but that only kicks the can down the road, eventually to display it needs to be rasterized. some vector viewers may incorporate antialiasing to hide the gaps by displaying averaged pixels (we do similar things for some cases such as images, but not here), but if you are viewing something with fewer pixels allocated to it than data points, some compromises will be made.

Basically the question becomes "is a bar plot actually an effective way of communicating the information?" and I'd argue that in general if you have 1000s of points, some amount of data reduction is basically necessary. That could be things like box or violin for a single variable as here. you can even configure those to highlight outlier values so if individual points are interesting those are communicated. "data reduction" can also mean binning, e.g. summing/averaging the values in sets of 50 or 100 or whatever makes sense for what you are communicating. this makes fewer bars but ones which can be individually visible. Such binning can be done using hist

There are any other number of additional ways to reduce the data to a representable set, what makes sense for your data is very context dependent, so I cannot say. (Especially on a np.random test variable, but domain knowledge is important context too, so even just knowing what it is would likely not be enough)

There is really nothing to be done here, this is a fundamental limit of rasterized plotting. Thus I'll close this. Thank you for reporting.

When bars are significantly narrower than your pixel they can disappear. Your options are:

• set an edge color which is sized in output space so your bars will have a minimum visible size, but it means that bars are going to overlap
• us a vector format (e.g. svg), but that leaves you at the mercy of the browser / viewer and how it renders things that are too small
• turn of pixel snapping
• save at a high enough dpi where your bars are bigger than an output pixel
• switch to using step or stairs which draws a single line rather than many rectangles (and will also have better render performance)

xref #27816
#13909