ax1.plot(np.random.rand(100, 100), 'o', rasterized=True) creates 100 "lines", and hence 100 individual rasterized images. If instead you flatten the array, to create only one single "line",
ax1.plot(np.random.rand(100, 100).flatten(), 'o', rasterized=True) the filesize reduces tremendously (in my case from 800kb to 20kb).

I'd say, usually one would plot such bunchs of plots via collections. Is there any special use case that would require you to plot that many lines?

In my use case I am actually not using ax.plot. I am using a scatter plot (with ax.scatter) to show 5000 points leading to a similar file size issue.

Do you have a minimal example for how scatter is used in that case? ax1.scatter(*np.random.rand(2, 5000), marker='o', rasterized=True) would create a 37kb file (compared to a 64kb file in case the axes is rasterized) for me, so it doesn't look like a problem at all.

Here is an example which closely resembles my use case:

import matplotlib.pyplot as plt
import numpy as np

plt.rcParams['figure.dpi'] = 200

data = np.random.rand(6, 832)
y = np.array([100, 200, 300, 400, 500, 600])

fig1, ax1 = plt.subplots()
for ind in range(data.shape[1]):
    ax1.scatter(data[:, ind], y, marker='o', rasterized=True)
fig1.savefig('rasterizing_markers.pdf')

fig2, ax2 = plt.subplots()
ax2.set_rasterized(True)
for ind in range(data.shape[1]):
    ax2.scatter(data[:, ind], y, marker='o')
fig2.savefig('rasterizing_axes.pdf')

I think this faces the same issue as you described earlier... this is creating 832 rasterized images resulting in large file size.

I tried coming up with a workaround i.e. starting with ax2.set_rasterized(True) followed by ax2.xaxis.set_rasterized(False) etc.(before savefig but after the loop containing ax2.scatter) to disable rasterization on x/yaxis and spines. But this approach did not work.

Did you try:

data = np.random.rand(6, 832)
y = np.array([100, 200, 300, 400, 500, 600])
y = np.tile(y, (1, 832))

fig1, ax1 = plt.subplots()
ax1.scatter(data, y, marker='o', rasterized=True)

thanks! it worked and makes sense. Should I close the issue?

Well, it depends on whether you consider this issue to be solved or if there is still something that matplotlib can or should do?

the only issue that remains (which came up when I was trying out approaches to solve the current issue) is, why doesn't the following work?

fig, ax = plt.subplots()
ax.set_rasterized(True)
ax.plot([1,2,3])  # plot/scatter/whatever plotting method is needed

fig.canvas.draw()  # to trigger tick label formatters
for xlab in ax.get_xticklabels():
    xlab.set_rasterized(False)

fig.savefig('file.pdf')

xtick labels still end up being rasterized!

In the last case I suspect (with out digging into the code to check) it is due to the way the draw tree works. At the top level we draw the figure by recursively walking the artist tree depth-first. We have code paths in the vector backends to switch temporarily rasterizing an Artist (and implicitly all of it's children) but not a way to switch back to using a vector backend on those children. The tick labels are children of the axis objects which are children of the Axes object.

Thanks everyone!