You changed pcolor and pcolormesh, but are showing an example with scatter. What do the before and after plots look like? I can't really evaluate what has changed.

Gah, thanks for spotting that 🤦 . I've updated the script, and added the timings I get. pcolormesh seems to suffer ~10% performance wise, but pcolor doesn't change.

I guess I was more interested in the visual change, which I think this has?

Here's before:

And after:

Using:

import numpy as np
import matplotlib.pyplot as plt

x = np.arange(512, dtype=float)
y = np.arange(512, dtype=float)
z = np.arange(len(x) * len(y)).reshape(len(x), len(y))

fig, ax = plt.subplots()
ax.pcolormesh(x, y, z)
ax.set_yscale("log")

plt.show()

Okay? I guess I'm not sure how you choose ymin in the case that y[0]=0. Is this the accepted solution for this for other undefined y/xlimits?

If one scatters the same x/y points used for the pcolormesh (that includes a zero value in the ys), there's similar behaviour:

I'm not sure why pcolormesh chooses a slightly lower y-limit, but I'd say that's in the weeds and this is a definitely improvement on past behaviour.

import numpy as np
import matplotlib.pyplot as plt

x = np.arange(512, dtype=float)
y = np.arange(512, dtype=float)
print(y[0])
z = np.arange(len(x) * len(y)).reshape(len(x), len(y))

fig, axs = plt.subplots(ncols=2)
ax = axs[0]
ax.pcolormesh(x, y, z)
ax.set_yscale("log")

ax = axs[1]
ax.scatter(x, y)
ax.set_yscale("log")

plt.show()

I'm not sure why pcolormesh chooses a slightly lower y-limit,

T.b.c: This may be a center vs. edge topic. Plot only chooses limits based on the points (centers of the mesh). The mesh may choose limits based on its edges.

I believe k[0]=0 is not special for us: Actually, the "center" coordinates of the mesh do not matter at all. AFAIK, we calculate the edges of the mesh and the limits are then determined from the positive edge coordinates (0 and negative coordinates are ignored). Note that even for all k>0 the edges can get negative (see #29615 (comment)).

Yes, I think that's right. If one scatters the edge coordinates of the mesh, the bottom limit is set similarly:

import numpy as np
import matplotlib.pyplot as plt

x = np.arange(512, dtype=float)
y = np.arange(512, dtype=float)
print(y[0])
z = np.arange(len(x) * len(y)).reshape(len(x), len(y))

fig, axs = plt.subplots(ncols=2)
ax = axs[0]
ax.pcolormesh(x, y, z)
ax.set_yscale("log")

ax = axs[1]
edges = np.concat([y - 0.5, [y[-1] + 0.5]])
ax.scatter(edges, edges)
ax.set_yscale("log")

plt.show()

The top limit is different because scatter adds some padding, but pcolormesh clips to the final edge.

Since this has two approvals with no changes since either, I'll merge this tomorrow unless there are any objections before then.