What happened:

When using a version of numpy that doesn't implement the shape argument to zeros_like, the zero fill in triu and tril is always float.

What you expected to happen:

The computed chunks should all have the same dtype as the array itself, which should come from the input argument.

Minimal Complete Verifiable Example:

import numpy as np
import dask.array as da

x = da.ones((6, 6), chunks=2, dtype=np.int64)
t = da.triu(x)
chunk = t[4:6, 0:2]
print(chunk.compute().dtype)

Prints float64 on numpy 1.15.1, but int64 (correct) on numpy 1.20.1.

Anything else we need to know?:

It's caused by the fallback code in zeros_like_safe, which ignores the dtype. There is a fix for this in #7277, but I thought I should file it as a separate bug for changelog tracking purposes. The fix is largely independent of the rest of #7227, so let me know if you'd like it cherry-picked into a separate PR.

#6997 might also fix it in passing.

Environment:

• Dask version: master (d540d73)
• Python version: 3.8
• Operating System: Ubuntu 20.04
• Install method (conda, pip, source): source