Slicing cube with dask.slicing (dask 2.0) runs into stack overflow

[valeriupredoi]

hey guys, we noticed that a particular type of cube slicing done by dask.slicing with dask=2.0.0 produces a stack overflow error. The problem is described in ESMValGroup/ESMValTool#1188 specifically ESMValGroup/ESMValTool#1188 (comment) and boils down to an example: take any cube with time-lat-lon dims:

cube[(0, slice(None, None, None), slice(None, None, None), slice(None, None, None))]

results in Fatal Python error: Cannot recover from stack overflow. and core dumping.

I am not 100% sure this is dask's fault, but if it is, we should raise it with them 🍺

[cpelley]

Does this present pre dask v2 I wonder?
I have an issue with memory with iris2 when slicing a cube with time but for me the proccess itself is just killed (due to memory). The same example works on iris1

[valeriupredoi]

@cpelley no, all works fine with the latest dask pre 2.0 (1.2.2). Fairly sure it's having to do with the code changes in dask.slicing for 2.0 but since those code changes are a LOT I am not able to figure out what is going on. Maybe you or @bjlittle can file an issue on the dask gitHub page? 🍺

[bjlittle]

Thanks @valeriupredoi

I'll investigate further. Related to #3349

[bjlittle]

Closed by #3355