I advocate for spot (or preemptable in GCP) because one is financially rewarded for practicing chaos engineering. If the system "has to" have one magic node always up, that's something which deserves attention and can lead to better system designs
The bad news is that such an exercise can often be literally more laborious than it pays back in cost savings, but engineering is filled with trade-offs so at least one can have a good explanation when someone points to a line item on a cloud bill and raises an eyebrow
