Too-Small Clusters Cause Jobs to Fail Without Logs

[time-less-ness]

To reproduce, create a cluster of t2.nano (or otherwise very small) instances, and try to exec any simple Skypilot job that just does echo "thing". All such jobs will fail. Log files will be empty.

Uncertain how to proceed. Best guess is: when running sky launch check RAM size of suggested nodes, and if it's less than 2GB (or so? I'm not sure the actual cutoff, but 1GB is definitely too small) then output a very loud WARNING that cluster jobs may fail without output because nodes are too RAM-constrained for underlying cluster orchestration to run properly.