I think with spot instances there is always the chance of getting errors because of eviction.

If all workers are evicted the query state is probably also lost, unless Databricks SQL keeps this stored somewhere.

For job clusters that is the case anyway, perhaps it works different on databricks sql.