A Spark task is much slower than its peers because one join key dominates the data. What should be investigated?

1. A missing table SELECT grant
2. An overly short cluster auto-termination setting
3. A uniformly undersized driver result buffer
4. Data skew in the join