One Spark task runs much longer than peers because a single key dominates the shuffled data. Which issue should be investigated?

1. Uniform load by definition
2. A table description typo
3. Data skew
4. A database password expiry as the only explanation