While monitoring a Spark job, you observe in the Spark UI that for a specific stage, the minimum and median task durations are nearly identical. However, the maximum task duration is approximately 100 times longer than the minimum. What is the most likely cause of this performance bottleneck?