The scenario described—where most tasks (minimum and median) finish quickly but one or a few tasks (maximum) take orders of magnitude longer—is a classic symptom of Data Skew.

Why this is correct:

Data skew occurs when the underlying data is not distributed evenly across partitions. The tasks responsible for the 'heavy' partitions become stragglers, processing far more records than their peers and preventing the stage from completing.

Why other options are incorrect:

• Network/Regional Latency: If cross-region data transfer were the issue, you would likely see an increase in the median duration across many tasks, not just a single outlier.
• Task Queuing: Thread pool issues would result in tasks waiting to start even when cores are available, affecting task scheduling rather than individual task execution duration.
• Disk Spillover: While spills to disk slow down tasks, they usually occur when executor memory is exceeded and tend to affect multiple tasks or lead to Out-of-Memory (OOM) errors rather than creating a single task that is 100x slower.