Let's see if you think like a senior data engineer. His Spark job takes 20 minutes every day. Today it took two hours. What is the first thing you will check? If you answer 'was', I will see the code.

I would start by inspecting the Spark job’s logs and metrics. First, I’d check the cluster health: are there executor failures, high GC pause times, or network bottlenecks? Next, I’d compare the current job’s DAG to the previous run to see if any new stages or partitions were added, which could increase shuffle overhead. I’d also verify that the input data size hasn’t increased; a larger dataset can easily double runtime. If the data volume is unchanged, I’d review recent code changes for expensive operations such as wide transformations, joins, or broadcast variables. Finally, I’d look at any recent changes to cluster configuration or Spark settings (e.g., executor memory, number of cores, shuffle partitions). By systematically checking infrastructure, data, and code, I can isolate the cause of the performance regression.