Data skew is a checkout line - how to unjam Databricks jobs without bigger clusters
The Databricks Data EngineerShow Notes
Your nightly Databricks job ran long, so you open the Spark UI. Two hundred tasks. One hundred ninety-nine finished in about a minute. One is still running, and it's been running for forty. You double the cluster and rerun it. Same forty minutes. Nearly twice the bill.
The cluster was never the problem. The rule that keeps your joins and group-bys correct is the same rule that lets one task carry the whole job, and no amount of extra workers moves that work off it. Sometimes the stuck task turns out to be worse than slow.
In this episode: - Why more workers and more shuffle partitions both leave the stuck task exactly where it was - Where heavy keys come from, and the one source that is usually a bug in a performance costume - Why plenty of skewed aggregations never slow down, and where adaptive query execution on Databricks does and doesn't split the load - When salting pays, when a different approach beats it, and when the right call is to leave skew alone
This episode is for Databricks data engineers whose reflex for a slow Spark stage is a bigger cluster. Whether you're mid-level and tired of paying for compute that changes nothing, or senior and about to be asked "how do you handle data skew?" in an interview, you'll walk away able to tell a skew problem from a capacity problem before you touch the cluster settings.
--- Helping 18,000+ Databricks data engineers become seniors: interview like seniors, execute like seniors, think like seniors.
Follow The Databricks Data Engineer for new episodes every week.
LinkedIn: linkedin.com/in/jrlasak Newsletter: dataengineer.wiki
#DataEngineering #Databricks #DataEngineer #CareerGrowth #ApacheSpark #DeltaLake
More from The Databricks Data Engineer
The rollout is done. Every production table migrated to Unity Catalog, every one owned by a team group, and the revenue table carries a certified check mark. The platform lead posts in Slack: fully governed. Monday morning, finance replies in the thread. The revenue on the finance dashboard and the revenue in the sales deck are about three percent apart. Which one's real?
Two proposals sit on a head of data's desk, both about the nightly pipeline that feeds the finance dashboard. One swaps the full reload for incremental loads with Auto Loader: it finishes in a fraction of the time and it's cheaper every month after the first. The other keeps the full reload, puts it on a bigger job cluster and starts it an hour earlier. It costs more. Forever. She asks one question, "What happens if this breaks on quarter close?", and approves the second one.
Sam is four years into a Databricks role, up for senior this cycle. Merged pull requests on the ingestion layer. The nightly job he rescued at three in the morning when the orders table stopped landing. Then his manager asks for one call where the answer wasn't obvious, he framed it, and he lived with the result. Sam starts three different sentences. None of them finish.
Open the workspace. Start a cluster. Get coffee. Come back, and sometimes it's ready. If you've been doing this a few years you can feel that wait in your shoulders, and it's the least interesting part of the story.
You cleared the recruiter screen. You cleared the Databricks technical round, and cleared it well. Then the final round, then the email that says strong candidate, tough decision. You go back looking for the mistake and there isn't one. You were good. Somebody was just easier to picture.