Wrapped up the LinkedIn DE loop about a month ago. I'm a senior DE with 7 years, mostly on large-scale batch and streaming systems. Targeting their data infrastructure org.
Process was four rounds after a phone screen, all virtual.
Phone screen: A recruiter call and then a technical screen with an L5 DE. The tech screen was 100% SQL: a multi-table join problem involving member activity aggregations, and one question about how you'd design a table schema for tracking job application events. I was expecting to talk Spark or Kafka but nope, they stayed firmly in SQL land for the screen. Good sign or bad sign depending on what you prep.
Round 1, coding: LeetCode-adjacent Python. Two medium problems. One involving graph traversal (BFS on a social network adjacency list), one on string parsing. Not particularly DE-specific honestly. Standard SWE coding bar.
Round 2, system design / data modeling: This was the interesting one. Prompt was roughly: design a data pipeline that powers LinkedIn's "People Also Viewed" recommendation feature, from data ingestion to the final feature store. They wanted end-to-end: what events you capture, how you batch vs stream, latency requirements tradeoffs, how you handle late data. I spent a lot of time on idempotency and at-least-once vs exactly-once delivery. They pushed hard on failure modes. What happens if your Kafka consumer lags? How do you backfill?
Round 3, SQL deep dive: Similar to DS loop but deeper. Window functions, recursive CTEs, one question on deduplication strategies for event streams. Know how to deduplicate with ROWNUMBER() partitioned by eventid.
Round 4, behavioral: The usual STAR method rounds. They care a lot about cross-functional collaboration stories, especially working with DS or analytics engineers to unblock their data needs.
Total time: recruiter reached out via LinkedIn, four weeks to offer. The offer was reasonable but I ended up declining for a competing offer. Would have accepted if they'd moved on base.