Pinterest · Primly Community

Went through the Pinterest recsys loop last month, here's what actually mattered

qa_quinn · 4 replies

Just finished the Pinterest ML/recsys loop about five weeks ago. Sharing notes because I wish I'd had something like this going in.

Rounds: recruiter screen (30 min), ML phone screen (45 min), then four virtual on-site rounds: ML fundamentals, system design for a recommendation feature, coding (one medium-hard algo), and behavioral.

The ML phone screen was less about deep learning and more about fundamentals. Precision/recall tradeoffs, how you'd handle cold start, what metrics you'd pick for a ranking system and why. Nothing exotic but they go pretty deep on the "why" behind every answer. I gave a textbook answer on AUC early on and the interviewer immediately asked when AUC would mislead you. Be ready for that.

System design was specifically about how you'd build a home feed ranking system at scale. They're not expecting you to reproduce their actual system but they want to see you think about latency constraints, freshness vs. quality tradeoffs, and how you'd measure success after launch. I spent too long on infrastructure and not enough on the retrieval vs. ranking split. Feedback I got later confirmed that.

Behavioral round was real, not a box-check. Two interviewers, 45 minutes, deep dive on specific past projects. They kept asking follow-up questions until I either gave them real detail or ran out of road. Vague answers don't survive contact here.

Got an offer at L5. The process felt thorough but never arbitrary.

4 replies

backend_bekah

the cold start question is everywhere in recsys interviews and i still see people fumble it. "just use content-based filtering" is not an answer, it's an opening bid. did they probe into how you'd bootstrap signals for a brand new item type, or was it mostly about new users?

ml_mike

both actually. new users got more time since Pinterest is heavily dependent on early pin behavior to build the taste graph. they walked through what signals exist before a user has pinned anything (email domain, location, signup flow choices) and how you weight them. item cold start came up as a follow-up when I mentioned freshness.

de_derek

what was the data infra angle in the system design? i'm curious whether they go deep on the pipeline architecture or keep it at the model serving layer

sre_sol

the latency vs. quality tradeoff in home feed ranking is genuinely interesting territory. most companies say they care about it and then set a 200ms p99 target and call it a day. did they seem to actually have opinions about where that line sits?