Canva · Primly Community

Canva data scientist interview (SQL + case + stats), full breakdown of what I saw

ds_dmitri · 4 replies

Went through the Canva DS interview loop for a product analytics DS role about six weeks ago. There's not much out there specifically on Canva's DS process so here's what I saw.

Total rounds. Four after the phone screen: a SQL/technical screen, a case study, a behavioral round, and a final stakeholder round.

SQL round (45 min). This was harder than I expected. Two problems. First was a standard aggregation + window functions question (running totals, rank within groups). The second involved a more complex multi-step analysis: given a user activity table, identify users who fit a specific retention pattern and then compute a metric across those users. Required subqueries and some careful thought about the join logic. If you're not fluent in window functions and CTEs, prep that first.

Case study (60 min). They gave me a product analytics case: a specific Canva feature's engagement has dropped 15% over the last 30 days. Walk me through how you'd investigate it. This is structured product sense + analytics, not ML.

I walked through: data validation first (is the drop real or a tracking bug), then segmentation (which user segments, which platforms, which geos), then funnel analysis, then hypothesis generation. The interviewer pushed on what data I'd want that I don't have access to, and how I'd prioritize my investigation given that collecting some data takes time.

Stats questions. They came up in the case study, not as standalone questions. I got asked how I'd test whether the drop was statistically significant, and a follow-up about what the risks were in running an A/B test on a feature that's already declining.

What I did not see. No ML modeling questions at all. No Python coding. This was a product analytics role and the interview matched that. If you're applying for an ML or applied science role the loop is probably different.

What mattered most. Structured thinking under pressure. The SQL problems can be learned. The case study is where they separate candidates who think like analysts from those who know how to query.

4 replies

analyst_ana

The 'is the drop real or a tracking bug' first step is so important and I always feel slightly ridiculous saying it out loud in case studies. Good to know they actually want to hear it. I've been trained by bad interviewers to skip past that.

de_derek

The A/B test follow-up on a feature that's already declining is a classic gotcha. The answer they're looking for is something about novelty effects, selection bias in who's still using the feature, and whether you can even trust a clean control/treatment split at this point. Did they probe on any of those?

ds_dmitri

Exactly those three. They seemed to care most about whether I understood that the declining users are a biased sample. I also mentioned the ethical concern of intentionally keeping some users on a degraded experience to run the test, which they liked.

qa_quinn

No Python coding at all is interesting. Some DS loops I've seen make you implement a logistic regression from scratch in 45 minutes, which tests almost nothing useful. Sounds like Canva's process is more practically calibrated.