went through oracle's DS interview loop for a role on their customer success analytics team. sharing the breakdown because the oracle DS process is weirder than most and i couldn't find anything specific when i was prepping.
the loop had 4 rounds:
round 1: sql screen (45 min, live coding). this was the hardest SQL i've done in any interview. not aggregate-and-group stuff. they gave me a schema for a multi-tenant cloud billing database (revenue, usage_events, customers, contracts) and asked: write a query to find customers where their actual usage in the last 30 days exceeded their contracted limit, broken down by product line. required: window functions, CTEs, date math, self-joins. if you're going into oracle DS, SQL is the main event, not a warmup.
round 2: stats / probability (45 min). two problems. first was a bayesian updating problem: you have a model predicting customer churn, the model is 85% accurate on the training set, how do you think about its actual usefulness in production. they wanted to hear about base rates, precision vs recall tradeoffs, not just 'accuracy is misleading.' second problem was about AB test design: how many observations do you need, what are your null/alternative hypotheses, and how do you handle early stopping if the metric looks bad on day 3.
round 3: case study (60 min). they emailed a dataset + business question 24 hours before. i had to do the analysis on my own and walk through my findings. the business question was something like: identify which oracle cloud regions are underperforming on customer adoption relative to their launch maturity. pretty open-ended. they cared about how i structured the problem, what i chose to measure, and whether my recommendations were actionable.
round 4: behavioral / stakeholder. very focused on: have you worked with sales teams, have you presented findings to non-technical leadership, can you explain a model result to someone who doesn't want to hear it.
obvious pattern: oracle DS heavily weights SQL and business communication. the stats round exists but it's not a research-lab level bar. if you're coming from a pure ML background, you'll need to shift gears.