Databricks sells AI. 24 of its 29 design questions are classic, not AI.

Posted on Oct 22 • Originally published at landthelab.com on Oct 22

24 of Databricks' 29 reported design questions are classic system design. Its 6 most reported questions include an ordered IP firewall with CIDR rules, and 5 are reported at no other company we track.

Databricks sells a data and AI platform. If you are also prepping for OpenAI or Anthropic, expect its design questions to look different from theirs. We counted every design question reported at Databricks (n=29): 24 are classic system design, the kind you find in a system design book, and 5 are AI system or ML design. Those 5 are retrieve tables and notebooks for a data-aware coding agent, a harmful content detection system, a design retrieval for a coding agent over a code base, an LLM content safety service and a GPU scheduling platform.

Snowflake, another data platform, has 20 classic of 23. OpenAI has 27 of 43 (63%), and Anthropic 6 of 22. If you prepped AI system design for the labs, you still need the classics for Databricks.

The design round runs 60 minutes, and you may write it in a Google Doc instead of on a whiteboard. The interviewer keeps extending the problem. Expect a smaller, more concrete system than at most big tech companies: API, schema, workflow, failure handling and a few scaling and concurrency choices.

These 6 are reported at Databricks more than any of its other questions, newest first:

5 are coding and one is design. The firewall, the store that tracks its own request rate and the log many threads write to are small systems you build inside the hour. Across all 74 questions, 46 are reported at no other company we track, 31 of the 40 coding ones among them. A shared question bank for big tech covers little of this.

The coding screen is 60 minutes of live coding in a runnable editor. Candidates describe one problem from a small set that keeps coming back: ordered IP firewall with CIDR rules, fastest single-mode commute on a city grid, snapshot set with frozen insertion-order iterators and key-value store that reports its recent request rate. 3 of the 4 are among the 6 reported the most. The interviewer pushes on edge cases and complexity, and Databricks makes a first estimate of your level after this round.

Staff backend candidates may get system design in this slot instead. Everyone else should be able to write all 4 from a blank editor, with tests, inside the hour.

The onsite has two algorithm rounds and a 60-minute concurrency round where you write correct, efficient multithreaded code. Most candidates call it the hardest round of the loop. The 7 questions tagged concurrency, newest first: durable event log writer shared by thousands of threads, a collaboratively edited playlist, thread-safe key-value store with atomic batch writes, blocking cache where reads wait for a missing key, a dependency-aware job scheduler, thread-safe single-machine chat server with channels and concurrent caching client for byte-range file reads.

Some loops make this a practical low-level round instead. The reported tasks include durable event log writer shared by thousands of threads, lazy array with deferred map and indexOf and hybrid run-length and bit-packing codec, plus file system tasks and a cache. If your language’s locks, condition variables and queues aren’t in your fingers, practise this round more than any other.

Databricks’ questions by the round candidates named (n=64 with a known round). The onsite coding rounds produce the most. The timed online test is reported for some candidates only, and every one of its 11 questions we hold was last asked by Sep 2024.

26 of the 74 questions were asked since Sep 1, 2026, 14 of them coding. Spend your time on the live rounds. Treat the online test bank as warm-up, not as a list of what you will get.

74 questions reported at Databricks, 65 of them with a last-asked date (newest Oct 15, 2026), taken Oct 21, 2026. 64 have a known round. The design mix compares Databricks with Snowflake, OpenAI and Anthropic (Databricks n=29, Snowflake n=23, OpenAI n=43, Anthropic n=22). “Reported the most” means rated high or very high on how much candidates report it. “Reported at no other company” means no other company in our data has it. A question counts as concurrency when it carries that topic tag. The loop (rounds, lengths, formats and what they grade) comes from candidate reports and public guides, compiled in Land the Lab’s loop for Databricks.

Limits: the sample is small. 10 questions have no known round, and 9 have no last-asked date. Reports lean toward software engineering roles. We don’t publish counts per question, only the tier. Loops vary by team and level.

How we made this: Land the Lab counted the reports by type, round and date. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.