Trace an event from click to analysis

WHAT THIS TESTS This is a foundational system-design question checking whether you understand how product analytics actually flow, not a textbook description of A/B math. The interviewer wants the stages, the components at each stage, and an appreciation that ingestion is asynchronous and built for scale and reliability.

A GOOD ANSWER COVERS The lifecycle begins on the client. An analytics SDK fires when the user clicks, capturing the event name, the experiment variant the user was assigned, a stable user id, a timestamp, and context like platform and app version. The SDK typically batches events and sends them over the network, retrying on failure with idempotency keys so a retry does not double-count. A collection endpoint, often a lightweight ingestion service behind a load balancer, validates the schema, attaches server-side metadata, and quickly acknowledges. From there events land in a streaming buffer such as a message queue or log, which decouples ingestion spikes from downstream processing. A processing layer consumes the stream, enriches, deduplicates, and writes the events into a data warehouse or columnar store. Finally, analysts or the experimentation platform query the warehouse, grouping conversions by variant to compute the difference between button colors. The variant assignment itself comes from a feature-flag or assignment service consulted at render time.

COMMON WRONG ANSWERS Proposing a synchronous insert into a production database on every click, which does not scale and couples the app to analytics. Forgetting batching and retries, so events are lost on flaky networks. Omitting the assignment step, so you cannot attribute clicks to a variant. Treating ingestion as instantly queryable rather than an asynchronous pipeline.

LIKELY FOLLOW-UPS How do you guarantee a user always sees the same variant? Deterministic hashing of the user id. How do you handle dropped or duplicate events? Idempotency keys and dedup in processing. What is the latency from click to dashboard, and why is it not instant?

ONE CONCRETE EXAMPLE A user opens the app; the assignment service hashes their id and shows the green button. They tap signup; the SDK records signup_clicked with variant green and a session id, batches it, and posts it to the collection endpoint, which acknowledges and pushes it to a queue. A stream job enriches and loads it into the warehouse, where the experiment dashboard later compares signup conversion for green versus the control blue.