Designing a Multi-Tenant Distributed Key-Value Store

Reported in interviews at LinkedIn

Problem statement

How would you build a shared key-value store for ML and analytics teams? It needs point operations, batches, conditional updates, expiry and optional ordered scans at high traffic. Acknowledged writes must survive an availability-zone outage, and some clients must immediately read their own writes.

Define the API, partitioning and replication guarantees.

Show hint

Which client requirement (scans, CAS, read-after-write) constrains each design choice most, and what does each one need from the storage layout?

Relevant roles: ML Engineer and MLOps Engineer.

Related free problems