Evaluating GPT‑4’s Ability to Design and Implement a Distributed Cache System for Uber

GPT‑4, the successor to the GPT‑3.5 model, was officially released and demonstrated by OpenAI CTO Greg Brockman, who showed the model turning a hand‑drawn sketch into a functional website and providing code solutions directly from error screenshots.

Former Google software engineer and Scaler teaching system designer Naman Bhalla used GPT‑4 to simulate an interview for a distributed cache system design for Uber. He asked GPT‑4 to produce a software requirement document, which GPT‑4 answered with a clear title, problem description, and a list of nine detailed requirements covering operations (Put, Get, Delete), configurable size and TTL, LRU eviction, multi‑node distribution, consistent hashing, horizontal scaling, quorum reads/writes, monitoring, a simple UI, and test cases.

GPT‑4 successfully generated the initial textual requirements and answered follow‑up questions about consistency strategies (e.g., session‑based, client‑side cache, sticky routing, read‑after‑write, versioning/timestamps, causal consistency) and listed typical cache features such as eviction policies, TTL, replication, and persistence.

When asked to provide Java code for the system, GPT‑4 initially supplied high‑level component outlines ( CacheNode , DistributedCache , ConsistentHashing , CacheClient ) and partial snippets. After iterative prompting, it delivered full implementations for CacheNode and other components, as well as a DistributedCacheTest JUnit test class covering basic put/get, delete, node addition/removal, request collapsing, and prefetching.

During the interaction, GPT‑4 exhibited occasional instability: it sometimes refused to provide complete code due to response length limits, repeated or omitted code, and produced an incorrect ExecutorService implementation for sticky key routing. With further clarification from Bhalla, GPT‑4 corrected these issues, refined the executor logic, and completed the missing test cases.

The overall assessment concluded that GPT‑4 can reach the level of an ordinary engineer in requirement gathering, design articulation, and code generation for a distributed cache, but it still requires human guidance to resolve context loss, non‑deterministic outputs, and incomplete implementations before the code becomes production‑ready.