What to Expect in the Runway System Design Interview

Runway's design round uses questions built from its own product: AI video features at scale. A system design interview asks you to plan the structure of a large software service. Candidates report questions about handling large video assets and using GPUs, the processors that run AI models. Machine learning candidates may get an ML system design round, which adds model serving and evaluation. Expect about an hour: requirements, architecture, then deep follow ups on cost and latency.

The Question Types Runway Asks

• Media pipelines. Upload, store, process, and deliver very large video files. A single project can hold gigabytes of assets.
• Generation job systems. A user requests a video, GPU workers produce it, and the user watches progress. This is the signature problem shape.
• Serving and cost. GPUs are expensive and limited. Designs must schedule work carefully and avoid idle hardware.
• Product features. Editing tools, sharing, and collaboration on creative projects. Here user experience drives the architecture.

The common thread is heavy compute plus impatient users. Creative users iterate: they generate, judge, adjust, and generate again. Latency, the delay before a result, decides whether that loop feels usable.

What the Interviewer Evaluates

• Requirements first. Who is the user, how large are the files, and how fast must results arrive.
• Async thinking. Generation takes seconds to minutes, so the design needs queues and progress updates. A queue is a waiting line for jobs.
• Cost awareness. Naming GPU utilization as a goal is a strong signal here. Utilization means the share of time hardware spends doing useful work.
• Failure handling. Jobs fail mid render. The design must retry without losing work or charging users twice.

A Walkthrough: Design a Video Generation Service

Here is a high level plan for the signature question.

Step 1: Requirements. Users submit a prompt and settings, then receive a video. Generation may take up to a few minutes. Users need progress updates and a gallery of results. The system must serve many users with a limited GPU pool.

Step 2: The request path. An API service accepts the job and stores its record. The job enters a queue with a priority, since paying users may go first. The user immediately gets a job id for tracking.

Step 3: GPU workers. A pool of GPU machines pulls jobs from the queue. Each worker loads the model once and reuses it across jobs. Workers report progress events, which the client receives through polling or a push channel.

Step 4: Storage and delivery. Finished videos go to object storage, a service for storing large files. A content delivery network, a set of servers near users, serves playback fast. Store thumbnails and metadata separately for quick gallery loading.

Step 5: Scale, cost, and failure. Autoscale workers on queue length, but note that GPU capacity is limited and slow to add. Checkpoint long jobs so a crashed worker resumes instead of restarting. Make job submission idempotent, meaning a retried request cannot create a duplicate charge.

State the core trade-off out loud. Bigger batches raise GPU efficiency but raise each user's waiting time. Naming that tension, and choosing based on user type, is the strong answer.

Common Mistakes in This Round

• Designing a synchronous request. A generation that takes minutes cannot block a web request. Missing the queue is the most common failure.
• Treating GPUs like web servers. They are scarce, costly, and slow to provision. Autoscaling assumptions from stateless services do not transfer.
• Forgetting the iteration loop. Users generate many drafts before one keeper. Design for cheap previews and fast retries, not single perfect runs.
• No cost discussion. At an AI company, cost per generation is a first class requirement. Bring it up before the interviewer does.

How to Prepare

• Learn the standard building blocks first. Grokking the System Design Interview covers queues, storage, and content delivery with worked examples.
• Practice one media design end to end. Design a video upload and streaming service, then add a generation step. This composite covers most of Runway's reported question space.
• Go deeper for senior loops. Grokking the Advanced System Design Interview treats partitioning, replication, and long running work in depth.
• Prepare the rest of the loop. See What is the Runway interview process like for every round, and Top Runway behavioral interview questions for the culture conversations.