Meta system design interviews usually push on product sense and scale at the same time. You are often expected to turn a vague product prompt into APIs, data models, ranking logic, consistency choices, and tradeoffs around latency, correctness, and cost.
Below are 20 of the most discussed Meta system design questions reported by candidates. I grouped them by theme because Meta tends to revisit the same ideas, location, feeds, ranking, real-time systems, and large-scale ML, through different product surfaces.
Location, ranking, and recommendations
Design a location-based radius top-K search
This question is a standard test of geospatial indexing, candidate retrieval, and ranking under latency limits. A strong answer usually starts with how to partition the earth, filter by radius, then rank and return the top K without scanning too much data.Design Nearby and Notification Ranking
Meta asks versions of this because it combines retrieval and ranking with user context, freshness, and feedback loops. Interviewers want to see whether you can split the system into candidate generation, feature computation, model scoring, and online experimentation.Design place-of-interest ML system
This is less about CRUD design and more about data pipelines, labeling, training, and online serving for place understanding. Good candidates talk about noisy location data, feature stores, model updates, and how to measure quality beyond click-through rate.Design nearby place recommendations
This prompt usually tests whether you can connect maps-style retrieval with recommendation ranking. Expect follow-ups on cold start, sparse user history, travel versus home context, and how to blend popularity with personalization.Design Place Recommendation System
Compared with nearby recommendations, this one often opens the door to broader recommendation design, embeddings, collaborative signals, and exploration versus exploitation. A sensible structure is retrieval, ranking, feedback collection, and periodic model refresh.
Social graph, feed modeling, and privacy
Model entities for feed content and shares
Meta cares a lot about data modeling, and this question gets right to the heart of social products. Interviewers are looking for clean entity boundaries, ownership rules, share semantics, versioning, and how the model supports feed generation and analytics later.Design a post privacy/visibility system
This is a good test of access control at massive scale. A good discussion covers policy representation, friend and group relationships, cache invalidation, permission checks in read paths, and what happens when privacy settings change after content is already distributed.Design an Online Judge and Live Comments
The two halves of this problem pull in different directions, isolated code execution on one side and low-latency fan-out on the other. Interviewers want to see whether you can separate components cleanly, define trust boundaries, and reason about real-time updates under spikes.
Auctions, banking, and money movement
Design an online auction system
Auction design is a good way to test correctness under concurrency. Expect to discuss bid ordering, anti-sniping behavior, idempotency, payment state transitions, and how to keep the winner correct when requests arrive at the same time.Design an online auction platform
This version usually broadens the scope from a single auction flow to a marketplace product. Strong answers address seller listing flows, search, bid histories, notifications, fraud controls, and operational concerns around peak traffic near auction close.Design a basic bank system API
This is a direct test of transactional thinking. Interviewers care about ledger design, double-entry principles, idempotent APIs, auditability, and what consistency level is required for balances, transfers, and statement generation.Design a mini banking system
This question often starts smaller than the API version but quickly moves into the same territory, concurrency, rollback, and immutable transaction history. If you lead with the ledger instead of a mutable balance table, you usually set the discussion on the right path.Schedule and cancel delayed payments
This is a scheduling and correctness problem dressed up as fintech. The interviewer is testing durable timers, cancellation semantics, retry behavior, exactly-once versus at-least-once execution, and how to recover after service or queue failures.
Event pipelines, storage, and data systems
Design batch and streaming ETL architecture
Meta uses data everywhere, so this question is a proxy for whether you understand modern data platforms. A solid answer contrasts batch and stream paths, deals with schema evolution and late events, and explains how downstream consumers get reliable, queryable data.Design tables for event-driven metrics
This sounds narrow, but it is really about data modeling for analytics at scale. Interviewers look for event schema choices, partition keys, aggregation tables, retention strategy, and how you prevent double counting when events are replayed.Design a price tracking system
This problem mixes ingestion, deduplication, change detection, and notifications. Good candidates define the monitored entities, explain how crawls or feeds update price state, and handle noisy changes, alert thresholds, and user subscriptions.Design an in-memory cloud storage system
Even though this came up as a take-home, it maps well to interview discussion because it exposes tradeoffs fast. The key areas are object layout, metadata indexing, replication, eviction or persistence strategy, and how reads and writes behave during node loss.
Large-scale ML and AI infrastructure
Architect an asynchronous RL post-training system
This is the kind of ML systems question Meta uses for senior candidates. The interviewer wants to hear how experience generation, reward modeling, policy updates, evaluation, and checkpoint management fit together when workers are asynchronous and data quality is uneven.Design a scalable MoE pretraining pipeline
Mixture-of-experts systems bring hard distributed-systems issues into ML training. A strong answer covers data sharding, expert routing, all-to-all communication costs, fault tolerance, checkpointing, and how cluster utilization changes as model scale grows.Design image and multimodal generation systems
This prompt tests whether you can think across the full ML product stack: data, training, serving, safety, and feedback. Expect pressure on inference latency, GPU scheduling, prompt and asset storage, abuse prevention, and evaluation for quality and policy compliance.
How to prepare for this set
A pattern runs through these questions. Meta interviewers usually reward candidates who define scope early, identify the heaviest read and write paths, choose data models before jumping into boxes and arrows, and talk clearly about tradeoffs.
For product-heavy prompts, spend time on ranking and feedback loops, not just storage. For money and auction prompts, correctness and idempotency matter more than fancy architecture. For ML prompts, separate offline training, online serving, and evaluation, then explain the interfaces between them.
If you want more practice, PracHub has 115+ Meta system design questions, including the 20 above and many more reported by candidates.
Top comments (0)