# Design a federated learning system: system design interview question

> A shared model must improve despite device loss, uneven participation, and privacy limits across repeated releases.

Source: https://sysround.com/questions/design-federated-learning-system
Updated: 2026-11-05


Difficulty: hard. Levels: Senior, Staff+.

## How the interviewer opens it

"Design a system that trains a shared model from data on user devices without collecting their raw examples. Devices have different capabilities and do not always remain connected."

## What to clarify first

- Task and users
- Round scale
- Payloads and read:write mix
- Device limits
- Latency
- Privacy threat model
- Privacy across releases
- Consistency and availability
- Durability and retention
- Existing systems and exclusions

## Scope of a strong answer

Core requirements:

- Select eligible devices and distribute a shared model
- Combine device updates without exposing individual updates
- Handle device loss and different device capabilities
- Release models and metrics within a defined privacy budget

Usually out of scope:

- Design a new model architecture
- Design cryptographic primitives
- Defend against compromised devices or a malicious coordinator
- Build account and device enrollment systems

## What interviewers probe

### Rounds with lost connections

Devices can disappear during several protocol phases. Retries must not count an update twice or apply a round to the wrong model.

### Which devices contribute

Fast, connected phones can dominate the model. Their data may differ from data on devices that rarely finish.

### What the server can learn

Model updates can reveal examples without containing raw text. Device loss also changes which aggregate the server can safely recover.

### Privacy across model versions

Repeated model and metric disclosures consume privacy budget. Overlapping time windows and repeated device contributions make a single-round guarantee insufficient.

## Common mistakes

- Treats the absence of raw examples as a complete privacy guarantee.
- Uses unlimited local work despite device limits and unequal data volumes.
- Combines updates from different model versions without defining the algorithm.
- Requires every invited device to finish a round.
- Claims random invitations remove all participation bias.
- Ignores cumulative privacy cost or treats rolling windows as calendar months.

## The bar by level

- **Senior:** Defines recoverable rounds and version rules. Explains secure aggregation, contribution bounds, rolling privacy accounting, and device bias. Uses payload estimates to choose storage and delivery.
- **Staff+:** Defines the trust boundary and disclosure rules precisely. Tests privacy assumptions against selection, overlapping windows, retries, metrics, and rollback. Balances model quality, device access, and enforceable release limits.

Practice this question with an AI interviewer: https://sysround.com/practice/new?problem=federated-learning


---
SysRound (https://sysround.com) is an AI practice tool for system design interviews. It has a voice or typed AI interviewer and a shared whiteboard. At the end, you get feedback on each skill and a model answer. The AI feedback can be wrong.