Adobe System Design Interview Questions

For responsiveness, keep heavy processing close to the user. Use WebAssembly/WebGL in the browser for real-time editing and offload complex rendering to GPU-enabled backend workers. Store edits as an operation log rather than overwriting the file—this allows non-destructive changes and easy undo/redo. Stream lower-resolution previews while full-res renders process in the background. Implement autosave checkpoints to prevent work loss.

An event collector at the edge ingests tool actions (edits, exports, opens) and sends them into a durable event log like Kafka. Real-time processing jobs aggregate usage metrics into a low-latency OLAP store for dashboards, while raw events are stored in a data lake for ML and audits. Include schema validation, a dead-letter queue for malformed events, and region-specific storage for compliance.

Use a multi-CDN setup with routing decisions based on user location and real-time network health. Pre-generate different renditions of assets (thumbnails, optimized formats) and store them at edge locations. Implement cache warming for frequently accessed files and signed URLs for secure delivery. Monitor tail latency and dynamically steer traffic toward the lowest-latency path using RUM data.

A token-bucket or leaky-bucket algorithm works well, backed by a fast in-memory store like Redis. Store limits per API key or user tier, and enforce limits at the API gateway. For global scale, use local limits in each region plus periodic reconciliation of usage totals. Return structured 429 responses with Retry-After headers, and provide an API for developers to check their remaining quota.

Model documents using CRDTs or Operational Transforms so edits merge automatically without conflicts. Keep active editing sessions in memory and sync updates via WebSockets. Each document gets its own “room” for real-time events, with snapshots periodically saved to persistent storage. Include presence indicators, cursor positions, and a change history so collaborators can see who did what, and allow offline edits to merge seamlessly when a user reconnects.

Index textual metadata (titles, tags, descriptions) in a search engine like Elasticsearch, while also storing visual embeddings for similarity search using an ANN index. Allow filtering by license, file type, orientation, and color palette. Combine lexical matching with visual similarity scores for ranking. Precompute related assets to speed up “you might also like” recommendations.

Break the encoding process into independent chunks so it can be parallelized across GPU/CPU workers. Store source files in object storage, generate intermediate mezzanine formats, and then produce multiple output renditions (bitrates, formats) for HLS/DASH. Use a job queue with autoscaling workers to handle spikes in demand. Persist progress checkpoints so long jobs can recover from failures.

Keep a prefix index in memory for instant matches, ranking results by popularity, personalization, and recency. Use debounce logic to limit unnecessary requests, and cancel stale queries as the user types. Log keystrokes and clicks to improve ranking over time. Support typo tolerance so small spelling errors don’t block results.

Maintain separate indexes for recent queries, entity search (assets, templates, fonts), and action shortcuts. When a user types, query all indexes in parallel, merge results, and apply a unified ranking policy. Cache recent results per user for instant rendering on first keystroke. Ensure permission checks happen before returning entity results to avoid exposing restricted content.