What to Expect in the Glean System Design Interview

The Glean system design interview asks you to design systems from its own product area: enterprise search. Enterprise search means one search engine over all of a company's internal apps and documents. Candidates report questions about indexing content from many workplace tools, search that respects user permissions, and ranking that mixes keywords with machine learning. The round runs about 60 minutes inside the onsite loop. It is graded on structure, trade-offs, and whether your design would actually protect customer data. You do not need to have built a search engine. You do need to reason carefully about data access, freshness, and scale.

The Question Types

• Federated indexing. Design a system that pulls content from many apps, such as chat, documents, and tickets, into one search index. An index is a data structure built for fast lookup.
• Permission-aware search. Design retrieval where every result respects what the searching user may see in the source app. This is the signature Glean problem.
• Hybrid ranking. Combine keyword matching with embeddings, which are numeric representations of meaning, and rerank the results.
• Fresh ingestion. Design incremental updates so the index reflects edits within minutes, despite per-app rate limits.
• Assistant orchestration. Design the layer that lets an AI assistant call search and other tools to answer a question.

A Walkthrough: Design Permission-Aware Enterprise Search

Step 1: Requirements. Suppose a company with 20 connected apps, millions of documents, and thousands of users. Queries must return in a few hundred milliseconds. The hard rule: no user ever sees a document they cannot open in the source app.

Step 2: Connectors and ingestion. Each app gets a connector, a program that reads content and changes through the app's API. Connectors differ in rate limits and data shapes, so run them independently with per-app queues. Normalize everything into one document format with text, metadata, and an access list.

Step 3: The index. Store documents in a search index with both keyword and embedding representations. Index the access information alongside the content. Access lists change often, so store them in a form you can update without reindexing the whole document.

Step 4: Permission enforcement. Present the two known options and compare them. Option one: copy access lists into the index and filter at query time. It is fast but risks staleness. Option two: check the source app at query time. It is always correct but slow and limited by rate limits. A common answer combines them: filter with indexed access lists, keep them fresh with change feeds, and re-verify the most sensitive sources. Say the failure rule out loud: when unsure, hide the document.

Step 5: Ranking. Retrieve candidates with keywords and embeddings in parallel, merge, then rerank with a model using signals such as recency and the searcher's team. Keep permission filtering before ranking, so the model never scores forbidden documents.

Step 6: Trade-offs. Close with the tension between freshness, latency, and safety. Naming permission staleness as the top risk shows you understood the company's actual business. If time remains, add monitoring: an alert on connector lag and a periodic audit that samples results against source permissions.

What Interviewers Listen For

• Safety first. A beautiful design that can leak one document fails. State the deny-by-default rule early.
• Realistic connectors. Per-app rate limits, partial failures, and retries. Uniform assumptions about sources signal inexperience.
• Ranking literacy. You need working knowledge of keyword search plus embeddings, not research depth.
• Numbers. Document counts, query latency targets, and index update delay, stated and defended.

How to Prepare

• Learn the standard blocks. Queues, indexes, caches, and fan-out appear in every question above. Grokking the System Design Interview teaches them in interview order.
• Add depth on data-heavy patterns. Grokking the Advanced System Design Interview covers partitioning and replication, which fit the indexing questions.
• Practice the walkthrough out loud. Design permission-aware search once on paper, then explain it in 20 minutes.
• Know the rest of the loop. See what is the Glean interview process like for the other rounds. Then prepare the paired stories with top Glean behavioral interview questions.