Indexes, storage, and distributed retrieval
Follow one search service as its requirements change.
Start with documents, indexes, and the read and write paths. Then follow a vector move, a tenant filter, a delete racing an indexer, and a missing shard. The opening map connects each mechanism to the answer it must preserve.
For backend, data, and AI engineers
Start with the map and vocabulary
- Connect documents, representations, indexes, and candidates across the read and write paths.
- Trace which payloads and postings an update must rewrite.
- Prove when skipping work is safe and when filtering needs an exact baseline.
- Reconcile versions, count resource costs, and preserve a complete answer across shards.
Begin with A Map of Search Systems . The primer introduces the search vocabulary; basic database queries are enough to start. Return to its shared vocabulary as you read. For the application context, read retrieval fundamentals .
A reading pass, then a worksheet pass
Keep one decision note
Read in order, follow the worked traces, and predict before opening each guided answer. Then solve the changed-input exercises with answers closed and carry each chapter's artifact into one cumulative decision note .
The 84 minutes estimate prose reading; exercises take additional time. Public engineering accounts and synthetic models support the lessons. The calculations are not engine benchmarks, and the book recommendations are study bridges.
Parts
Articles in this series
- A Map of Search Systems Connect documents, indexes, read and write paths, and the terms behind a correct answer.
- When an Index Becomes Your Data Model Trace what a vector move rewrites when placement becomes document identity.
- Doing Less Work, and Doing Work Faster Prove a safe skip, then separate candidate count from execution cost.
- Finding Neighbors Inside the Eligible Corpus Choose a filtered search path against the exact eligible answer.
- From Durable Write to Searchable Snapshot Reconcile replacements and deletes before ranking; recover from committed authority.
- Paying for Retrieval Across Storage and Shards Count bytes and dependencies, then defend a complete answer across shards.