Build with MongoDB

Choose MongoDB for: flexible/evolving schemas, document-centric data (JSON APIs), rapid prototyping, horizontal scaling needs, and when embedding related data improves performance. Choose PostgreSQL for: complex transactions, strict relational integrity, or when your team knows SQL. Many modern apps use both for different purposes.

Atlas is MongoDB's fully managed cloud service. It handles: provisioning, patches, backups, scaling, and monitoring. Use Atlas for: production deployments (removes ops burden), serverless workloads, and when you want vector search. Self-host only if you have strict data residency requirements or existing infrastructure expertise.

Think in terms of access patterns, not normalization. Embed related data that's accessed together (reviews inside products). Reference data that's accessed separately or grows unboundedly. Design for your queries—MongoDB rewards query-driven schema design. Use schema validation for critical fields.

The aggregation pipeline is MongoDB's data processing framework. Chain stages like $match (filter), $group (aggregate), $project (shape), $lookup (join), and 30+ others. More powerful than SQL GROUP BY—handle complex transformations, time-series analysis, and graph traversal in a single query.

Create a vector search index on your embedding field in Atlas. Store embeddings (from OpenAI, etc.) as arrays in documents. Use $vectorSearch aggregation stage to find similar documents. Combine with $match for hybrid semantic + filter search. Great for RAG—store content and embeddings together.

Create indexes for frequently queried fields. Use compound indexes matching your query patterns. Analyze with explain() to verify index usage. Avoid large document arrays that grow unbounded. Project only needed fields. For complex queries, use covered queries that read only from indexes.

Embedding: store related data inside the document (faster reads, atomic updates, good for 1:1 or 1:few). Referencing: store a reference (ObjectId) and query separately (avoids duplication, handles 1:many or many:many). Use $lookup for reference joins. Most apps use both patterns strategically.

MongoDB supports multi-document ACID transactions since 4.0. Use sessions for transaction scope. Transactions work across collections and shards. However, MongoDB's document model often avoids needing transactions—embed related data that updates together. Use transactions when you truly need multi-document atomicity.

Analyze SQL schemas and queries. Flatten normalized tables into documents. Consider embedding vs referencing based on access patterns. Use MongoDB Relational Migrator for semi-automated conversion. Test performance with realistic data. Often, a MongoDB-native redesign outperforms literal SQL translation.

Realm (now Atlas Device SDK) provides: local database for mobile/desktop, real-time sync with Atlas, offline-first architecture. Use for: mobile apps needing offline support, reactive data binding (Swift, Kotlin, React Native), and real-time collaboration. It's a client-side database that syncs, not just an ORM.

Atlas includes: automated daily backups with point-in-time recovery, cross-region backups, and queryable backup snapshots. For self-hosted: use mongodump for logical backups, filesystem snapshots for large databases. Test restores regularly. Configure replica sets across availability zones for high availability.

Change streams provide real-time notification of database changes. Subscribe to insert, update, delete, and replace events on collections, databases, or whole clusters. Use for: event-driven architectures, cache invalidation, ETL pipelines, and real-time dashboards. Resumable with resume tokens for reliability.