Scaling PostgreSQL without Microservices: Lessons from Notion’s 480 Shards

Source: Dev.to

TL;DR: Scaling Notion’s Monolith

• Application‑Level Sharding – 480 logical shards mapped to a smaller set of physical nodes.
• Shard Router – Implemented in TypeScript using space_id % 480 to route requests instantly.
• PgBouncer – Acts as a traffic controller, pooling connections to prevent overload.
• Zero‑Downtime Migrations – “Shadow Write” strategy moves billions of rows while keeping the app live.

Chapter 1: The Problem with Monolith

In Notion’s early architecture, a Node.js backend was paired with a single PostgreSQL instance. Over time, they encountered:

• CPU Saturation – Daily spikes regularly hit 90 %+ utilization.
• The Vacuum Problem – Autovacuum couldn’t keep up, risking a Transaction ID wraparound, which would halt writes to protect data integrity.

Chapter 2: Why Not Microservices?

The typical microservice argument is “split the code, split the load.” Notion chose a different path, keeping a monolithic codebase while scaling the data layer.

Chapter 3: The 480‑Shard Blueprint

Logical Shards

• Partition Key – space_id is used so all data for a workspace stays together, enabling fast joins.
• Setup – 480 independent schemas (logical shards) are distributed across 32 physical AWS RDS instances.

Benefits

• When a server becomes overloaded, a logical schema can be moved to a new server, providing linear scalability.

Chapter 4: The Great “Shadow” Migration

Backfill

• Historical data is moved to the new shards in the background.

Double Writing

• New changes are written to both the old database and the new shards simultaneously.

Cutover

• After a comparison engine verifies data parity, traffic is switched to the new shards with zero downtime.

Future Improvements

• Scaling to 96 Nodes – Expanding the physical node count to further distribute load.
• Blocks All the Way Down – Continuing to treat every piece of content as a block for fine‑grained sharding.
• Data Lakes & Connection Hubs – Exploring centralized analytics and integration points.

Feel free to share your thoughts in the comments!