Pageserver

Stable

How Pageserver stores data history, serves soft basebackup, and allows PostgreSQL to start with only a minimal local data set on the pod.

Updated: March 21, 2026

Pageserver is an internal SPG99 storage service. It exists so that a database can start quickly, stop, and start again without being tied to the permanent local disk of a specific Compute instance.

Put simply, Pageserver is the service of data history and bootstrap for PostgreSQL in SPG99. It:

  • stores timeline storage metadata;
  • publishes layers and the manifest to S3-compatible storage;
  • receives WAL history from Safekeeper;
  • serves soft basebackup to Compute for cold start;
  • handles remote read-through on local misses.

What matters most in the new model

Pageserver now serves not a heavy full working snapshot of user relation files, but a thin startup image.

This means:

  • only the startup minimum for PostgreSQL is kept locally;
  • system catalogs and relmap are materialized immediately;
  • user relation files do not have to exist locally in full;
  • Compute fetches them through Pageserver as needed.

In practice, this is exactly what allowed the amount of local data on the pod to shrink to the minimum required.

How Pageserver participates in the database lifecycle

Database creation

The Control Plane creates a new timeline in Pageserver.

First start

On the first startup, Compute requests a soft basebackup from Pageserver. Pageserver chooses the best bootstrap source and serves the startup image.

Running database

While the database is running, Pageserver:

  • keeps the WAL subscription;
  • materializes history;
  • serves page read-through;
  • publishes new layers and the manifest.

Restart

On a later restart, Pageserver again serves startup state, without requiring a full permanent local set of user files on compute.

Why this is fast

Pageserver accelerates the restart path for several reasons:

  • Compute does not depend on a permanent durable disk;
  • startup uses a thin soft image;
  • relation pages are fetched lazily;
  • bootstrap is streamed;
  • storage history is materialized gradually.

Why this is reliable

  • durability is not stored only in the local Compute PGDATA;
  • the WAL source is Safekeeper quorum;
  • Pageserver assembles working state from storage history, not from one “only” local worker.

Short summary

In today’s SPG99, Pageserver is the service that made the following possible:

  • soft basebackup;
  • the minimal local pod disk;
  • fast cold start;
  • safe restart without tying the database to one specific worker.