Regions & replicas
How Silos places databases across regions and materializes on-demand edge replicas near your users.
Silos can run a database close to your users. A primary serves writes, and replicas can materialize in other regions to serve reads locally — hydrating from storage and staying current by following the WAL. This is how you get low-latency reads without standing up and managing a multi-region cluster yourself.
On-demand edge replica materialization and global WAL streaming are rolling out. The mechanism is implemented, but multi-region production availability is still maturing. Treat region count and replica behavior described here as the model we're building toward, and check the Replication overview for current status before depending on it in production.
Regions
A region is a geographic location where Silos can run a database. There is a canonical set of roughly ten regions available today, with broader coverage rolling out over time.
You may see references to "100+ regions" — that's a target as the edge network rolls out, not a count of live regions today. We label rolling-out capabilities explicitly so you always know what's available now versus planned.
On-demand replicas
Rather than pre-provisioning replicas everywhere, Silos materializes them on demand near where reads are coming from. When a replica is needed in a region, it hydrates from the latest snapshot and then follows the WAL to stay current. This keeps the cost model lean — you don't pay to keep idle replicas running all over the world.
Read and write paths
- Writes go to the primary. There is a single authoritative writer for a database.
- Reads can be served by a nearby replica for lower latency, or routed to the primary when a query demands the freshest data.
How fresh a read needs to be is a per-query choice. Silos supports tunable consistency — eventual, read-your-writes, bounded-staleness, and strong — so you decide the trade-off between latency and freshness on each query. See Consistency levels.
How replicas stay current
A replica follows the primary's WAL. As the primary writes, WAL is streamed out and applied on replicas, so they converge on the primary's state. Global sync is not instantaneous — it's bounded by physics — so a replica may briefly lag the primary, and the consistency level you choose controls how that lag is handled for a given read.
Real-time WAL distribution across regions is part of what's rolling out. Until it's fully production-ready, treat global, always-on replica freshness as a maturing capability rather than a guarantee.