Read-your-writes

How Silos uses LSN tokens to guarantee a session sees its own writes, even when reads come from a lagging replica.

Read-your-writes is the default consistency level. It guarantees that within a session, any read reflects that session's own prior writes — so the classic "I just inserted a row and my next query doesn't see it" problem doesn't happen, even when reads are served by replicas that lag the primary.

The mechanism: LSN tokens

Silos uses log sequence number (LSN) tokens to make this work:

A write returns an LSN

When you commit a write, the primary records its position in the write-ahead log as an LSN. That LSN is, in effect, a watermark: "the database is at least this current."

A read can carry a minimum LSN

A subsequent read can pass a minimum LSN — the position it needs the data to have reached. The default session behavior tracks your latest write's LSN automatically.

The router enforces it

The router checks the candidate replica's position:

If the replica has caught up to the required LSN, it serves the read locally.
If the replica is behind, the router waits for it to catch up, or routes the read to the primary, which is always current.

This gives you local-latency reads when the nearby replica is fresh enough, and a correct fallback when it isn't — without you having to reason about replica lag.

What it guarantees (and what it doesn't)

Guarantees: within a session, you never read older than your own latest write.
Does not guarantee: that you see writes made by other sessions instantly. Those may still be slightly behind on a replica. If you need the freshest cross-session data, use strong consistency, which reads from the primary.

When to use a stronger level

Read-your-writes is the right default for most application logic. Step up to bounded staleness if you need a freshness bound across sessions, or strong if every read must reflect the absolute latest committed state globally. See Consistency levels.

LSN-token read-your-writes is implemented, but it operates over the rolling-out replication path. Until cross-region streaming is production-proven, the dependable way to always read your latest write across regions is to read from the primary. See the Replication overview.