Consistency Models: From Strong to Eventual (and What CAP Doesn't Tell You)Limited sources1 source listed so far, so check specifics before you rely on them.This rates where the sources come from. It doesn't prove every claim.
Linearizable, causal, bounded staleness, session, eventual — what each one buys you, Cosmos DB's 5 levels, SQL isolation, read-your-writes, and sagas instead of 2PC.
By Amit UpadhyayPublished Updated 13 minArticle 1 of 6 in System Design
AI assistant · answers from this article
1 of 5
One database
Nothing to add yet
There is one copy of the data, so there is nothing to be out of date. Consistency models only matter once you make a second copy.
On this page
Consistency models, in one sentence
A consistency model is the promise a system makes about how fresh a read is, once the data lives in more than one place.
You reach for it when data is copied to several places, such as replicas, caches or regions, so reads can come from a copy that is behind. Not before: it costs more routing logic, extra load on the primary, and the need to state consistency requirements explicitly.
The cheat sheet
What to use when
You do not need all of this on day one. Each row is one signal and the piece that answers it. Stop at the row your system has reached, or pick one to see it in the picture above.
Where you areThe signalWhat to useWhat it costs
None. Reads and writes go to the same place, so every read is current.Nothing. Normal transactions are enoughNothing
The database is busy mostly with reads.Read replicasA read can be slightly behind the last write
They save, reload, and see the old value.Send that user’s reads to the primary for a few seconds after they writeStep 4 ↓More load on the primary
Checkout acts on an address or a balance that must be current.Read from the primary for that operationStep 1 ↓That read does not get faster by adding replicas
During a burst of writes, stale reads multiply.Measure replica lag and stop routing to a replica that is too far behindLess read capacity until it catches up
Start With One Copy
With one database there's nothing to be inconsistent with. Every read sees the last write, and a transaction is all the guarantee you need. Consistency becomes a question the day you make a second copy of the data: a read replica, a cache, another region. From then on a read can come from a copy that hasn't heard the news yet.
So the skill isn't picking the strongest model. It's deciding, for each operation, how fresh the answer has to be, and paying for strength only where being wrong costs something. The table above lists the signals in the order most systems meet them. This article gives you the vocabulary first, then the answer to each situation and a way to check it.
If users see stale or conflicting data
You used "eventual consistency" as an excuse for a bug
You can't explain the difference between linearizable and serializable
You read your own write and saw the old value, and assumed the cache was broken
Your Cosmos DB account is set to "Strong" everywhere because "safer"
This is the consistency menu — what each model actually guarantees, what it costs, and how to match it to your workload.
Pro deep dive13 min full read
Keep reading with Pro
This is one of our advanced production deep dives. Pro unlocks the rest of it, every other deep dive, and the See, Animate, Test and Build modes on every topic.
What it covers
Linearizable, causal, bounded staleness, session, eventual — what each one buys you, Cosmos DB's 5 levels, SQL isolation, read-your-writes, and sagas instead of 2PC.