AI, agents, databases, and serverless at AWS. Views are my own.
Joined October 2013
- Tweets 2,802
- Following 749
- Followers 21,224
- Likes 5,493
Make sure you (and your database vendor) are benchmarking from where your application is actually running!
Multi-region active-active applications see similar trade-offs between consistency and latency, with a higher multiplier (cross-region RTT can be tens of milliseconds).
Different database architectures come with different trade-offs.
DSQL has slightly higher in-AZ read latency than Aurora Postgres (by a few milliseconds), but incurs significantly fewer cross-AZ round-trips for multi-AZ applications.
If you haven't read the post, check it out here: brooker.co.za/blog/2025/11/0…
Updated my "DSQL: Simplifying Architectures" blog post with some more thoughts on active-active, and vector workloads.
What I'm guessing is that the wrapping logic wraps the Unicode modifier sequence onto the next line, so it no longer modifies the emoji.
Marc Brooker retweeted
DSQL: Simplifying Architectures brooker.co.za/blog/2025/11/0…
The Aurora DSQL team is on Discord. If you'd like to learn more about DSQL, or have questions, come join us:
New paper by Nancy Lynch summarizing her career's influence on the field of distributed computing.
arxiv.org/pdf/2502.20468
If you don't know who she is, she's the L in FLP and DLS. @MarcJBrooker has a good summary article: brooker.co.za/blog/2014/05/1…
Marc Brooker retweeted
The real AI challenge isn't building agents—it's operating them at scale. AWS's @MarcJBrooker says the gap is operational. Promising prototypes fail when they hit real-world security, reliability & cost requirements.
Get the full story: go.aws/4owP7EX
New blog post, looking at some of the decisions we made when building Aurora DSQL, and why we made them the way we did.
All systems decisions are trade-offs at some level. This post looks at how we chose the trade-offs that simplify building great systems in the cloud.
For multi-region clusters, you do still need to point your application at the right region. But we're going to make that complexity go away, too.
With Aurora DSQL, your application traffic automatically gets handled in-AZ, no configuration or hand-holding needed.
And no trade-off between consistency and latency either. DSQL reads are in-AZ and strongly consistent.
I'm kind of happy with the results of this game, because they support our intuition that Strong Snapshot Isolation (equivalent to Repeatable Read on single-machine Postgres) is the right default for Aurora DSQL.
A huge benefit of distributed databases like Aurora DSQL is that you can get both strong consistency *and* read scale-out without compromises. Along with high availability, including region- and datacenter fault tolerance.
There are compelling reasons in classic DB systems to have eventually consistent read replicas. But they do force builders to deal with all kinds of weird edge cases around consistency.
