The follow-up, on a harder question about the changing role of junior software engineers: brooker.co.za/blog/2026/03... (And a 19th century book about trains 🚂)

New blog post, on the changing role of principal/staff+ IC leaders: brooker.co.za/blog/2026/03... Core message: get building, get learning.

New blog post, about a great little way to add fairness to distributed systems. Combining stochastic fairness queuing, best-of-two, and shuffle sharding. Read it here: brooker.co.za/blog/2026/02...

Now on YouTube, Monday's CMU Database Seminar on Aurora DSQL. Thanks again to the awesome CMU folks (especially Andy Pavlo) for inviting me to do this talk. www.youtube.com/watch?v=mK2h...

Today's Postgres vs. World Seminar Speaker: Marc Brooker (@marcbrooker.bsky.social) will present the architecture of Amazon's Aurora DSQL Postgres-compatible serverless OLTP DBMS. Zoom talk open to public at 4:30pm ET. YouTube video available after: db.cs.cmu.edu/events/pg-vs...

The cost of turning written business logic into code has dropped to zero. At best, near-zero. The cost of integrating services and libraries, the plumbing of the code world, has dropped to zero. At best, near-zero. What does that mean for the future? New blog post: brooker.co.za/blog/2026/02...

The world if MacOS supported SECCOMP_RET_USER_NOTIF

I'm still learning how easy it is to go from "software I wish existed" to "software that exists" now, especially for anything that's a script, automation, or small utility. Everything I internalized in 20 years about the cost of building things has changed in less than two years.

I’m happily building DSQL persistence for @temporal.io. @marcbrooker.bsky.social and team have gifted us a fab service. I readily accept I may be messing up my integration 😂 Guidance greatly accepted 🙏 (Oh! @kiro.dev has proven invaluable in this ❤️)

New blog post "Pass@k is mostly bunk" brooker.co.za/blog/2026/01... Pass@k, especially for k>3, can make basically anything look good.

In my first blog post of the year: The Box. Constraining what AI agents can do is required for making them safe and reliable. Pushing that control out of the agent itself, into its environment (The Box), is the only way to get the outcome we need. brooker.co.za/blog/2026/01...

Moving Towards Spec-Driven Development with Kiro and Agentic AI. Listen to @marcbrooker.bsky.social on The Real Python Podcast @realpython.com realpython.com/podcasts/rpp...

Finally, the session i have been most looking forward to at #reinvent2025! @marcbrooker.bsky.social discussing virtualization techniques used with services like #AgentCore - a 500-level Chalk Talk (with no slides). We need more of these types of sessions @awscloud.bsky.social

On the success of ‘natural language programming’ via @marcbrooker.bsky.social brooker.co.za/blog/2025/12... #programming #genAI #aI

New short blog post, touching on how I'd redesign UUIDv7 to keep the database performance benefits while mitigating most of the downsides. Short version: replace unix_ms with H(unix_ms, unix_ms >> 13 | salt). Get yer extremely niche RFC commentary here: brooker.co.za/blog/2025/10...

New blog post, reflecting on nearly seven years since the Firecracker launch, and how we're using Firecracker to power serverless databases (in Aurora DSQL) and infrastructure for AI agents (in Bedrock AgentCore). Here's the post: brooker.co.za/blog/2025/09...

NULL BITMAP: Using Huffman Coding for Sampling Discrete Distributions: buttondown.com/jaffray/arch...

Seattle friends - in case you missed it: lu.ma/umtmqb38 The Seattle Systems group is hosting their very first meetup! @marcbrooker.bsky.social will be presenting Aurora DSQL. Excited to see more systems groups popping up (thanks @eatonphil.bsky.social). My colleague @devhawk.net is co-organizing.

People often ask me about the differences in architecture between Amazon Dynamo (the 2007 SOSP paper), DynamoDB (the AWS serverless NoSQL database), and Aurora DSQL (the AWS serverless SQL databases). I memoized the response on my blog. brooker.co.za/blog/2025/08...

If you're into databases and you're in Seattle, @marcbrooker.bsky.social is talking at Seattle Systems on Aug 21: lu.ma/umtmqb38

LLMs are more powerful, more dependable, more efficient, and more flexible when deployed as a component of a carefully designed system. New blog post: brooker.co.za/blog/2025/08...

I had a great time talking to Ben and Will from Antithesis on the BugBash podcast about software testing and validation at AWS. In this wide-ranging hour we talk deterministic simulation testing, fuzzing, formal methods, the history and future of testing, and much more.

Fun conversation with @jkxosound.com about where I think things are going in the AI agents space, Strands Agents, and rats 🐀

Prototyping agents is easy, fast, and fun (especially if you use Strands). But how to you get them into production with all the scalability, security, and resilience you need? We've been building Bedrock AgentCore to answer exactly that question. Check it out: aws.amazon.com/blogs/aws/in...

I sat down with Darko Mesaros to talk about Kiro, specification-driven development, and putting AI-powered development in context of the history of computing. Is an LLM a compiler? What is a spec? How are specs like SQL? Check it out: www.youtube.com/watch?v=OdD9...

I've been using Kiro as my primary IDE for a couple months, and am super happy to see it launch. Spec-driven development is super powerful. More reliable, more repeatable, and easier to control than prompt-based vibe coding. Kiro's hooks are also pretty great. I use them to keep clippy happy :)

Finally got to fixing the "related posts" feature in my blog. The Amazon Titan embedding model on Bedrock, plus cosine similarity, does a nice job:

🚀 New AWS Bites episode! We tried Amazon #DSQL so you don’t have to (but you might want to) 👀 Luciano and Eoin put Aurora DSQL through its paces in #AWS Bites #Podcast Episode 145 🧵 Here’s what we found:

New blog post, with some career advice: brooker.co.za/blog/2025/06...

Here's the answer. On the left, we see the great "exponential benefit at linear cost" impact of redundancy. But that quickly tapers off as the single point of failure dominates. What can this teach us about building good HA architectures?

Wrote a monte carlo simulator in Python. Took 300 seconds to run. Asked Q CLI to translate it to Rust, compile, and run. Took 48 seconds, including the translation, and gave exactly the same result. Is Claude 4 the best Python compiler?

Availability intuition quiz! I have an architecture with N redundant components, each available 99% of time time, and one non-redundant component available 99.999% of the time. Sketch, on the provided axes, the end-to-end availability curve versus N.

New blog post, about some of my favorite papers from this year's HotOS program. Love Little's law? Hate fork? Think static analysis of shell programs sounds fun? Had enough of the f threshold model? Check it out for some papers to read: brooker.co.za/blog/2025/06...

"You will never experience eventuality. Eventual consistency is an abstract concept, not a guarantee you can count on."

Interested in how we test systems at Amazon? In a new article in Communications of the ACM, Ankush and I write about the evolution of systems correctness and testing practices at AWS. We cover classic and formal approaches, and new ones like deterministic simulation. cacm.acm.org/practice/sys...

Aurora DSQL is now Generally Available! I couldn't be more proud of the team that brought DSQL to life, and excited about how our customers are going to build on it. 🧵

In a new guest post from #AWS Sr. Principal Engineers @nikomatsakis.com and Mark Bowes take us inside Aurora DSQL's development from scaling write operations without two-phase commit, to overcoming garbage collection hurdles, and embracing Rust. www.allthingsdistributed.com/2025/05/just...

Last week I had the privilege to give a keynote at the International Conference on Performance Evaluation (ICPE'25). It wasn't recorded, so I captured the key points as a blog post: brooker.co.za/blog/2025/05...

Recently, one of our testing tools at AWS found an issue with transaction visibility in PostgreSQL, where the order that transactions become visible differs between the primary and replicas. Sergey Melnik dives deep on the AWS Database Blog: aws.amazon.com/blogs/databa...

I enjoyed @alexmillerdb.bsky.social's "Decomposing Transaction Systems" blog so much I had to write up how I'd decompose Aurora DSQL in a new blog post: brooker.co.za/blog/2025/04...

This is an excellent insight into the nature of transaction systems. Worth reading if you have any interest in databases.

I generally like Mitutoyo products, but this travel comb sucks.

Exciting to see another great step forward in the Rust standard library verification contest with the addition of VeriFast. www.linkedin.com/posts/r%C3%A...

@marcbrooker.bsky.social breaks down how they've architected a fully ACID-compliant database service that combines simple, serverless management with high availability and massive scale on AWS Aurora DSQL. youtube.com/shorts/dScUi...

It's widely known that sharing a queue across multiple servers, rather than queue-per-server, often helps reduce latency and improve utilization. But when is one queue better? In my new blog post, I look at one case: different classes of work. Read it here: brooker.co.za/blog/2025/03...

I had a great time chatting to John Kutay on the What's New In Data Podcast about Aurora DSQL, the architectural choices we made, and the trade-offs in the database space. Check out our conversation here: www.youtube.com/watch?v=Q3CY...

What Justin is talking about here is close to how Aurora DSQL's internal storage works. Even in our case (building SQL on top of the storage), pure K/V leaves a lot of optimization opportunities on the ground.

Love the explanation of achieving coordination-free local properties, ie, avoiding coordination, while still providing useful isolation levels at scale and across regions.

Wrote a blog post reflecting my thoughts on DeepSeek, NSF funding and system research communities in general. Apologies for the bold claims -- hope they can invite some discussions. blog.xiangpeng.systems/posts/system...

One challenge of handling partial/gray failures in distributed systems is telling 'healthy' from 'unhealthy'. Even in terms of error rate, it can take a surprisingly high number of samples to differentiate between normal and abnormal hosts. Here's what it looks like: