Jason Gorman

Testing is often an unnecessarily emotive subject in the software industry. I particularly recall a trip to a 'Software Craftsmen' London meetup in the early 2010s - the culture disintegrated into a shark tank when a brave soul dared to even slightly question (respectfully) the received wisdom du jour. As a fresh graduate, was I then going to ask any questions I had? Was I heck! I didn't go back, safe to say. (I instead found the London Clojurians, a much friendlier bunch - my kinda people 😊) Fast forward to today: I wanted to define a more objective language to talk about testing - a recognition that, despite all of the religious fervour, there can be a pragmatic middle ground that adequately mitigates the risks at a reasonable cost. So, rather than vague lamentations around 'we need time to improve our tests', 'we don't have enough tests', etc, the dimensions in this JUXT article give us a more precise, objective language to talk about the value and costs that our tests bring, prioritise certain areas over others, or quantify our improvements. https://lnkd.in/eu8rp8Wa Have I missed any big ones?

13

9 Comments

And now for some readings of user level RCU .. and a landmark paper that led to the implementation of liburcu - the user space RCU library, a Christmas Day evening read .. Why RCU is difficult in user space ? RCU, particularly its high-performance Quiescent-State-Based Reclamation (QSBR) variant, is easier in kernel mode because the kernel scheduler automatically detects quiescent states whenever a CPU context-switches, enters user mode, or idles, allowing grace-period tracking without any explicit cooperation from kernel code. In user mode, applications lack this built-in mechanism, so threads must explicitly register and periodically report quiescent states (e.g., by calling specific functions), imposing invasive global constraints on the entire application. These requirements make user-level RCU harder to adopt broadly, as they complicate library design and require modifications to all potentially reading threads, which is impractical in many user-space programs. User level implementations of RCU : This paper contributes to user-level RCU by formally describing efficient and flexible implementations that overcome the limitations of prior approaches, which either imposed high read-side overhead or severely restricted application design. It presents multiple classes of RCU (including QSBR, memory-barrier, signal-based, and bullet-proof variants) with detailed algorithms, performance analysis, and comparisons to locking, directly forming the foundational basis for the liburcu library's core flavors and enabling its widespread adoption in user-space applications.

73

3 Comments

Recording is available on Ministry of Testing website. "This session is part practical demo, part conversation, and full of insights into how you can think critically about inputs, outputs, boundaries, and system behaviour. Ben and Nat also reflect on when to use these techniques, how to combine them, and why even the basics of testing remain valuable in modern contexts." https://lnkd.in/dq3BVi7z

8

Balancing speed and safety with AI tooling. Josh Nesbitt and I discuss if Test Driven Development (or Behaviour Driven Development - remember cucumber!?) is actually more important now than it ever was.

20

4 Comments

Continuing my series of posts attempting to (pseudo-)formalise the practice of Test-Driven Development, I want to look at a “rule” of TDD that’s actually helpful whether you’re going test-first or test-after. (Link in comments so as not to anger the algorithm gods)

16

2 Comments

A director of software engineering asked me recently what good coordination between PCCP and MLOps actually looks like. He was struggling because despite having a PCCP and a MLOps pipeline, progress felt slow and brittle. This was a result of misalignment. When teams treat the PCCP as the governing contract and MLOps as the execution layer, alignment becomes much easier. Here’s what that looks like in practice: 𝗔𝗹𝗶𝗴𝗻𝗲𝗱 𝗰𝗵𝗮𝗻𝗴𝗲 𝗰𝗼𝗻𝘁𝗿𝗼𝗹. A shared framework that makes it obvious which model, data, and performance changes are already authorized under the PCCP and which ones require additional review or regulatory involvement. 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝗱 𝗲𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗮𝗻𝗱 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲. Your MLOps pipelines enforce the PCCP rules in real time. Every approved change generates the validation evidence and traceability you need, without teams stopping to build documents by hand. 𝗖𝗹𝗼𝘀𝗲𝗱 𝗹𝗼𝗼𝗽 𝗺𝗼𝗻𝗶𝘁𝗼𝗿𝗶𝗻𝗴 𝗮𝗻𝗱 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲. Training data, model versions, and real world performance are connected end to end and feed directly back into the PCCP risk controls. In practice, this only works when PCCP and QMS requirements are embedded in the tools engineers already use, not bolted on afterward as a documentation exercise. I’m curious how others are approaching this. How are you managing model updates in MLOps, without or without a PCCP?

29

Clearing System Design interviews at the Staff+ level remains a mystery. Even if you go for best-in-class courses, like HelloInterview by my colleague Stefan, you're often left on your own to show your experience. Sys Design is hard - you need to master foundations, manage time and comms, make solution future-proof, etc. Common prep materials make it worse with overly complex examples like social platform ranking systems or Dropbox-style sync protocols. The truth is: on Staff+ interviews they ask relatively simple questions to test depth, not speed. You won't waste time drawing 20 components - they want to see how deeply you think instead. I don't know the full formula to clear Sys Design rounds at this level, but can share what was working for me personally: 1/ Learn the fundamentals: things like GFS/Dynamo/RocksDB papers, behavior of distributed systems under high load, data locality, over-provisioning, clocks, replication lag, consistency models, use scientific notation, etc. Example from Anthropic Sr Staff interview: - Me: "I don't remember the exact metric name, but we trigger provisioning when we detect this resource utilization pattern." - Interviewer: "Oh, we call it metric X, and pattern Y, and the way you're describing it is how it works in our system." Pro tip: share some anecdotes, e.g., "we don't need persisted state from the get-go, but it was a nightmare to revoke stateless sessions on my previous project". 2/ Become good at sizing the problem. Know the population of key countries/markets, understand seasonal usage patterns across days/weeks, distributed systems fundamentals - like maximum dataset size for RDBMS, DB index overhead, key latency numbers, etc. Example from Google L7 interview: - Me: "Let me quickly quantify the system. For this problem, I'm estimating the population size to be in the mid-hundreds of millions, let's say it's 500M, but I will round it up to 1B." - Interviewer: "Actually, in my reference notes I have 500M, so let's keep it like this, so it's easier for me to compare the final numbers." - In the end, the final numbers checked. Pro tip: casually drop some fundamental reference, e.g. "By Little's Law, RAM won't bottleneck this cluster given our latency and throughput targets". 3/ Start and finish simple. Most real systems avoid expensive services like Kafka, complex client/server protocols, or real-time UI updates. Example from my current role's interview: - Me: "I'd just hash the input and denormalize the schema to avoid excessive DB upserts. Not the fanciest approach, but it works". - Interviewer: "Oh, this is exactly how our system actually works". Pro tip: research the company's tech stack, and don't be shy if you're not familiar with it: "I haven't used GCP, but we solved similar problems with AWS service X".

110

5 Comments

👻 Kiro 👻 has had a whole host of great new features recently. There are the really big ones like Skills support and the constant evolution with new models like Opus and Sonnet 4.6, but one of my favourites is ACP... The Agent Client Protocol lets you take advantage of one of the best coding assistants 💪 around and use the underlying agent in a variety of different clients. I've been exploring how you can take advantage of Kiro for use cases that aren't all about writing code 💻 and how you and your teams can all reap the benefits from its huge capability. I've got my latest post in the comments on some of my favourite new ways to use Kiro with ACP. Let me know what you think. 🚀 #kiro #genai #kirocli

27

2 Comments

I mentioned one of my “down the pub” software development opinions to James Lewis at EventCentric. He suggested that I mention it to Kent Beck and so, while I remember. But definitely, the tone here, is “crazy talk over a pint.” ———- XP optimises software development for folks with ADHD. The Extreme Tuesday Club is likely a hotbed of ADHD diagnoses. TDD, CI, Small Releases: we need those rewards to keep coming in regular doses. Pair Programming: “Stay on target, stay on target” Refactoring, Boy Scout Rule: “Ooh, while I am here, I should fix that as well. Now where was I? Oh the test list.” Sustainable Pace: We can only focus so long in one stretch, even with the above. Planning Game: one story at a time, don’t get distracted by other things. Business value, not “shiny thing.” Small stories because we get bored easily. Whole team working so that we can get to that shiny story faster. Osmotic Communication: Quick, let me blurt out my idea before I get distracted and forget I wonder if there is a correlation between XP teams achieving flow, and XP teams filled with neurodivergent people who have figured a coping mechanism for productivity.

312

67 Comments

We’ve been trying to find a senior software engineer with good TDD skills. The process includes a coding exercise, currently remote. Recently we had a candidate who started well but got stuck as the exercise got more complicated (to be clear, this is a very basic domain with no algorithms involved). The strange thing was they couldn’t understand the code they’d written so far and kept looking off to the side. We suspected they were using a bit of LLM help. It all crashed rather badly. The irony is that we don’t mind if people do use AI, as long as it’s open and they can explain what they’re doing. I’m also wondering what they thought would happen if they’d got the job. The way we work, any such dependency would have shown up on the first day.

216

33 Comments