Krishna Vishal

That's me. From being an ML Engineer to contributing to a DB project. Never imagined this timeline but here we go.

Rust is thriving at @TheASF 🦀 🥇 Apache DataFusion - 8.5k+ ⭐ 🥈 Apache OpenDAL - 5k+ ⭐ 🥉 Apache Iggy - 4.1k+ ⭐ The future of open-source data infrastructure is written in Rust. iggy.apache.org #Rust #OpenSource #Apache #DataEngineering #Streaming #Iggy

@vaibhaw_vipul Happy birthday, Vipul!

Turned 31 today, here are the few points I would have told my 24 year old self. Life isn’t random, even when it feels like it is. What you do today will come back to you, in ways you don’t always see immediately. Your actions have consequences and over time, they stack up. So don’t take your choices lightly, even the small ones. Also, at some point, you have to figure things out on your own. Learn how to learn. Be curious, be resourceful, and don’t wait for perfect guidance. That skill alone will carry you further than anything else. Another thing, Starting over? It’s not failure. You’re going to have to do it more than once. Careers, relationships, ideas - sometimes things won’t work out. That’s normal. Most importantly, be patient. Real growth is slow. The things that truly matter - skills, money, relationships, respect - they compound. But only if you give them time!

The threats are no longer working. I'm scared.

Very cool!! What is the thread count used here? Because I've found that as you vary thread counts parking_lot becomes slower than std::mutex after a certain point. Also is there a possibility to share MWE repro of this? I've a WIP mutex implementation that should shine in high contention and short critical section case and I want to benchmark it in this case.

I had a beautiful opportunity to use a (userspace) spinlock. The critical section was very small (like one line) but high contention. I thought a spinlock would be perfect, but it turns out it isn't! Spinlocks come with their own T&Cs, suitable for only specific and few workloads. I benched with parking_lot, which spins first and only then tries to acquire the lock. This turned out to be the best compared to a spinlock and std mutex. I also found matklad's excellent blog post supporting this. There goes my dream of using a spinlock in a production database. 🪦

@penberg Turso lore drop.

What are Turso, Turso Cloud, and libSQL? I get this question a lot, and AI agents seem confused about it, too. tl;dr: Turso is a new SQLite-compatible database, rewritten from scratch in Rust, currently in beta. libSQL is our open-source fork of SQLite, which powers Turso Cloud today. Turso Cloud is our managed database service, and we're actively working towards adding Turso to it. Still confused? That's fair. The names evolved as the product did. Here's the full story in chronological order. Back in October 2022, we announced a fork of SQLite called libSQL. At the time, we were building an edge-replicated database based on SQLite for our internal use. However, we faced a major problem that there was no way to upstream the core enabling technology, WAL virtualization. We decided to fork SQLite and make it open to everyone's contributions. With the fork out of the way, we kept building what would become Turso Cloud. One of the most memorable moments was when ThePrimeagen tested our database on his stream sometime in February 2023, while we were still in private beta. The whole team was watching the stream and monitoring our systems, hoping everything would work out fine. In April 2023, we launched a public beta. With our own fork of SQLite, we spent a lot of time early on trying to address one of its biggest limitations: the single-writer transaction model, which limits scalability and isn't the greatest DX. Me, @sarna_dev, and @iavins built a Hekaton-based MVCC implementation and a prototype of libSQL supporting BEGIN CONCURRENT. However, the integration wasn't great, and there was a lot to do with Turso Cloud, so we shifted priorities there. During that same time, I began exploring rewriting SQLite. I initially used Zig, thinking I would reuse large parts of the SQLite front-end, for example. However, I also wanted to make the database core fully asynchronous with io_uring support, and it turned out to be very hard to reuse any of the SQLite code for such a fork. With some skill issues during comptime (as ThePrimeagen and Jarred Sumner once framed it), I ended up ditching Zig and, in late August 2024, switched to Rust. I called the project Limbo at the time, but later it was renamed Turso. The whole thing was more of a side quest than anything at that point, mostly exploration. In early 2024, we had a bunch of people asking for vector support in libSQL. @SivukhinN added support for approximate vector search, similar to pgvector, but we had to compromise on the DX because the changes to libSQL would have been very intrusive. Around this time, we realized that forking SQLite would only take us so far, mostly because the main test suite SQLite uses is fully proprietary, and we had not been able to obtain access to it. After some back-and-forth, we ended up making my SQLite rewrite side quest public. Surprisingly enough, early people like @Peristocles1, @jussisaur, and @PThorpe92 showed up and started contributing to the project. As we saw more uptake, we made it an official company project in December 2024, renaming it to Turso. (As a side note, we did end up rewriting the libSQL server components using TigerBeetle-style DST architecture with S3-based diskless architecture in August 2024 and migrating from Fly to AWS. But that part fortunately added very little confusion to users, as it's mostly a Turso Cloud internal detail.) Today, Turso is in beta with early customers working toward production deployments. The libSQL fork has been an incredible success with 800k weekly downloads for the JavaScript SDK and running production workloads at scale. We're actively maintaining libSQL, but the future is Turso. Built from scratch with no proprietary test suite in the way, Turso supports features like concurrent writes that a fork of SQLite never could, and we're working on bringing it to Turso Cloud.

Compiler is correct 99.99% of the every time.

Happy to see that our recent blog post about thread-per-core & io_uring migration has been mentioned by @ThisWeekInRust this-week-in-rust.org/blog/2026/03/0…

@jussisaur Did you ever find out who was sprinkling in these sleeps lmao?

@EigenVectorizer i fucking shit you not, at my last webdev job one of the biggest user-facing perf improvements during my tenure was removing an inexplicable sleep(1000ms) from the spinner component's default properties

working on query planners is great because you can change a few lines of code and get a 50x, 500x or 5000x perf improvement the fact the original query plan was eye-wateringly horrific is irrelevant. 5000x bro. if ur not making 5000x improvements you're NOT going to make it bro

Thread-per-core + io_uring? See how we moved Apache Iggy from Tokio's work-stealing to a shared-nothing* architecture to make the most out of modern hardware and provide ultra-low, predictable tail latencies. iggy.apache.org/blogs/2026/02/… #io_uring #iggy #rust #streaming #apache #asf

@jussisaur peak male fantasy.

cant wait to be laid off and go to work on a farm shoveling pigshit so i don't have to read one of these you are falling behind FUD posts ever again

@DeepknowledgeU Gaddamn bruh. Baasha level flashback indeed.

Never bothered to preserve certificates or medals but my mom managed to preserve a few of my meals from my running days. 2 Ultras, 3? Marathons, 4 21ks (and countless unofficial runs). While others were grinding LC, I was grinding em miles

@PThorpe92 Average CS program in US has turned into a scam anyway.

The takeaway was supposed to be that the knowledge is the important part... not the degree. There is infinite resources on the internet to teach yourself whatever you want.

@Finance_Weights It is quite high. Do a retest in a week. And also get a test for hs-CRP, it's a marker for low grade inflammation. If hs-CRP is elevated in multiple tests then increased cholesterol is to be taken even more seriously.

Anyone understands how to interpret this? ChatGPT simply says consult a cardiologist. ~My body fat is not too high (16%) ~LDL cholesterol is very high 185 ~Triglycerides are amazingly good (44), almost falls in elite levels Do I need to consult one or not ?

Yuuge performance gains on the way for tursodb. Been deep in the planner/optimizer for like a week now.

@amuldotexe ROFL

Claude code in the hands of a PM

@PThorpe92 Merry christmas homes. How far have we come since last December.

Spending the day doing some refactoring and performance work. This is the last Christmas that I'll have to spend away from loved ones :) 133 days

Trying to squeeze 5GB/s from localhost has turned into a deep learning exercise. The best I'm getting is 1.5GB/s.

Performance Hints Over the years, my colleague Sanjay Ghemawat and I have done a fair bit of diving into performance tuning of various pieces of code. We wrote an internal Performance Hints document a couple of years ago as a way of identifying some general principles and we've recently published a version of it externally. We'd love any feedback you might have! Read the full doc at: abseil.io/fast/hints.html

The lack of "feature design" is why so many products over time feel hollow or messy. This isn't visual design. This isn't architectural design. I thought that a short video lecture of what feature design is and a real case study of applying it in Ghostty would be helpful. Feature design is the planning step behind how you're going to solve one or more user problems with a product feature: what that feature looks like, how it feels, and not just how its going to tactically solve these specific problems, but how that solution is going to interface with the edges of other features that currently exist or are planned to exist in the future.