Developer at YDB, passionate about performance. Interested in databases and distributed systems. Opinions are my own.
Joined August 2021
- Tweets608
- Following203
- Followers569
- Likes3.9K
Pinned Tweet
io_uring easily beats AIO and gets faster with every kernel — until both suddenly get 30% slower.
Join a database developer’s unexpected journey into the Linux kernel and IOMMU.
medium.com/ydbtech/how-io-ur…
I'm actively writing C++ code which uses both Actor model combined with C++-20 coroutines. For additional inspiration rewatched a great talk by @benh about Actor combined with futures/promisses vs. Actors and lazy continuations. Highly recommend the talk: a very useful complicated things are explained in a very simple way.
youtu.be/P6Y-Z1uPp4c?si=3kW-…
Evgenii Ivanov retweeted
excellent read
ARM or x86? Your choice has a big impact on Postgres performance.
Here's what to know before you pick:
planetscale.com/blog/when-to…
Evgenii Ivanov retweeted
Powerful people imagine a future. Organizations reduce it to KPIs. Everyone optimizes for the numbers, and the world becomes what we measured instead of what we meant.
Evgenii Ivanov retweeted
Have you ever wondered what a data race is? How a race detector works? If there are bugs in your race detector? We take a look.
Paywall has expired, give it a read!
theconsensus.dev/p/2026/09/0…
"unknown future" - nowadays, that's for sure :)
At least these 150 days of suggested NVMe testing are within its lifetime warranty :)
I asked Astra to polish my fio scripts, which I use to evaluate NVMe devices. Its findings sounded reasonable, it did the job, started full run, I switched to Luna to follow the script execution.
After ~ few hours the benchmark was still running. Luna estimated its execution in half of year (!!!), while Astra thinks 1 month is enough (!!!).
I asked Astra to polish my fio scripts, which I use to evaluate NVMe devices. Its findings sounded reasonable, it did the job, started full run, I switched to Luna to follow the script execution.
After ~ few hours the benchmark was still running. Luna estimated its execution in half of year (!!!), while Astra thinks 1 month is enough (!!!).
Also, it removed all the useful stuff:
"The review confirms several major regressions beyond runtime: I removed your repeat-and-median workflow and warmup, applied 80% occupancy only to the medium profile, and made sustained tests use the entire target. The tests even assert that 255 separate preparations should occur—they validated the wrong design. I’m checking the statistics and profile definitions next so the proposed fix addresses more than the timeout."
"You’re right. I implemented conditioning before individual test points and a repeated full-capacity write requirement
that made the runtime wildly exceed your intended benchmark. I’ll review the code against your original scripts and
your stated runtime, identify the methodological and reporting mistakes, and propose concrete fixes with a realistic
time budget. The device run is stopped; this review will be read-only."
Evgenii Ivanov retweeted
Even if an AI did exactly what I told it to do, precisely and without error, I would still work hard as an engineer.
Brooks told us why in 1986: "Even perfect program verification can only establish that a program meets its specification. Much of the essence of building a program is in fact the debugging of the specification."
What he meant is that engineering is not the execution of commands... It is very much design. And design is ridiculously difficult.
You need to decide what you will build and how you will build it.
Of course, you can outsource the design to the AI. But that's assuming that the AI knows what must be built.
Does it?
"I, Asimov" is an incredible book. I love sci-fi, and Asimov has always been one of my favorite authors, but this autobiography made me admire him on a whole new level.
It’s the true story of a man who dedicated his life to writing. After reading it, I admire Asimov not only as a writer, but as a human being: for his curiosity, discipline and passion, humor, and sheer love of his craft.
And IMO, the book is seriously underrated on Goodreads.
Evgenii Ivanov retweeted
The Consensus tracks over 120 of the best developer blogs, and that number grows each week. There's now a dedicated feed page for this group. And RSS as well. No subscription needed.
theconsensus.dev/feeds/devel…
Can I/O outlive a dead process? With io_uring, yes. A killed and reaped writer can still modify storage. Here’s how - and how to stop it.
blog.ydb.tech/is-there-i-o-a…
I’d love to hear from @axboe on this one :)
Can I/O outlive a dead process? With io_uring, yes. A killed and reaped writer can still modify storage. Here’s how - and how to stop it.
blog.ydb.tech/is-there-i-o-a…
The book is awesome, highly recommended 👍
A user asks a question, and an LLM directly searches through your data sources to put together a response. This search approach has become widespread.
But remember that you can pre-filter the data for the LLM with good old full-text search and similarity search. This way, you’ll burn fewer tokens and put less garbage into the LLM’s context window.
If Postgres is one of your data sources, then the Just Use Postgres book shows how to take advantage of various search capabilities, including full-text and similarity search.
An LLM is an expensive and nondeterministic search tool. Know how to use additional search capabilities to reduce the cost and get more predictable responses. The book is 45% off on August 30th: hubs.la/Q03-d27Y0
