Dmitriy Ryaboy

Love this term "cognitive debt" as explained by Margaret Storey in margaretstorey.com/blog/2026/02... and reshared by @simonwillison.net in simonwillison.net/2026/Feb/15/... . I suppose changescape.dev is me grappling with how to hold cognitive debt at bay.

Working on a bit of Rust code and it's so confusing how this faster and more efficient language is so much slower to iterate on and sends my laptop fans spinning so much harder than the Python & TypeScript stack I've been spoiled by.

Last decade's tools are the wrong paradigm for dealing with the speed and volume of changes happening to this decade's codebases. We need a better changescape: changescape.dev

That awkward moment when your tool doesn't find 25 planted bugs it was supposed to find in a benchmark, but does find 5 that were there unintentionally.

Sigh. So what actually moved the needle was fixing their cdc and not writing so many updates to the upstream in the first place. Which probably would've made BQ work just fine too. It's so often not the migration but the rearchtecture it forces. www.bolna.ai/blog/bigquer...

The founding of the East India Company, in 1600: "a barrel of beer a day was authorised ‘for the better holding together of the workemen from running from ther worke to drinke’." 400 years later, tech offices pretty much running the same playbook.

Kind of embarrassing how many ML people are reacting to Jev with "lol, LLM noobs discovered classifiers / BERT".

Viking laws, man. (Well, technically, post Viking. 11th century Norway).

Write about how software engineering is changing inside your company / team / project. Don’t use AI. Win $10K and get published in The Pragmatic Engineer. You have a month: submit by 4 October. Details: blog.pragmaticengineer.com/how-software...

Totally wild that they gave Terminator a "neural net cpu" back in 1991

Claude consistently making code 25% longer when I tell it to DRY things up is equal parts frustrating and hilarious.

@phokion.bsky.social just finished Whalesong. Excellent as always! The wait for #4 begins.

Dealing with bare metal providers and, man, I remember now why "cloud" vendors get away with their margins. Any old Twitter people remember NTT America?

Just parse the AST, they said. It'll be easy, they said. SQL is a pretty small and intuitive language, they said.

Me: "Problem is, LLMs rarely give a straight call. They equivocate." LLM: "You are half right."

I knew SQLGlot is a solid sql parser, but it wasn't till I started working on dblect that I learned how powerful it is. medium.com/@squarecog/t...

We need a new shorthand: TS;DR (too slop, didn't read)

Who would've thought that the 2 fastest python SQL parsers would be created by 2 different guys named Tobias M.

dblect 0.1.0 released! Test your dbt workflows and find hidden dangers that linting and unit testing alone won't warn you about. Pypi: pypi.org/project/dble... . Github: github.com/dvryaboy/dbl... Intro blog: medium.com/@squarecog/t...

Introducing dblect: medium.com/@squarecog/t...

A cool thing about the dblect tooling is that you can `noqa` it inside macros. Something sqlfluff doesn't support right now, but i found it necessary for significant dbt usage.

I'm getting pretty close to the 0.1.0 release of dblect. If you are into the whole actually testing your dbt workflows thing, check out the README: github.com/dvryaboy/dbl...

I've long held that people underestimate the power of just grepping the codebase. And boy did claude code prove me right... 😅

Meet the SEMI JOIN. A join in the streets, a filter in the sheets.

Rough order of planned `dblect` releases: 1. Structural hazards: static analysis across dbt models 2. Logical types + contracts: eg no summing amounts without grouping by currencies 3. Var/flag worlds: awareness of flag values that change generated SQL 4. PBT: generate examples and test execution

If you build excel file support by pretending it's just a csv wrapped in xml, you're gonna have a bad time.

OMG my baby tool `dblect` is actually somewhat useful, check this out @matsonj.com : github.com/matsonj/nba-...

I'm working on a tool, dblect, to help test dbt project changes for bugs in meaning, not just sql validity or data diffs. Last weekend it got the ability to call out bugs like summing up money amounts without grouping by currency. This is good stuff, man.

Anthropic was practically daring US gov to apply export controls.

Hypothesis is catching such beautiful and weird edge cases. Saving me a lot of future embarrassment. Thanks @drmaciver.bsky.social and the rest of the team!

I'm writing something that lets you say "the columns amount:int and currency:varchar make up the type Money". The right name for that is probably SemanticType, but I'm concerned folks will get confused and think this is connected to Semantic Layers. Does DomainType work?

Two days ago, I joked "Codeim, I mean Windsurf, I mean whatever it is called now - probably Devin." Yesterday they actually renamed it to Devin. What should I do with my newly discovered powers?

FiveTran's m&a strategy to basically own the airspace above BigQuery and Snowflake is kind of fascinating.

The phrase "earning its keep" is no longer earning its keep.

What if you could have nice types on top of relations, with type inference through your sql workflows. With type refinements. And you could write correctness contracts on them. And it was statically checked. And maybe you could also generate PBT style tests on top. Would that be nice?

Types are hard, databases are hard, lineage and type propagation on top of database tables is hard... It's all hard, man.

Pop quiz, how many different sql parsers do FiveTran products run now?

You all should know there is a 2011 paper called "Don't Thrash: How to Cache Your Hash on Flash" by Bender et al.

It's been like 15 years since the Twitter API started putting ids as both numbers and strings cause js sucks, and yet there are Real Serious Companies still tossing 64 bit int ids around in json, like loaded footguns.

Dbt is pig without piglatin. Fight me.

Old sigmod and vldb papers are gold. I'm working on a thing at the intersection of program analysis and db theory (aka pbt for dbt) and it's hella fun. What a gem the Olston, Chopra, Srivastava 2009 sigmod paper was. Undervalued.

This is a pretty intriguing idea for future proofing file formats. It does assume wasm is future proof, of course, but that feels like a safer bet than "assume readers are updated"

If you love this sort of thing, read up on C-store, which introduced this idea in 2005 and commercialized it in Vertica. Stonebraker, Sam Madden, Daniel Abadi. Parquet was also partially inspired by Vertica (and Google's Dremel, and PaX by Natassa Ailamaki et al) :-).

ML is just applied stats. Stats is just applied algebra. LLM is just ML backward and with an extra L.