Camuel Gilyadov

yes, ofc ! radix-tree metadata is backed by slatedb.. no point in hand rolling custom impl there.. thx to you, @criccomini n maintainers 🙏 slatedb might not be suitable for actual large kvblocks - write amplification, compaction concerns.. wanna iterate n see.. nothing is finalized yet #L218" target="_blank" rel="nofollow noopener">github.com/Venkat2811/wom…

A month ago, Michael Malis started rewriting PostgreSQL in Rust with AI. PG is 1.3M LOC. At 250k LOC, 1/3 of 50k regression tests passed in 2 weeks; a week later, 450k LOC passed 2/3. The goal wasn't just to find out if AI could pull it off, but also to improve on PG limitations.

@sattyyouneed I remember this. It is not about LLM generating code but about weights and Software 2.0. concept.

Imagine being this early

@gwenshap @theo Files, pages, single-threadness

@theo What’s wrong with Postgres? 😂

Crazy how the web is still the only platform that doesn’t suck to develop for.

@apurva1618 what is the third thing?

Success like tpuf's is so hard to replicate. You have to get 3 things right at once: a focused product that solves a painful enough problem to get people to switch, and then has space to grow. We started Responsive 3 years ago, reached 1M ARR fast, and still pivoted. Responsive v1 had a focused product that solved a real problem for people building Kafka apps (hence the quick revenue). But the market wasn't growing and the wedge didn't open up. Responsive v2 (@opendatadbs) is the opposite. It's ambitious, has loads of space to grow, and we have 3 competitive databases on a common foundation already. But an ambitious vision, even with a correct technical thesis behind it, is not an automatic reason for anybody to switch. Comparing Responsive v1 and v2's problems, v2 is the better one to have: we are in an obviously growing space with a lot of optionality, so at the very least we are likely to fail better.

Looked at this way, LLMs are the triumph of symbolic AI, just with more noise.

tpuf is a super interesting case study. 19 months ago there were like 10 vector database startups that had explosive initial revenue and then had started to peter out in the 5-15m ARR range Consensus quickly forms that vectors aren’t a standalone database category and that existing databases would just implement a vector type and that would be the end of it Then tpuf enters the scene and does 1 —> 100m while raising 20-100x less than the competition Does anyone even remember the names of the other startups? I don’t of course tpuf is no longer just a vector database, they’re going deep on general purpose search now, but still just incredible stuff Clickhouse is another exemplary case study

Yep, they totally bootstrapped it, well done! Aren't WarpStream similar, they also got to a good traction without much external funding. QuickWit were lean too. That said I think calling them all a managed database perhaps a bit exaggerating? No SQL, no optimizer, no executing arbtrary plans, no schemas and etc. Developing a full natively diskless HTAP SQL DBMS that the industry needs so much will be a real test.

@camuelg @apurva1618 thats exactly why they break my mental model. they're profitable WITHOUT significant VC funding. they don't need the crazy high margins, which is exactly the point.

I've written before on the broken economics of the database market, and while I'm thrilled for turbopuffer this tweet shatters my world view. My theory (w/o the nuance) was that you can't run a profitable company selling managed databases on cloud hardware long-term. I wrote about it here: bitsxpages.com/p/the-broken-e… @Sirupsen does exactly that, and there's signs that it's going to last long-term too. Not sure what the magic sauce is, it's probably a combo of: - object store native infra is actually easier to run with favorable cost structures on the cloud - staying lean with a ridiculously high talent team - they landed some huge customers early on, who are themselves growing bananas - AI makes it possible to stay lean long-term and automate a lot of business process Kudos to the turbopuffer team, it's probably time to rewrite my "Broken Economics" post just a few months after releasing it.

@almoggavra @apurva1618 We've barely unlocked the power of diskless. Incumbents can't pass the savings on to customers because taking an immediate revenue hit is unacceptable. It’d be interesting to look at their gross margins; I bet they aren't as high as VCs demand.

@apurva1618 Agreed, many have attempted this path but few succeed.

July 22 in Denver, CO @ApacheDataFusio meetup: Join myself, Andy Grove, Brent Gardner to find out about DataFusion, meet other users, and talk about Databases. We still have speaking slots left (10-15 minutes) and would love to have you join us luma.com/jsu6faie

Forget Turing machines. Finite automata are all you need.

@julianhyde Lol... the ai scaling will stop working one day

WD-40 is amazing. Just imagine what WD-50 will be capable of.

Free will in 2026

@almoggavra Latency?

how many systems could be replaced with an S3 file now that If-None-Match is available? Was just chatting about kafka schema registry and this came to mind... one file per schema seems braindead simple.

The next 12 months will be dramatically better for infrastructure companies upstream of Anthropic and OpenAI than for application-layer companies downstream of them.

there are more startups than ideas

@Trace_Cohen The shadow IT

The most valuable workflows inside a company are usually invisible to outsiders because they exist in spreadsheets, tribal knowledge, Slack messages, approvals, and exceptions.

This reminds me of computerization. The amount of "work" people could execute on computers increased by a huge factor, but their productivity did not. The amount of work "needed" to arrive at the same high-level outputs exploded.

The essential engineering cheatsheet of 2026: agent → while loop subagent → nested while loop agent harness → the rest of the code cloud agent → all the above, on EC2

Headless is such a nice rebrand for API lol