Zod is the default validator for https://github.com/gajus/slonik.

Zod alone accounts for a significant portion of the CPU time.

> In the context of the network overhead, validation accounts for a tiny amount of the total execution time.

> Just to give an idea, in our sample of data, it takes sub 0.1ms to validate 1 row, ~3ms to validate 1,000 and ~25ms to validate 100,000 rows.

Surprised not to see more people ask about performance profile of v4.

Zod is great in terms of API, but a no-go in terms of performance.

We ended up writing babel plugins (https://github.com/gajus/babel-plugin-zod/) and using it along with zod-accelerator, which improves performance, but breaks in various edge-cases.

I guess people are not asking because the article contains benchmarks

What's a relatively small dataset?

For someone that could be 1m records for others that could be 1bn records

Below 10M documents for single node. Below 100M documents for clustered setup. Total data size (including indices) that can comfortably fit in available RAM.

TLDR

> So, we now focus on the new NIST standards of FIPS 203 (ML-KEM), FIPS 204 (ML-DSA) and FIPS 205 (FLH-DSA).

With what hardware? These PQC standards are still slow and definitely not ready for deployment.

That’s a lot longer post than I anticipated on the subject.

Just so I understand the other side of the argument: where is the purported inefficiency of keeping DST coming from?

According to this article, a rather weak argument about people not liking to wake up before sunrise based on questionable correlation of commute times to sunrise times, ignoring factors such as average commute lengths, dominant (historical) industries, effective natural light at different times in modern housing.

From that it makes an (incorrect) assumption about the value of AM sunlight over PM sunlight and declares that all-year DST is pointless.

In my opinion the only argument against all-year DST which holds any water at all, and even then not much, is the concern about kids going to school in the dark. However, since many places don't have enough winter daylight to go around, trade-offs need to be made and kids are probably better off on-net having daylight time during their free time instead of while eating their toast inside and commuting to school.

Kids being forced to start school in basically the middle of the night is another especially American phenomenon that requires a separate solution, I feel.

What solutions have the non-Americans developed to deal with it?

Start school at a later time? Have other types od take home work if a day is shortened as well?

What’s better/different/same from many tools in this space like langfuse etc

This is specifically directed at agent traces and not necessarily other LLM use cases. We work on a lot of automated analysis and error detection mechanisms (see https://github.com/invariantlabs-ai/invariant/tree/main/anal...) on such agent traces, which can be nicely shown and highlighted in line with the trace in Explorer. Also, agent builders value collaboration a lot. They send around traces like pastebins to point out specific issues and failure modes of their agents. Explorer makes it easy to point to specific points in long traces and annotate them.

The idea of llm generating flashcards is kinda neat. Not an area I am personally exposed, but I can see use cases for this

thanks a lot! yes, the idea was to really dial in on features that can help people learn faster!

All fair insights and learnings. Good luck with your next one

All interesting and rational points. Thanks!

What's the SQL equivalent of this?