Got it, thank you!

Lovely. And then once found, it's reproducible? Does adding additional debug information break the repro?

What flavor of stuff are you finding?

I think that they've done such a good job with the software that I don't miss Android Auto (supporting CarPlay alone gets you 58% of the US market and leaves out 41%). As long as they continue investing to stay ahead, it's the right decision, imo.

Needs something like Community Notes to help deal with misinformation. www.lesswrong.com/posts/sx9wTy...

But yea, lots of that scaling is premature, so! 😅

If the data that is going into the calculation is within the same Postgres instance, then transactional boundaries are preserved... if it's not, and you're making network calls, then totally. At scale (either for team boundaries or performance) personalization data is likely to be living elsewhere.

I like these buckets, but one exception here is when ordering by recency: immediately after creating something, regardless of result size, you expect to be able to find it in time order.

We don't yet, but I suspect that we'll support re-ranking directly in ParadeDB. Elastic has a simple DSL for some re-ranking use cases: for us, the natural fit would be arbitrary (PL)SQL.

Rust extensions are also easier than ever to build: defining a fast(er) UDF to compute a score would work too.

It's also important to remember that your first pass of ranking is still very important: you can only re-rank whatever amount you over-query by in that first pass (maybe the top 10k), then re-rank them.

This is true for some of the highest scale search systems, yes. But we've also seen it applied for the wrong reasons: for example, team boundaries where one team owns search infra and another that owns re-ranking and/or faceting.

I'm fine with the popover, because it's very difficult to preserve formatting when translating: and if you can't preserve formatting, then doing it online seems less valuable.

Until you compact an undo log or WAL by applying it to create a new snapshot! Not all use cases can afford to keep history forever.

You can definitely do this with Vortex, although I'm not sure whether it does it by default for strings.

Hooray! Thank you.

In Python with mypy (which is only one of three+ type checkers), there is an option that you can toggle to allow calling untyped from typed.

We're all in awe that we didn't know this existed. Thanks!