What if SELECT, FROM, WHERE were functions?
remy.wangI've been looking into Prela the last few days and really like it :)
Conceptually, the two key operators `.select()` and `.and()` connect somewhat to the lineage of arrow notation [1], fork algebras [2], etc.
Left-to-right composition:
(>>>) :: (b -> c) -> (c -> d) -> (b -> d)
Fanout:
(&&&) :: (b -> c) -> (b -> d) -> (b -> (c, d))
The database-specific detail is enumerating the universe of entities to maintain shared correlations and using projections as ways to access information, so, in the above notation:
movie :: Movie -> Movie
title :: Movie -> String
movie >>> title :: Movie -> String
The specialization to then have Rust types be able to generate inlined query execution code at compile-time is quite neat. But, I wonder where the limitations of this approach are seen? In contrast to something like Soufflé which has a much more complicated optimizer but generates C++ code ultimately, it seems difficult to always rely on the Rust compiler. Likewise, some other projects like Crepe rely on Rust macros more heavily. And, the embeddability is a plus, but it does not look too easy currently to use from other languages as you would need to compile the query independently and provide something like an FFI wrapper.
Relatedly, some other posts about relational language design from the last couple days: https://news.ycombinator.com/item?id=49342530, https://news.ycombinator.com/item?id=49363617
And this tutorial looks interesting: https://northeastern-datalab.github.io/relational-language-t...
--
[1] https://ghc.gitlab.haskell.org/ghc/doc/users_guide/exts/arro... [2] https://www.cosc.brocku.ca/Faculty/Winter/JoRMiCS/Vol1/PDF/v...
Thank you! Prela’s `.and` operator is exactly the fork in fork algebra. I also suspect you will get `.select` and `.and` if you lift >>> and &&& to the category of Rel.
> I wonder where the limitations of this approach are seen?
Compile time. Rustc takes forever to compile, much longer than the time it takes to run the query. There’s plan to build a JIT for Prela, which would also improve the interop.
Most of the language design is orthogonal to the embedded implementation though, and Prela could very well be implemented in a vectorized engine.
Got it! One more question if I may: in https://arxiv.org/pdf/2607.26356, all the JOB queries are around a second or less, if I read correctly.
My understanding is the query:
movie
is inlined to something by the compiler approximately like.with(keyword.eq("my-kw")) .select(title)for movie in 0..movie_count {
}let lo = keyword_offsets[movie]; let hi = keyword_offsets[movie + 1]; for pos in lo..hi { let kw = keyword_ids[pos]; if keyword_text[kw] == "my-kw" { emit(movie, movie_title[movie]); break; } }So there's no index lookup on keyword, right? E.g., to use a hash index on keyword to find the resulting movie rows. If I wanted to do so, would I re-normalize the data in some way? This is what surprised me: that even without the secondary indexes, it is still the same (or more) performant than DuckDB.
Perhaps the index-lookup version would use something like
let movies_by_keyword: HashIdx<_, _> =
but it does not seem like it does (even though I assume DuckDB may).keyword.text().inv().collect();Yes basically. But DuckDB actually does not build an index on the keyword text. It only builds primary key indices automatically on data load. The standard benchmark schema specifies fk indices but not one for the keyword text.