But for core text search and indexing - tantivy does indeed support tokenization, stemming, fuzzy matching, and so on. For the core use case of performing a text search on a large corpus of text a lot of the functionality one would need is there - as you can see if you look at the list of queries in the benchmark. From the list of queries there's little I would miss from our "enterprise" use of lucene today - speaking for myself.
One might argue "oh but if you add some obscure features that 5% of people use like a real enterprise solution then it will slow down by 2x" - but I doubt it.
To me, tantivy and the likes are a technology demonstrator - they show that lucene could be significantly faster if it wasn't in java.
Lucene is a Swiss army knife for building you something to get you the best results possible. Those aren't enterprise features if you actually care about what your users do with search results. For example because your sales are directly correlated to that. The trick with Lucene is to do as much as you can possibly get away with as opposed to doing as little as you can do to call it a day that seems to be the strategy here for proving X is faster than Y with this particular benchmark. It's faster because it does less useful work.
The proof of concept here would be building something that is as good and as fast. By not even trying to benchmark how good things are, you kind of make the point that you either don't know or don't care enough to even bother to measure it.
And since the article is called "why C is faster than Java" and the main argument is literally "here's this thing implemented in C that is fast", it's an excellent counterpoint to go "here's this other thing in Java that is fast".
A historical accident isn’t a counterpoint.
I am not a big Java fan. We are not talking a Haskell or Rust or similarly safe language here. However, Java is still far ahead of C in that regard.
For these use cases, you could swap out lucene for something faster or more “low level” in implementation without affecting the typical application using it much, as all interaction happens across some rest API.
If you replace a more or less safe implementation, with a buggy, memory unsafe C version of it, you are not going to make anyone happy about the performance improvement. Always assuming, that you put in "only" the same amount of time and number of people, not more, than the original implementation.
The amount of things you need to keep in mind with all the memory management and other details when writing C code can also make you blind to other issues like proper input validation and implementation of invariants of application settings. You have to put a lot of energy into making sure you are not leaking memory and avoiding memory safety issues, so you lack that energy when it comes to higher level issues. This is directly affecting developer productivity.
But the point at issue is that by using java you are trading off speed (potentially for other benefits).
By the way - tantivy at least is implemented in rust so it shouldn’t suffer from memory safety issues just as lucene shouldn’t. And the rust language will provide some guarantees that java doesn’t - like freedom from data races.
Most of the libraries you mention don't come close to even implementing a tiny portion of Lucene. Classic case of apples and oranges. Also, good examples of the niche things I was talking about. Benchmarks like you mention are kind of self serving like that. They measure something but not everything and probably very selectively. They are faster at what exactly? Under what circumstances? Why? As soon as you answer those questions, what inherent limitations do the Lucene developers have replicating that?
A lot here boils down to how the underlying search engine implements tokenization, stemming, language analyzers, fuzzy matching, and a few other things that you'd need to build a search engine that doesn't suck. The benchmark conveniently does not specify any of that; presumably because it lacks many of those features or has extremely naive implementations of those things. That would be the kind of corner cutting I was talking about. What kind of relevance ranking is being used here? How good is it? Was that even evaluated or considered? Hint, Lucene gives you many options here. Search quality and performance are the big trade off here. Anything that trades off quality over performance is going to look good until you look at the quality.
Also things like the index size and document volume are not specified. Or the hardware this ran on. Or the JVM configuration, compiler flags, etc. It's like benchmarking a formula 1 car by how quick it is at parallel parking. Yes, a T-Ford is going to be faster at that maybe. But is that even meaningful to look at?