perf: switch jaq-json hasher to foldhash

[tranzystorekk]

This change replaces the hasher used in jaq-json with foldhash.

My personal motivation is that foldhash seems to be entering the ecosystem as an upcoming state-of-the-art, but there are other considerations:

• The original motivation for ahash way back in 0.5.0 was 6114f30 and now foldhash has zero runtime dependencies, which allows dropping zerocopy from Cargo.lock
• Benchmarks show that foldhash is at worst comparable to ahash, so no apparent perf harm in the replacement

I am aware that this change might feel insubstantial or arbitrary, so please treat it as a loose suggestion and feel free to close immediately if uninterested :)

[01mf02]

Hi @tranzystorekk, thanks for your PR!

I ran a quick benchmark:

$ ./bench.sh ./jaq-ahash ./jaq-foldhash 
{"name": "empty", "n": 512, "time": {"./jaq-ahash": [0.43], "./jaq-foldhash": [0.39]}}
{"name": "bf-fib", "n": 13, "time": {"./jaq-ahash": [0.45], "./jaq-foldhash": [0.45]}}
{"name": "defs", "n": 100000, "time": {"./jaq-ahash": [0.06], "./jaq-foldhash": [0.06]}}
{"name": "upto", "n": 8192, "time": {"./jaq-ahash": [0.00, 0.00, 0.00], "./jaq-foldhash": [0.00, 0.00, 0.00]}}
{"name": "reduce-update", "n": 16384, "time": {"./jaq-ahash": [0.01, 0.01, 0.01], "./jaq-foldhash": [0.01, 0.01, 0.01]}}
{"name": "reverse", "n": 1048576, "time": {"./jaq-ahash": [0.05, 0.04, 0.04], "./jaq-foldhash": [0.05, 0.04, 0.04]}}
{"name": "sort", "n": 1048576, "time": {"./jaq-ahash": [0.13, 0.12, 0.12], "./jaq-foldhash": [0.12, 0.13, 0.12]}}
{"name": "group-by", "n": 1048576, "time": {"./jaq-ahash": [0.52, 0.55, 0.55], "./jaq-foldhash": [0.55, 0.54, 0.54]}}
{"name": "min-max", "n": 1048576, "time": {"./jaq-ahash": [0.21, 0.21, 0.22], "./jaq-foldhash": [0.21, 0.22, 0.22]}}
{"name": "add", "n": 1048576, "time": {"./jaq-ahash": [0.48, 0.48, 0.50], "./jaq-foldhash": [0.47, 0.49, 0.48]}}
{"name": "kv", "n": 131072, "time": {"./jaq-ahash": [0.12, 0.13, 0.12], "./jaq-foldhash": [0.12, 0.13, 0.12]}}
{"name": "kv-update", "n": 131072, "time": {"./jaq-ahash": [0.13, 0.14, 0.13], "./jaq-foldhash": [0.14, 0.14, 0.13]}}
{"name": "kv-entries", "n": 131072, "time": {"./jaq-ahash": [0.60, 0.60, 0.60], "./jaq-foldhash": [0.58, 0.59, 0.58]}}
{"name": "ex-implode", "n": 1048576, "time": {"./jaq-ahash": [0.53, 0.51, 0.51], "./jaq-foldhash": [0.53, 0.54, 0.53]}}
{"name": "reduce", "n": 1048576, "time": {"./jaq-ahash": [0.77, 0.78, 0.78], "./jaq-foldhash": [0.78, 0.80, 0.78]}}
{"name": "try-catch", "n": 1048576, "time": {"./jaq-ahash": [0.29, 0.29, 0.29], "./jaq-foldhash": [0.29, 0.30, 0.31]}}
{"name": "repeat", "n": 1048576, "time": {"./jaq-ahash": [0.15, 0.16, 0.15], "./jaq-foldhash": [0.16, 0.15, 0.15]}}
{"name": "from", "n": 1048576, "time": {"./jaq-ahash": [0.33, 0.31, 0.31], "./jaq-foldhash": [0.31, 0.32, 0.31]}}
{"name": "last", "n": 1048576, "time": {"./jaq-ahash": [0.04, 0.04, 0.04], "./jaq-foldhash": [0.04, 0.04, 0.04]}}
{"name": "pyramid", "n": 524288, "time": {"./jaq-ahash": [0.37, 0.34, 0.35], "./jaq-foldhash": [0.34, 0.36, 0.35]}}
{"name": "tree-contains", "n": 23, "time": {"./jaq-ahash": [0.07, 0.07, 0.07], "./jaq-foldhash": [0.07, 0.07, 0.07]}}
{"name": "tree-flatten", "n": 17, "time": {"./jaq-ahash": [0.80, 0.79, 0.79], "./jaq-foldhash": [0.79, 0.77, 0.79]}}
{"name": "tree-update", "n": 17, "time": {"./jaq-ahash": [0.71, 0.71, 0.77], "./jaq-foldhash": [0.71, 0.71, 0.70]}}
{"name": "tree-paths", "n": 17, "time": {"./jaq-ahash": [0.43, 0.45, 0.44], "./jaq-foldhash": [0.47, 0.48, 0.46]}}
{"name": "to-fromjson", "n": 65536, "time": {"./jaq-ahash": [0.04, 0.04, 0.04], "./jaq-foldhash": [0.04, 0.04, 0.04]}}
{"name": "ack", "n": 7, "time": {"./jaq-ahash": [0.61, 0.65, 0.63], "./jaq-foldhash": [0.59, 0.58, 0.56]}}
{"name": "range-prop", "n": 128, "time": {"./jaq-ahash": [0.36, 0.37, 0.37], "./jaq-foldhash": [0.35, 0.36, 0.35]}}
{"name": "cumsum", "n": 1048576, "time": {"./jaq-ahash": [0.28, 0.29, 0.28], "./jaq-foldhash": [0.29, 0.30, 0.29]}}
{"name": "cumsum-xy", "n": 1048576, "time": {"./jaq-ahash": [0.48, 0.50, 0.49], "./jaq-foldhash": [0.49, 0.50, 0.50]}}

The most relevant ones are probably "kv*", and we see very little difference there.

I'm a bit divided about this change: Having fewer dependencies is always nice. :) However, the given hasher shows up in the public API of jaq-json, which means that it cannot be updated without breaking the API.

As evidence for the Rust ecosystem shifting to foldhash, I see hashbrown migrating to it. Is there any other evidence that supports your observation?
I also found: https://thaliaarchi.github.io/notes/pl/langs/rust/foldhash_vs_rustc-hash.html
(Perhaps @thaliaarchi would also like to comment on this topic?)

I also did not test foldhash without the std feature, but this would need to be done too, because jaq-json should not depend on std by default.

Just throwing in an idea: While looking at hashbrown right now, I saw that it has a feature default-hasher. This might be also a nice idea for jaq-json --- making jaq_json::Val generic over its hasher and providing a default one only if default-hasher is enabled. I would gladly accept a PR for such a change, because it would even more effectively decrease the number of dependencies. ;)

I tend to be more conservative with jaq changing dependencies nowadays. Intuitively, I would not change the hasher right now, but let me think a bit about it.

[thaliaarchi]

Since both hashbrown and rustc have switched to foldhash, I'm quite convinced that it's suitable for general-purpose, but fast hashing. As noted in my comparison that you found, rustc uses a variant of foldhash which simplifies it for the data typical to a compiler and it even improves upon the old FxHash. I've switched to foldhash in my projects.

As for your concern about dependencies, I think the default-hasher default feature you describe would be a good solution. If library users care about the implementation, they can disable the default hasher and supply their own. That would be a good solution for users who want a stable hasher. I suspect the churn would be significantly less than you expect.

[01mf02]

Thanks for your input, @thaliaarchi. I have thus decided to use foldhash. API stability is not such a big issue for this change after all, because it only affects jaq-json, but neither jaq-core nor jaq-std. Thanks also again to @tranzystorekk for your PR!