Add xan and qsv benchmark

[xcpky]

xsv is now unmaintained, the author recommends xan and qsv. Could we add them to the benchmark?

[liquidaty]

Thanks for the suggestion.

xan is super cool! I had not seen it before, so thanks for introducing.

It looks like xan, unlike xsv (or qsv), treats unbalanced quotation marks differently from Microsoft Excel (and Numbers, for folks using Mac). For example, hi","there" will be parsed by xan as a single column, but will be parsed by Excel/Numbers as two columns. Doesn't mean we can't include it in the benchmarks, but worth noting, since Excel parity was one of the most important core requirements that zsv was created to meet (and surprisingly enough, we've encountered many cases in the real world of CSV data with unbalanced dbl-quotes!).

I don't suppose xan comes with a flag for turning off that behavior? My quick look did not reveal any such thing but perhaps I missed. If it exists, we could do a truly apples-to-apples comparison (maybe one day we'll add a flag to zsv to go the other way, which will surely also make it faster...). Otherwise we will have to live with apples to oranges

[xcpky]

(maybe one day we'll add a flag to zsv to go the other way, which will surely also make it faster...)

I think this is good! For me I don't care about Microsoft and Apple sh*t, and I'm sure many don't either.

[jqnatividad]

Hi @xcpky , qsv maintainer here.

Would love to see qsv included in the benchmarks.

FYI, qsv has a highly tuned csv fork that also uses a lot of tweaks to make CSV parsing faster:

https://github.com/dathere/qsv/blob/47dfece1f1ed48f6a668a54a1165cbaa40aa757f/Cargo.toml#L298-L307

Would be interested to see how it fairs against zsv.

And while talking about benchmarks, have you considered using hyperfine for more robust results?

[liquidaty]

Thanks @jqnatividad. I'm guessing you meant to direct your prior msg to me (if not, disregard this note). Had not seen hyperfine before, thanks for the tip! Looks interesting. Does it also measure memory footprint? In any case, on first blush it seems to me that using hyperfine would be a different way to get measurements that, at the end of the day, we are already getting. Perhaps a better way, but on the other hand, a way that would require additional work to get to the same (at least current) end state. And perhaps not "same" exact end point, but "same enough" within an acceptable margin of error. or does it do more than that, and I'm missing the point?

Re the tuned qsv fork, will consider... what's the easiest way to install on mac?