Update HyperLogLog library to original Airlift

[pranjalssh]

There has been considerable performance improvements to original airlift library. Namely:

airlift@a932881
airlift@23b97cb

I copied all HyperLogLog files from airlift/airlift to prestodb/airlift - and made some API changes necessary for our internal PrivateLpcaSketch. API changes were "read-only" - so we don't have to worry about correctness issues in PrivateLpcaSketch:

DenseHll.eachBucket(BucketListener listener)
HyperLogLog.getMaxBucketValue()
HyperLogLog.getNumberOfBuckets()

Benchmark before and after showing 70X improvement for benchmarkMergeWithSparse which matches our expectations:

Benchmark                                   Mode  Cnt   Score   Error  Units
BenchmarkDenseHll.benchmarkInsert           avgt   50   1.441 ± 0.114  us/op
BenchmarkDenseHll.benchmarkMergeWithDense   avgt   50  21.630 ± 0.541  us/op
BenchmarkDenseHll.benchmarkMergeWithSparse  avgt   50  22.805 ± 0.227  us/op

Benchmark                                   Mode  Cnt   Score   Error  Units
BenchmarkDenseHll.benchmarkInsert           avgt   50   1.409 ± 0.035  us/op
BenchmarkDenseHll.benchmarkMergeWithDense   avgt   50  11.615 ± 1.016  us/op
BenchmarkDenseHll.benchmarkMergeWithSparse  avgt   50   0.335 ± 0.009  us/op

Commit 1: Update benchmarks and some tests(so we can compare)
Commit 2: Ports over HyperLogLog changes

[pranjalssh]

IDK why they removed @ NotThreadSafe. I'll let someone more familiar with Java decide what's better

[rschlussel]

This was changed as part of airlift@1eb30dc.

It looks like the background is that javax.annotations was moved and renamed to jakarta.annotations (rename was due to some licensing disagreement between oracle and eclipse see jakartaee/rest#760). But Jakarta.annotations doesn't have the NotThreadSafe annotation (see discussion here: jakartaee/common-annotations-api#91). So seems like this change was basically part of upgrading that library.

[feilong-liu]

LG. Have you run any tests to make sure the result is the same before and after change?

[pranjalssh]

I've run unit tests so far. Note that Presto will not use this airlift code until we release this and update the version in Presto's pom. Once we merge and release this we can run correctness suites before landing in presto

[pranjalssh]

Tried several queries to verify correctness and saw performance gains as well. Going to push now and figure out how to release later

[agrawaldevesh]

Just wow !! Would love to learn more about the performance improvement we saw on a real presto query with this change.

[rschlussel]

I didn't really try to understand all of it, but it's been reviewed already in airlift, and its well tested and looks reasonable.

Can you add a co-author line to the commit messages to credit whoever wrote the commits you took. Also would be nice to link those commits so its easy to follow to get more info.

[rschlussel]

This is a correctness fix. We should call this out explicitly in the commit message

[pranjalssh]

Good point, I'll put :
"Improved accuracy and performance of hyperloglog functions"

[rschlussel]

Do you know why this number is 4? Is it because BITS_PER_BUCKET is 4 (I see that's used in calculating the size of the deltas array)? If so, we should use the static variable explicitly here so they can't get out of sync.

[pranjalssh]

That makes sense - but ideally we should keep most of the code in sync with airlift/airlift so its easier to copy in future as well. So we can change there first and then copy it. Wdyt?

[rschlussel]

that's fine with me.