update OffheapIncrementalIndex to put facts off-heap

[himanshug]

and also let user define max limit in terms of off-heap size instead of max number of rows

[fjy]

@himanshug what are the performance implications of this? last time we tried mapdb we had a lot of problems with serde and it decreased ingestion rates substantially

Also, please remember to submit proposals for these types of changes

[himanshug]

@fjy OffheapIncrementalIndex is not available for indexing[yet], but only optionally used in GroupBy queries which tries to optimize things by not sorting results in IncrementalIndex when possible (see #2571 ). with that, here are some numbers based on https://github.com/druid-io/druid/blob/master/benchmarks/src/main/java/io/druid/benchmark/IncrementalIndexAddRowsBenchmark.java .

OffheapIncrementalIndex with sortFacts = TRUE
Benchmark Mode Cnt Score Error Units
OffheapIncrementalIndexRowsBenchmark.normalFloats avgt 20 45.928 ± 1.221 us/op
OffheapIncrementalIndexRowsBenchmark.normalLongs avgt 20 46.847 ± 2.234 us/op
OffheapIncrementalIndexRowsBenchmark.normalStrings avgt 20 51.218 ± 1.358 us/op

OffheapIncrementalIndex with sortFacts = FALSE
Benchmark Mode Cnt Score Error Units
OffheapIncrementalIndexRowsBenchmark.normalFloats avgt 3 15.450 ± 9.408 us/op
OffheapIncrementalIndexRowsBenchmark.normalLongs avgt 3 13.112 ± 1.487 us/op
OffheapIncrementalIndexRowsBenchmark.normalStrings avgt 3 22.428 ± 24.327 us/op

OnheapIncrementalIndex with sortFacts = TRUE
Benchmark Mode Cnt Score Error Units
IncrementalIndexAddRowsBenchmark.normalFloats avgt 200 15.176 ± 0.625 us/op
IncrementalIndexAddRowsBenchmark.normalLongs avgt 200 14.611 ± 0.570 us/op
IncrementalIndexAddRowsBenchmark.normalStrings avgt 200 21.284 ± 0.394 us/op

note that in groupBy query processing sortFacts is TRUE only once in the very end before returning results to end user and with sortFacts = FALSE off-heap performance is not that bad.

this impl of off-heap is different from previous one as it does not serde any strings and keeps dimension dictionary on-heap wherease old impl tried to put everything off-heap.
note that on-heap version can not be created without very large heaps but off-heap version (even if 3x slower with sortFacts being true) allows larger groupBy queries with ability to upper bound on off-heap size.

[nishantmonu51]

I wonder if we should default to maxOffHeapSize always and ignore maxResults here ?

[himanshug]

copy/paste error, fixed, it is only supposed to rely on max off-heap limit and not on maxRows.

[himanshug]

@fjy @nishantmonu51 ?

[nishantmonu51]

all my comments are addressed, 👍

[himanshug]

i guess its not needed after the new groupBy query strategy.