This reverts commit 7c6d3981bf.
Reason for revert: high contention at bucket16Pool on systems with big number of CPU cores.
This slows down query processing significantly.
This speeds up bucket32.addBucketAtPos() when bucket32.buckets contains big number of items,
since the copying of bucket16 pointers is much faster than the copying of bucket16 objects.
This is a cpu profile for copying bucket16 objects:
10ms 13.43s (flat, cum) 32.01% of Total
10ms 120ms 650: b.b16his = append(b.b16his[:pos+1], b.b16his[pos:]...)
. . 651: b.b16his[pos] = hi
. 13.31s 652: b.buckets = append(b.buckets[:pos+1], b.buckets[pos:]...)
. . 653: b16 := &b.buckets[pos]
. . 654: *b16 = bucket16{}
. . 655: return b16
. . 656:}
This is a cpu profile for copying pointers to bucket16:
10ms 1.14s (flat, cum) 2.19% of Total
. 100ms 647: b.b16his = append(b.b16his[:pos+1], b.b16his[pos:]...)
. . 648: b.b16his[pos] = hi
10ms 700ms 649: b.buckets = append(b.buckets[:pos+1], b.buckets[pos:]...)
. 330ms 650: b16 := &bucket16{}
. . 651: b.buckets[pos] = b16
. . 652: return b16
. . 653:}
This partially reverts fb82c4b9fa
It has been appeared that the additional memory allocation may result in higher GC pauses.
It is better to spend CPU time on copying bigger bucket16 structs instead of increasing query latencies due to higher GC pauses
* Remember the last accessed bucket on Has() call.
* Inline fast paths inside Add() and Has() calls.
* Remove fragile code with maxUnsortedBuckets inside bucket32.
Issues fixed:
- Slow startup times. Now the index is loaded from cache during start.
- High memory usage related to superflouos index copies every 10 seconds.
This should improve inverted index search performance for filters matching big number of time series,
since `lib/uint64set.Set` is faster than `map[uint64]struct{}` for both `Add` and `Has` calls.
See the corresponding benchmarks in `lib/uint64set`.
