Brian Brazil
971dafdfbe
Coalesce series reads where we can.
...
When compacting rather than doing a read of all
series in the index per label name, do many at once
but only when it won't use (much) more ram than writing the
special all index does.
original in-memory postings:
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4 1 1202383447 ns/op 158936496 B/op 1031511 allocs/op
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4 1 1141792706 ns/op 154453408 B/op 1093453 allocs/op
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4 1 1169288829 ns/op 161072336 B/op 1110021 allocs/op
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4 1 1115700103 ns/op 149480472 B/op 1129180 allocs/op
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4 1 1283813141 ns/op 162937800 B/op 1202771 allocs/op
before:
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4 1 1145195941 ns/op 131749984 B/op 834400 allocs/op
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4 1 1233526345 ns/op 127889416 B/op 897033 allocs/op
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4 1 1821942296 ns/op 131665648 B/op 914836 allocs/op
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4 1 8035568665 ns/op 123811832 B/op 934312 allocs/op
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4 1 71325926267 ns/op 140722648 B/op 1016824 allocs/op
after:
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4 1 1101429174 ns/op 129063496 B/op 832571 allocs/op
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4 1 1074466374 ns/op 124154888 B/op 894875 allocs/op
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4 1 1166510282 ns/op 128790648 B/op 912931 allocs/op
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4 1 1075013071 ns/op 120570696 B/op 933511 allocs/op
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4 1 1231673790 ns/op 138754288 B/op 1022791 allocs/op
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
2019-12-12 08:38:14 +00:00
Brian Brazil
373a1fdfbf
Reread index series rather than storing in memory.
...
Rather than building up a 2nd copy of all the posting
tables, construct it from the data we've already written
to disk. This takes more time, but saves memory.
Current benchmark numbers have this as slightly faster, but that's
likely due to the synthetic data not having many label names.
Memory usage is roughly halved for the relevant bits.
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
2019-12-11 22:23:39 +00:00
