See the corresponding benchmark in Prometheus - 23c0299d85/tsdb/head_bench_test.go (L52)
The benchmark allows performing apples-to-apples comparison of time series search
in Prometheus and VictoriaMetrics. The following article - https://www.robustperception.io/evaluating-performance-and-correctness -
contains incorrect numbers for VictoriaMetrics, since there wasn't this benchmark yet. Fix it.
Benchmarks can be repeated with the following commands from Prometheus and VictoriaMetrics source code roots:
- Prometheus: GOMAXPROCS=1 go test ./tsdb/ -run=111 -bench=BenchmarkHeadPostingForMatchers
- VictoriaMetrics: GOMAXPROCS=1 go test ./lib/storage/ -run=111 -bench=BenchmarkHeadPostingForMatchers
Benchmark results:
benchmark old ns/op new ns/op delta
BenchmarkHeadPostingForMatchers/n="1" 272756688 364977 -99.87%
BenchmarkHeadPostingForMatchers/n="1",j="foo" 138132923 1181636 -99.14%
BenchmarkHeadPostingForMatchers/j="foo",n="1" 134723762 1141578 -99.15%
BenchmarkHeadPostingForMatchers/n="1",j!="foo" 195823953 1148056 -99.41%
BenchmarkHeadPostingForMatchers/i=~".*" 7962582919 8716755 -99.89%
BenchmarkHeadPostingForMatchers/i=~".+" 7589543864 12096587 -99.84%
BenchmarkHeadPostingForMatchers/i=~"" 1142371741 16164560 -98.59%
BenchmarkHeadPostingForMatchers/i!="" 9964150263 12230021 -99.88%
BenchmarkHeadPostingForMatchers/n="1",i=~".*",j="foo" 216995884 1173476 -99.46%
BenchmarkHeadPostingForMatchers/n="1",i=~".*",i!="2",j="foo" 202541348 1299743 -99.36%
BenchmarkHeadPostingForMatchers/n="1",i!="" 486285711 11555193 -97.62%
BenchmarkHeadPostingForMatchers/n="1",i!="",j="foo" 350776931 5607506 -98.40%
BenchmarkHeadPostingForMatchers/n="1",i=~".+",j="foo" 380888565 6380335 -98.32%
BenchmarkHeadPostingForMatchers/n="1",i=~"1.+",j="foo" 89500296 2078970 -97.68%
BenchmarkHeadPostingForMatchers/n="1",i=~".+",i!="2",j="foo" 379529654 6561368 -98.27%
BenchmarkHeadPostingForMatchers/n="1",i=~".+",i!~"2.*",j="foo" 424563825 6757132 -98.41%
The first column (old) is for Prometheus, the second column (new) is for VictoriaMetrics.
Prometheus was using 3.5GB of RAM during the benchmark, while VictoriaMetrics was using 400MB of RAM.
Production workload shows that the index requires ~4Kb of RAM per active time series.
This is too much for high number of active time series, so let's delete this index.
Now the queries should fall back to the index for the current day instead of the index
for the recent hour. The query performance for the current day index should be good enough
given the 100M rows/sec scan speed per CPU core.
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.
Continue trying to remove NFS directory on temporary errors for up to a minute.
The previous async removal process breaks in the following case during VictoriaMetrics start
- VictoriaMetrics opens index, finds incomplete merge transactions and starts replaying them.
- The transaction instructs removing old directories for parts, which were already merged into bigger part.
- VictoriaMetrics removes these directories, but their removal is delayed due to NFS errors.
- VictoriaMetrics scans partition directory after all the incomplete merge transactions are finished
and finds directories, which should be removed, but weren't still removed due to NFS errors.
- VictoriaMetrics panics when it finds unexpected empty directory.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/162
Incremental aggregate functions don't keep all the selected time series in memory -
they keep only up to GOMAXPROCS time series for incremental aggregations.
Take into account that the number of time series in RAM can be higher if they are split
into many groups with `by (...)` or `without (...)` modifiers.
This should reduce the number of `not enough memory for processing ... data points` false
positive errors.
The origin of the error has been detected and documented in the code,
so it is enough to export a counter for such errors at `vm_index_blocks_with_metric_ids_incorrect_order_total`,
so it could be monitored and alerted on high error rates.
Export also the counter for processed index blocks with metricIDs - `vm_index_blocks_with_metric_ids_processed_total`,
so its' rate could be compared to `rate(vm_index_blocks_with_metric_ids_incorrect_order_total)`.
Slow loops could require seeks and expensive regexp matching, while fast loops just scans
all the metricIDs for the given `tag=value` prefix. So these operations must have separate
max loops multiplier.
The fastest tag filters are non-negative non-regexp, since they are the most specific.
The slowest tag filters are negative regexp, since they require scanning
all the entries for the given label.
