promscale
promscale copied to clipboard
Better defaults for cache sizes
To achieve good performance we are currently recommending our users to tweak following:
--metrics.cache.series.initial-size=30000000
--cache.memory-target=90%
Let's make sure that cache memory defaults work well enough out of the box. Cache tweaks should be only used for extreme
cases.
We should validate our findings/improvements by running benchmarks.
Fixes in https://github.com/timescale/promscale/pull/1705 mean that great majority of users shouldn't have to tweak metrics.cache.series.initial-size
.
Benchmark numbers look pretty good now with default settings so cache.memory-target
can stay the same.
Reopening issue to validate existing DefaultSeriesCacheSize = 250000
setting which seems low.
Base on memory heap profile existing DefaultSeriesCacheSize = 250000
uses around 70MB. Maybe we should reconsider increasing it to 500K or maybe even 1M? (140MB or 280MB)? That would make defaults more appropriate for realistic data loads and would mean less overhead of growing cache. @VineethReddy02 What are your thoughts?