Parca retains profile samples in memory and rotates the in-memory block out when it reaches a certain size in bytes based on configuration settings.
We recommend a rule of thumb of 60 bytes of memory for each sample. By default the Parca-Agent profiles each core at 19Hz, which means you can expect Parca to require 60 bytes 19 number of cores. So if you were profiling a 100 cores at a time you'd expect Parca to ingest 114Kb/s.
Flags to adjust the active memory footprint
|By increasing the granule size you'll decrease nominal memory footprint but increase memory usage during compaction.|
|Adjust the number of bytes held in memory before being rotated out. This is the easiest way to adjust Parca's memory footprint.|
|Smaller row groups increase memory footprint but increase query speed. Increasing row group size results in lower memory usage but slower queries.|