FlowSphera 1.3.2 release notes
ClickHouse gets half of the RAM (+3 more changes)
Faster pages over long periods on busy networks (reported: a VM with 4 CPUs and 8 GB of RAM, ~720 records/s, 40 million records a day — Overview for 24 hours took ~20 s at 100 % CPU and some panels failed). Measured on 40 million synthetic records a day, 4 vCPUs:
- ClickHouse gets half of the RAM (1–8 GiB; was a quarter rounded down to whole GiB, so 8 GB of RAM gave 1 GiB
and the panels of one page did not fit into it together).
--upgradeapplies it to existing installs (ClickHouse restarts once, the collector keeps the records meanwhile). - Top lists read the data once: the total for the percentages comes from the same query (it was a second full pass for every list); what an outside address is (service, country, city, AS) is looked up only for the addresses shown; distinct addresses are counted by a 64-bit hash (the same numbers, less than half the work).
- Top conversations and Top hosts → main service no longer run out of memory on millions of host-peer pairs: exact while they fit, otherwise the heaviest candidates are found with a bounded-memory count and their numbers are exact.
- The same request once: widgets of a page that ask the same question share one answer (Overview asked for the trend twice, and the map overlay repeated the top lists); answers over 6 hours or more are kept for 60 s on the server, so reloads, other users and other pages with the same period do not read the day again.