dragonflydb
untracked memory or a leak
A dragonfly process slowly grows in RSS while used memory stays stable. See the attached files:
info log + arena logs: info_all_101208.txt arena_101208.txt arena_backing_101208.txt
The process uses 16 threads. These show that the leakage is in the backing arena. Further analysis shows that most of the arena memory is spent on allocations with sizes higher than 3000.
Allocation tracking of memory track add 3000 20000 1
jdbfly.txt.tar.gz
More investigations is needed.
My working assumption is that we do not track/clean up redis parser memory. I see lots of memory allocations coming from there. maybe even we have a bug that causes the parser to allocate more and more.
It seems that we track redis parser correctly. I am attaching more allocation logs but now it seems more and more like a leak.
Just thought I'd mention that I'm seeing this in production on v1.22.1. Virtual memory steadily grows over time (in blocks), while real memory usage stays relatively stable.
