Open vadim2404 opened 1 year ago
@knizhnik created a branch with in-memory cache neondatabase/neon#4891
ref https://neondb.slack.com/archives/C03TN5G758R/p1692042048475039?thread_ts=1690842628.716389
I've been able to reproduce the kernel panic with a relatively simple setup (at least, not requiring Neon's entire stack :sweat_smile:). And, it looks the panic requires tmpfs usage to occur, and does not trigger with equivalent load from anonymous mmap usage.
Below are the notes I wrote for myself while trying to debug this, along with the various files referenced there.
Should we merge mmap version of the cache and remove our tmpfs hacks?
Yes, once reviewed & passing tests, etc. I'm not familiar enough with C to review it.
This has been mitigated by moving the file cache to disk, but the underlying kernel issue is still present, so: leaving this issue open, but reducing priority.
We recently updated the kernel (#638), and should check whether this issue is now fixed because of it.
Environment
Steps to reproduce
Expected result
Actual result
Other logs, links