On Mon, Aug 22, 2022 at 05:44:26PM -0700, Adam Williamson wrote:

Hey folks! I apologize for the wide distribution, but this seemed like a bug it'd be appropriate to get a wide range of input on.

There's a bug that was proposed as an F37 Beta blocker: https://bugzilla.redhat.com/show_bug.cgi?id=1907030

it's quite an old bug, but up until recently, the summary was apparently accurate - dnf would run out of memory with 512M of RAM, but was OK with 1G. However, as of quite recently, on F36 at least (not sure if anyone's explicitly tested F37), dnf operations are commonly failing on VMs/containers with 1G of RAM due to running out of RAM and getting OOM-killed.

The discussion in the bug indicates that this memory growth is related to loading of the full filepath dataset. We have been discussing splitting out the non-primary-filepath-data (i.e. paths that are not /etc, /usr/bin, /usr/sbin), out into a separate lazilly-loaded file. If we manage to do that, we'll kill two birds with one stone:

- initial download of repo metadata on every freakin' dnf operation can go down from 80 to 20 MB - the peak memory use will go down

Apparently DNF5 makes this possible. My vote is: yes, this is an issue. No, we shouldn't block F37 on this. Apparently the only reasonable way to tackle this is with a major rework of DNF, so let's get it right with DNF5.

Zbyszek

On 8/22/22 20:44, Adam Williamson wrote:

Hey folks! I apologize for the wide distribution, but this seemed like a bug it'd be appropriate to get a wide range of input on.

There's a bug that was proposed as an F37 Beta blocker: https://bugzilla.redhat.com/show_bug.cgi?id=1907030

it's quite an old bug, but up until recently, the summary was apparently accurate - dnf would run out of memory with 512M of RAM, but was OK with 1G. However, as of quite recently, on F36 at least (not sure if anyone's explicitly tested F37), dnf operations are commonly failing on VMs/containers with 1G of RAM due to running out of RAM and getting OOM-killed.

There's some discussion in the bug about what might be causing this and potential ways to resolve it, and please do dig into/contribute to that if you can, but the other question here I guess is: how much do we care about this? How bad is it that you can't reliably run dnf operations on top of a minimal Fedora environment with 1G of RAM?

I think we need to stop thinking about this as just a single computer problem.

In a containerized world you could have numerous containers running dnf operations on a machine at any given time. Making the dnf operations take less memory means we'll hit many fewer issues as single machines are able to run many dnf operations at a given moment.

This obviously has some overlap with our stated hardware requirements, so here they are for the record:

https://docs.fedoraproject.org/en-US/fedora/latest/release-notes/welcome/Har...

that specifies 2GB as the minimum memory for "the default installation", by which I think it's referring to a default Workstation install, though this should be clarified. But then there's a "Low memory installations" boxout, which suggests that "users with less than 768MB of system memory may have better results performing a minimal install and adding to it afterward", which kinda is recommending that people do exactly the thing that doesn't work (do a minimal install then use dnf on it), and implying it'll work.

After some consideration I don't think it makes sense to take this bug as an F37 blocker, since it already affects F36, and that's what I'll be suggesting at the next blocker review meeting. However, it does seem a perfect candidate for prioritized bug status, and I've nominated it for that.

I guess if folks can chime in with thoughts here and/or in the bug report, maybe a consensus will emerge on just how big of an issue this is (and how likely it is to get fixed). There will presumably be a FESCo ticket related to prioritized bug status too.

Thanks folks!