For the most part, I'd rather have native packages. I'm not deeply philosophically opposed to secondary packaging systems, and only mildly opposed to "ship the whole dependency tree in an archive" software distribution methods (lookin' at you NextStep/OS X style bundles), and see their potential especially on platforms with no/bad native package managers or to bring in specific software that would pose a compatibility problem for the host system... but they never seem to work nearly as well as native packages, and the two big players on Linux have problems.

As far as I'm concerned, they're just taking the old last-ditch practice of "I have this piece of recalcitrant software that is incompatible with the rest of my system, so I'll throw it in /opt with its entire dependency tree," replacing opt with a bunch of bind mounts, and doing so with varying degrees of additional tooling.

The sandboxing is a nice idea, but it seems like in practice the models on both snap and flatpack are simultaneously restrictive in ways that make them annoying-to-unusable for many tasks, and too sloppy to provide reliable security guarantees.

They make debugging problems harder because you can't check functionality from another program because they likely don't share libs. ldd is a lot easier than spelunking around with eg. flatpak list --app --columns=application,runtime until you find a "peer" to test.

If I need a one-off piece of software that is a compatibility nuisance on my host distro (but not so much of a nuisance it needs to go in a container or VM, which is a pretty narrow window), I'll usually reach for an AppImage because unlike the other two, they're actually fairly standalone and don't involve a big invasive runtime/tooling system.

The Immutable-core OSes that depend on them are kind of the same way at the moment. Fundamentally a pretty neat idea, but so far I find them super frustrating in practice. Nix is ...different... to work with, but is likely a more elegant scheme to solve the same class of problems.