Agreed 30s isn’t enough, but for long running processes LXD has async operations (which use websockets to monitor the progress of the operation), so the initial header response should still be <30s.
I’ve tested importing large instances that take over 30s (and indeed our own automated tests do this for VMs as well) and we don’t see the issue. I’ve also tested launching instances from images that take over 30s to download (by slowing down the network), and don’t see the issue there either.
If you can reproduce this issue then it would be valuable to see the server log output from lxc monitor --type=logging --pretty at the same time as running the command that is failing with the --debug flag.
I started to go over the 7210 lines in the log to make sure I don’t leak something private, but it was taking too much time and since then LXD 5.3 was released, so I’ll try the new version.