Ah, Ok.

I do as (or a similar workflow): I rsync the content directory and let Hugo on the server render. My sites are public, but perhaps they're just much smaller or not as popular; Hugo renders even my largest site in about a second, but for a large, slow, heavy-use production situation I could see a push-and-swap process for a more atomic site update.

I don't see the degradation you do, but there are so many possible variables.

My biggest gripe about Hugo is how limited it is in supporting source document formats. There's no mechanism for hooking in different formats, and the team is reluctant to merge PRs for other formats. When I started with Hugo, I had a large repository of essays spanning a decade and written in a variety of markup, from asciidoc (which I used for years), to reST, to markdown; and markdown is by far the worst. I was faced with converting everything to markdown, which was usually a lossy process because markdown is so limited, or not publishing all of that history. And now we have djot, which is almost the perfect plain text markup language, but I again have to first do a lossy conversion to markdown to get Hugo to consume it. It low-key sucks, and I'm actively looking for an alternative that has a more flexible AST-based model for which new formats can be added; something that consumes a format like pandoc's AST.