Yes, a known issue with redefining records and other things to which you might hold a reference somewhere.

so they are just bad to use with a repl and browser hot-reload or is there some other technique a-la defonce?

With hot code reload in a browser it should work just fine. At least, AFAICT it has always been working for me. The technique that I prefer on the CLJ side is to never reload things by hand. Instead, I rely on the "reloaded" workflow where I make a set of changes, send a specific keystroke, and my REPL reloads all the necessary namespaces while tracking all of the dependencies. For that, I use clojure.tools.namespace.repl and Integrant (helps with stateful things, and there are alternatives).

I came across this using hot-reload with re-frame/shadowcljs. Saving the ns with defrecord makes all the keys in app-db useless

All the keys? That sounds bizarre given that usually the keys there are keywords or other scalars.

all the keys of that type

Hmm. Yeah, I have never seen that probably only because I almost never use records. But sounds logical. Code references should get refreshed by hot code reload, but data references cannot be when that data is in defonce, just like re-frame.db/app-db.

Came across this old thread, and thought I'd chime in. This technique won't work if you change an actual record's definition, but Potemkin's defrecord+ (and related macros) are designed for exactly this. They skip re-evaluation if the defrecord's body is unchanged. Dunno about cljs tho.

Yeah, that's only for CLJ. And not necessary with a reloaded workflow. ;) With CLJS, I would simply avoid storing records altogether.

"Doctor, it hurts when I do that." "Then stop doing that...forever."

Well... yeah. :) Of course, sometimes it does make more sense to use a record instead of a map. But it's a relatively rare case. And most of the time you can still find a way to use plain maps without sacrificing anything.