Absolutely true! When I worked on this at $prevCo, it was tricky and sometimes caused bugs, unexpected behavior, and confused colleagues.
I will say that a proper implementation of a normalized cache on the client must have an ~inherent (thus generated) understanding of the object graph. It also must provide simple control of "whether to cache" at each request. Most of the problems we experienced were a result of the first constraint not being fully satisfied.
My impression is that Apollo does a good job on both of these but I haven't used it so I can't say.
I'll also note that the approach of "when one component makes an update, tell all other components on the page to refetch" sounds like a recipe for problems too – excess server/db load, unrelated data changing in front of users' eyes (and weird hacks to prevent this), etc.
Of course, with the wundergraph architecture, it sounds like answer to these questions would simply be to load a given table only once per page – which means no more defining queries on the "panel" components in the dashboard, for example.
All tricky tradeoffs! The right answer depends on what you're building. The Wundergraph approach sounds pretty cool for a lot of cases!
Does it lead to greater complexity somewhere, and all the issues around making that complexity bulletproof? Sure. But the user experience is so viscerally different that some will demand it. I think it’s admirable to work on getting that complexity correct and properly abstracted so that it can be re-used easily.
But I don't think it addresses why I'd want a "smart" GraphQL client: normalized caching on the client.
Say I have a dashboard where multiple panels on a given page make their own requests, since the panels are shared between many pages. But they share some objects. If I get updated data in a request from one panel, I'd like to see that update in all panels, without triggering more requests.
Side note that a magic layer to have each of those components combine their requests into one would actually hurt performance, since it's better to load the requests in parallel. And manually merging them into one would be quite a chore.