Also worth nothing that I think I prefer @nomeata’s approach of allowing the canister to request that the proxy “upgrade” the call from query to
update.
“upgrade” means to make the call again as an update call, so in general it has the potential to always make 2 calls instead of 1, but it makes no assumptions about how HTTP methods map to the IC.
The compelling examples against my approach from the linked discussion include the need to use HTTP POST but not update state, e.g. an image transcoding service.
In that case my approach is inefficient because it makes an update call when a query would do.
The real problem with my approach is that if you need to support an existing API client which makes something like a GET request but then the canister needs to make a state modification. @nomeata’s approach handles that just fine, although there the inefficiency is in making 2 calls.