This sounds interesting but also looks like a significant change.
My original idea was to use one client instance per request, but your proposal seems to be reasonable.
After brief research, the above statement can be split into these:

let endpoint = tonic::transport::Channel::from_static(service_url);
let channel = endpoint.connect().await?;
let mut blob_client = BlobServiceClient::<Channel>::new(channel);

As I understand, the Channel is "encouraged to be cloned" and the BlobServiceClient is a wrapper over tonic::client::Grpc<Channel> which derives Clone so it is also cloneable.

So my questions are:

1. Where and how would you store the client instance? Should it be provided in the same way as e.g. DatabaseClient is provided to the service? (D6178)
2. When should it be initialized? Lazily when first used?
3. How does Tonic handle network connection? Is it kept alive? How are disconnections handled then? Should I care about it at all? (I think I need to read more about http2)

If this is going to be a significant code structure change, I'd prefer to create a Linear task and do this as a follow-up after this stack is landed.

Awesome, thank you!

I created a Linear task for this: https://linear.app/comm/issue/ENG-2688/backup-service-reuse-single-blob-client-instance