Subscriptions are currently implemented by running the user-specified query at regular intervals (1 sec by default) and only returning a response to the client if the result changes. The exact throughput numbers would depend on the type of query that you are running, the amount of data returned and the compute resources that you have.
Dgraph, in general, has very good query throughput so they should be scalable. Also, to support a large scale you can always shard data and bring up more Alpha nodes.
It should be ok to have a subscription return large amount of data, but you might want to configure the polling interval. If the result of your query is changing frequently and transmitting the data over the wire takes more than a second (default polling interval) then the updates that you would be receiving might be delayed.
We are also doing some performance benchmarking in the next few weeks to test some of this stuff. We’ll share the results from those as soon as we have them.