Open lcdr opened 1 week ago
Just having gzip'ed responses should reduce most of the overhead of JSON repeating syntax/names. There's also BSON if that's not sufficient.
I'd prefer not to roll our own format
true but it annoys me that the server first does a bunch of work to serialize key names which are then sent over the network only for the client not to need them in the first place
anyway this is very low priority so i doubt i'll get to this in any case
Currently GraphQL uses JSON as the network format. This works, but is inefficient, as each entry repeats its field names. Since the GraphQL client specifies the response structure, it already knows the field names. As long as there are no optional fields, this means we can transfer data in a much more compact manner by skipping field name serialization entirely and reconstructing them on the client.
In addition, in the case of arrays, we wouldn't need to send opening and closing braces and commas for each entry, since with a constant number of fields known in advance we can section the array on the client.
If we ever start supporting conditional fields we could still serialize those as standard JSON, resulting in a hybrid format that is both flexible and compact.
This would help reducing network overhead to something comparable to the SQL API's CSV format.