sequitur | Library of autoencoders for sequential data | Machine Learning library

We're creating an IM app for iOS devices, and we're using APN notifications to inform the user each time one of their chats has a new message. Reading the documentation, Apple advises, "Because the delivery of remote notifications is not guaranteed, never include [...] data that can be retrieved by other means in your payload."

This seem like a bit of a non sequitur to me. Just because the data can be retrieved by other means, is this a reason not to put it in the notification payload? If we include chat message bodies in our notification payload (where the size of the body is going to be no larger than, say, 1KB), we can cache the message and display it as soon as the user opens the app, instead of the app having to send off to the server for the message, introducing an extra delay.

Sure, APS notifications may come out of order, and delivery isn't guaranteed, so we'd use message dates to order the messages and call the server to get any messages that weren't delivered through APS. But for the messages that did get through via APS, I can't see why we wouldn't just include the entire message body in the notification.

Apple's documentation gives the example of an email, where the email body would not be delivered in the APN body but downloaded separately by the app. However, emails are much larger than IM bodies, typically, and can be multiple megabytes in size. Our IM bodies would be much smaller, so that isn't a good example.

Am I missing something in that there is a technical reason not to include such smallish IM message bodies? It kind of makes me wonder why notifications can be up to 4KB in size if you're not supposed to bundle this kind of stuff.