Enforce one transaction per notification in the networking

text/0056-one-transaction-per-notification.md

@@ -0,0 +1,96 @@
+# RFC-0056: Enforce only one transaction per notification
+
+|                 |                                                                                             |
+| --------------- | ------------------------------------------------------------------------------------------- |
+| **Start Date**  | 2023-11-30                                                                                  |
+| **Description** | Modify the transactions notifications protocol to always send only one transaction at a time|
+| **Authors**     | Pierre Krieger                                                                              |
+
+## Summary
+
+When two peers connect to each other, they open (amongst other things) a so-called "notifications protocol" substream dedicated to gossiping transactions to each other.
+
+Each notification on this substream currently consists in a SCALE-encoded `Vec<Transaction>` where `Transaction` is defined in the runtime.
+
+This RFC proposes to modify the format of the notification to become `(Compact(1), Transaction)`. This maintains backwards compatibility, as this new format decodes as a `Vec` of length equal to 1.
+
+## Motivation
+
+There exists three motivations behind this change:
+
+- It is technically impossible to decode a SCALE-encoded `Vec<Transaction>` into a list of SCALE-encoded transactions without knowing how to decode a `Transaction`. That's because a `Vec<Transaction>` consists in several `Transaction`s one after the other in memory, without any delimiter that indicates the end of a transaction and the start of the next. Unfortunately, the format of a `Transaction` is runtime-specific. This means that the code that receives notifications is necessarily tied to a specific runtime, and it is not possible to write runtime-agnostic code.
+
+- Notifications protocols are already designed to be optimized to send many items. Currently, when it comes to transactions, each item is a `Vec<Transaction>` that consists in multiple sub-items of type `Transaction`. This two-steps hierarchy is completely unnecessary, and was originally written at a time when the networking protocol of Substrate didn't have proper multiplexing.
+
+- It makes the implementation way more straight-forward by not having to repeat code related to back-pressure. See explanations below.
+
+## Stakeholders
+
+Low-level developers.
+
+## Explanation
+
+To give an example, if you send one notification with three transactions, the bytes that are sent on the wire are:
+
+```
+concat(
+    leb128(total-size-in-bytes-of-the-rest),
+    scale(compact(3)), scale(transaction1), scale(transaction2), scale(transaction3)
+)
+```
+
+But you can also send three notifications of one transaction each, in which case it is:
+
+```
+concat(
+    leb128(size(scale(transaction1)) + 1), scale(compact(1)), scale(transaction1),
+    leb128(size(scale(transaction2)) + 1), scale(compact(1)), scale(transaction2),
+    leb128(size(scale(transaction3)) + 1), scale(compact(1)), scale(transaction3)
+)
+```
+
+Right now the sender can choose which of the two encoding to use. This RFC proposes to make the second encoding mandatory.
+
+The format of the notification would become a SCALE-encoded `(Compact(1), Transaction)`.
+A SCALE-compact encoded `1` is one byte of value `4`. In other words, the format of the notification would become `concat(&[4], scale_encoded_transaction)`.
+This is equivalent to forcing the `Vec<Transaction>` to always have a length of 1, and I expect the Substrate implementation to simply modify the sending side to add a `for` loop that sends one notification per item in the `Vec`.
+
+As explained in the motivation section, this allows extracting `scale(transaction)` items without having to know how to decode them.
+
+By "flattening" the two-steps hierarchy, an implementation only needs to back-pressure individual notifications rather than back-pressure notifications and transactions within notifications.
+
+## Drawbacks
+
+This RFC chooses to maintain backwards compatibility at the cost of introducing a very small wart (the `Compact(1)`).
+
+An alternative could be to introduce a new version of the transactions notifications protocol that sends one `Transaction` per notification, but this is significantly more complicated to implement and can always be done later in case the `Compact(1)` is bothersome.
+
+## Testing, Security, and Privacy
+
+Irrelevant.
+
+## Performance, Ergonomics, and Compatibility
+
+### Performance
+
+Irrelevant.
+
+### Ergonomics
+
+Irrelevant.
+
+### Compatibility
+
+The change is backwards compatible if done in two steps: modify the sender to always send one transaction per notification, then, after a while, modify the receiver to enforce the new format.
+
+## Prior Art and References
+
+Irrelevant.
+
+## Unresolved Questions
+
+None.
+
+## Future Directions and Related Material
+
+None. This is a simple isolated change.