These change should improve privacy, discourage some resource wasting behavior from parties trying to attack user privacy, and provides additional node bandwidth control.
When Bitcoin Core was changed to not use fixed sleeps in network message handling this mostly broke the trickle logic: Instead of picking a new random node to bypass the delay every 100ms, it now chooses a new one every time through the message processing loop which can be much faster and can be remotely triggered.
As a result a vast majority of invs simply blew through without any delay.
This patch drops the old 1/4 random selection logic and elects two trickle peers which will get immediate forwards. These peers are selected at random from the top four peers ordered by their ability to take unfiltered relays, being outbound, being network nodes, and highest uptime. All other peers have their INVs triggered on a shared random delay (with a mean of 1 second).
The selection criteria tries to find a stable set of non-attacker controlled nodes that work and stick with them.
Two are used so that even if one was the original source of the only transactions ready to send the transactions will not be guaranteed to end their journey here. It also should improve robustness to dysfunctional nodes.
When the upload limit is passed, this responds by increasing the batching interval by another second. Preliminary testing suggests this significantly reduces the number of transaction getdatas the node receives.
Finally, this sorts transaction INV before sending them to reduce the potential for information to leak in INV ordering.
