Re: [quicwg/base-drafts] Required state for retaining unacked RETIRE_CONNECTION_ID frames is unbound (#3509)

[Kazuho Oku <[e-mail]>]

Just to give another example of how an endpoint might end up required to track the retirement of more than active_connection_id_limit:
* A server tries to supply three active CIDs to the peer. At the moment, it has provided CID<sub>0</sub>, CID<sub>1</sub>, CID<sub>2</sub> to the client.
* The client sends RCID(CID<sub>1</sub>),  RCID(CID<sub>2</sub>). The server provides NCID(CID<sub>3</sub>), NCID(CID<sub>4</sub>). The two NCID frames are delivered to the client, but the ACK for the RCID frames gets lost.
* The client decides to probe two paths concurrently, by using CID<sub>3</sub> and CID<sub>4</sub>, but shortly after the packets are sent using those paths, the underlying network for those two paths disappear.
* The client now needs to make sure that RCID frames carrying 4 CIDs reach the peer.

In this example, we might argue that it would be possible for the client to figure out that CID<sub>3</sub> and CID<sub>4</sub> were provided as substitutions for CID<sub>1</sub> and CID<sub>2</sub>. But do we want to implement logic that detects such condition into our code? Note that tracking of substitutions can become tricky; if RCID(CID<sub>1</sub>) and RCID(CID<sub>2</sub>) were sent in different packets, it would be impossible for the client to see if CID<sub>4</sub> was provided as a supplement for CID<sub>1</sub> or for CID<sub>2</sub>.

-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/quicwg/base-drafts/issues/3509#issuecomment-606408197