Re: [quicwg/base-drafts] are flow control frames really idempotent? (#1612)

[Marten Seemann <[e-mail]>]

I think the sequence of values sent in packets 1-4 should be valid, and that an implementation MUST NOT error when receiving these.

When retransmitting a packet, you can do two different optimisations:

1. you can update the offset in the MAX_DATA frame, and thereby save a few bytes that you'd otherwise spend on sending a new MAX_DATA frame.
2. you can retransmit the MAX_DATA frame as is, and instead keep track that packet X is a retransmission of packet Y. When receiving an ACK for X, you then know that Y doesn't need to be retransmitted (and vice versa).

According to my measurements, 2. is a much better optimisation than 1. However, it can lead to exactly the sequence that you described. 
It might be possible to combine 1. and 2., but this will create **a lot** of complexity: If the retransmission of a packet contains an updated MAX_DATA frame, you can't just cancel the retransmission for that packet if the original packet is acknowledged, since you need to make sure that the MAX_DATA containing the updated offset is delivered reliably.

I see little value in enforcing that peers implement optimisation 1., since the expected performance gains are minimal, while 2. offers a real benefit on lossy connections.

-- 
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/1612#issuecomment-409083746