- From: Kazuho Oku <kazuhooku@gmail.com>
- Date: Thu, 11 May 2017 20:35:33 +0900
- To: Stefan Eissing <stefan.eissing@greenbytes.de>
- Cc: Willy Tarreau <w@1wt.eu>, Mark Nottingham <mnot@mnot.net>, Erik Nygren <erik@nygren.org>, "ietf-http-wg@w3.org Group" <ietf-http-wg@w3.org>, "Ponec, Miroslav" <mponec@akamai.com>, "Kaduk, Ben" <bkaduk@akamai.com>
2017-05-11 20:33 GMT+09:00 Stefan Eissing <stefan.eissing@greenbytes.de>: > >> Am 11.05.2017 um 13:31 schrieb Kazuho Oku <kazuhooku@gmail.com>: >> >> 2017-05-11 17:19 GMT+09:00 Stefan Eissing <stefan.eissing@greenbytes.de>: >>> >>>> Am 11.05.2017 um 07:33 schrieb Willy Tarreau <w@1wt.eu>: >>>> >>>> Hi Mark, >>>> >>>> On Thu, May 11, 2017 at 10:23:12AM +1000, Mark Nottingham wrote: >>>>> If an origin doesn't have robust retry/replay protection in place for >>>>> non-idempotent requests, it seems operationally simpler and safer for them to >>>>> disable 0RT, rather than refusing it on a request-by-request basis. That's >>>>> the discussion I think we should have here... >>>> >>>> That's exactly the situation I'm facing for now with haproxy. A few >>>> users have asked us to support 0RTT and by lack of way to 1) decide >>>> which requests are really safe, and 2) tell the client it must replay >>>> them using 1RTT, for now I refused to enable it. The load balancer >>>> and the origin server will have a different view of the acceptability >>>> of 0RTT, and all the chain must be able to accept or reject them, and >>>> let the client retry. >>> >>> Even the "origin server" might not be aware what the application's >>> committal and guarantee here is. >>> >>> My thoughts for an implementation is: >>> >>> - it has to work without the "upper" layer / next hop being aware of it >>> - it has to fail in a defined HTTP way. The HTTP request is tagged as >>> possibly replayed, regardless of the actual transport. The answer >>> needs to also work on that transport. >>> - The negative answer to a 0-RTT request might come early, might come >>> late. For h2, other streams might have been opened, even answered, >>> in the meantime. >>> - The sender selecting 0-RTT should only do so, if it understands the >>> retry answer. (Once that is defined) >>> - The sender may well want to select 0-RTT only if it considers the >>> data to be safe for replays *and* it expects the server to come to >>> the same conclusion. >>> - So, ideally, sender and receiver have the same notion about what HTTP >>> data is acceptable for 0-RTT. >> >> This is an interesting discussion! >> >> I believe that there is no need for us to require a _client_ to resend >> a HTTP request, even in case it sends a HTTP request in 0-RTT and then >> turns out that the application running behind tells the "origin >> server" that it cannot handle 0-RTT request. >> >> IMO what the origin server should do is buffer the 0-RTT request >> (note: in TLS 1.3, a server can cap the size of 0-RTT data), and if >> the application refuses to handle the request due to the fact that it >> has been sent in 0-RTT, wait until the client proves itself to be a >> legitimate client (by sending an 1-RTT data), and then resend the >> buffered request to the application. > > Hmm, how many RTTs will this proof take? 1RTT. The latency will be the same as when the client did not use 0-RTT. OTOH, the obvious benefit of the proposed approach is less use of bandwidth since there is no need for a client to resend the request. > > >> In HTTP/2, the proof can be obtained by sending a PING frame from the >> server after sending ServerFinished message (of TLS 1.3) and waiting >> for the response to the PING frame. >> >> So, while I agree that it is beneficial to have an agreement on how >> the interaction scheme between the origin server and the application >> running behind (possibly as an informational RFC), I do not see a >> strong reason that we need to introduce some kind of profile due the >> introduction of 0-RTT data in TLS 1.3. >> >>> -Stefan >>> >>>> >>>> I tend to think that a 4xx status code would make sense and would be >>>> useful to pass the verdict back to the client. For example we could >>>> return "418 not idempotent". >>>> >>>> Willy >>>> >>> >>> >> >> >> >> -- >> Kazuho Oku > -- Kazuho Oku
Received on Thursday, 11 May 2017 11:36:06 UTC