- From: James M Snell <jasnell@gmail.com>
- Date: Sat, 21 Jul 2012 15:11:01 -0700
- To: Roberto Peon <grmocg@gmail.com>
- Cc: ietf-http-wg@w3.org
- Message-ID: <CABP7RbcT31WXTDYp5r6Wqj7isGYY+Eie1YjMepAFLWjigZ9uXg@mail.gmail.com>
On Jul 21, 2012 12:41 PM, "Roberto Peon" <grmocg@gmail.com> wrote: > > > On Jul 21, 2012 12:10 PM, "James M Snell" <jasnell@gmail.com> wrote: > > > > Continuing my review of the SPDY draft... have a few questions relating to SPDY and load balancers / reverse proxy set ups... The intent is not to poke holes but to understand what the SPDY authors had in mind for these scenarios... > > Poke away! > > > > > 1. Imagine two client applications (A and B) accessing an Origin (D) via a Reverse Proxy (C). When a client accesses /index.html on Origin D, the Origin automatically pushes static resources /foo.css, /images/a.jpg and /video/a.mpg to the client. > > > > Basic flow looks something like... > > > > A RP O > > | | | > > | | | > > |==================>| | > > | 1)SYN | | > > |<==================| | > > | 2)SYN_ACK | | > > |==================>| | > > | 3)ACK | | > > |==================>| | > > | 4)SYN_STREAM (1) | | > > | |================>| > > | | 5) SYN | > > | |<================| > > | | 6) SYN_ACK | > > | |================>| > > | | 7) ACK | > > | |================>| > > | | 8) SYN_STREAM(1)| > > | |<================|-- > > | | 9) SYN_STREAM(2)| | > > | | uni=true | | > > |<==================| | | > > | 10) SYN_STREAM(2) | | | > > | uni=true | | | Content Push > > | |<================| | > > | | 11) SYN_REPLY(1)| | > > |<==================| | | > > | 12) SYN_REPLY(1) | | | > > | | | | > > | |<================| | > > |<==================| 13) DATA (2,fin)|-- > > | 14) DATA (2,fin) | | > > | | | > > | | | > > > > My question is: what does this picture look like if Client's A and B concurrently request /index.html? > > > > With HTTP/1.1, static resources can be pushed off to CDN's, stored in caches, distributed around any number of places in order to improve overall performance. Suppose /index.html is cached at the RP. Is the RP expected to also cache the pushed content? > > It could if the pushed content indicated it was cacheable using http/1.2 caching semantics/headers. > > > Is the RP expected to keep track of the fact that /foo.css, images/a.jpg and /video/a.mpg were pushed before and push those automatically from it's own cache when it returns the cached instance of /index.html? > > No, but a smart implementation would, cache headers allowing... > Ok... so what if the cache headers don't allow? e.g. /index.html is cacheable, /images/foo.jpg is not... /index.html would be served from the cache and images/foo.jpg would never be pushed unless the cache is smart enough to know that it has to go grab it and push it to the client. Not saying that's a bad thing, just acknowledging that for push to work, caches and proxies would need to become much more aware of the content that's being passed around. > > If not, when the caching proxy returns index.html from it's cache, A and B will be forced to issue GETs for the static resources defeating the purpose of pushing those resources in the first place. > > > > In theory, we could introduce new Link rels in the same spirit as http://tools.ietf.org/html/draft-nottingham-linked-cache-inv-03 that tell caches when to push cached content... e.g. > > Yup > > > > > SYN_STREAM > > id=2 > > unidirectional=true > > Content-Location: http://example.org/images/foo.jpg > > Content-Type: image/jpeg > > Cache-Control: public > > Link: </index.html>; rel="cache-push-with" > > > > What does cache validation look like for pushed content? E.g. what happens if the cached /index.html is fresh and served from the cache but the related pushed content also contained in the cache is stale? > > Same as any normal cache expiry. > Just want to make sure I'm clear on it... Cache wants to push /images/foo.jpg to the client, notices that the cached version is stale, it has to send a GET to the origin to revalidate, it gets back the updated version then pushes that version back to the client. > > > > I'm sure I can come up with many more questions, but it would appear to me that server push in SPDY is, at least currently, fundamentally incompatible with existing intermediate HTTP caches and RP's, which is definitely a major concern. > > Given that the headers on the pushed objects indicate cacheability I'm either misunderstanding or missing something :) > > > > > As a side note, however, it does open up the possibility for a new type of proxy that can be configured to automatically push static content on the Origin's behalf... e.g. A SPDY Proxy that talks to a backend HTTP/1.1 server and learns that /images/foo.jpg is always served with /index.html so automatically pushes it to the client. Such services would be beneficial in general, but the apparent incompatibility with existing deployed infrastructure is likely to significantly delay adoption. Unless, of course, I'm missing something fundamental :-) > > Jetty has done this already publically, btw. > Good to know, thank you. I'd missed that. I'll dig into their implementation shortly to see what they've done. > > > > 2. While we on the subject of Reverse Proxies... the SPDY spec currently states: > > > > When a SYN_STREAM and HEADERS frame which contains an > > Associated-To-Stream-ID is received, the client must > > not issue GET requests for the resource in the pushed > > stream, and instead wait for the pushed stream to arrive. > > > > Question is: Does this restriction apply to intermediaries like Reverse Proxies? > > Yup. RPs are both servers and clients (depending on the side one is on) and the respective requirement are still in force. > > > For instance, suppose the server is currently pushing a rather large resource to client A and Client B comes along and sends a GET request for that specific resource. Assume that the RP ends up routing both requests to the same backend Origin server. A strict reading of the above requirement means that the RP is required to block Client B's get request until the push to Client A is completed. > > Only per connection. The intention and purpose of this requirement is to eliminate the potential race whereby the server is pushing an object, but the client is requesting it as well (because it doesn't yet know the object is already on its way). > Ok, yeah I understand the reasoning behind it. The spec should clarify that the restriction is only per connection. Even still, I can see a number of ways where this blocking could potentially trip developers up if they don't understand or aren't fully aware of how it works. Namely, developers would need to avoid pushing relatively large and/or non-cacheable resources via a single URI unless those resources are ALWAYS pushed and not intended to be retrieved directly by GET. Quite interestingly... server push opens up the possibility of push-only resources that do not allow any direct requests. That is, I could set up an server to always push /images/foo.jpg while blocking all methods including GET. Obviously that would have "interesting" cache validation issues.... I'm not saying this would be a good idea, just that it would be possible. hmmm..... > > Further, the spec is not clear if this restriction only applies for requests sent over the same TCP connection. Meaning, a strict reading of this requirement means that even if the RP opens a second connection to the Origin server, it is still forbidden to forward Client B's GET request until Client A's push has been completed. > > We should clarify that the invariant here exists to prevent the race mentioned above and any behavior is allowed so long as that race is prevented. > +1 Overall, the spec could stand a lot more clarity with regards to intermediaries and caching. There is quite a bit that is implied and after a bit of noodling makes sense but it would be great if some of those things were spelled out a bit more explicitly. > -=R > > > > > > > - James > > > > > > On Jul 21, 2012 12:10 PM, "James M Snell" <jasnell@gmail.com> wrote: >> >> Continuing my review of the SPDY draft... have a few questions relating to SPDY and load balancers / reverse proxy set ups... The intent is not to poke holes but to understand what the SPDY authors had in mind for these scenarios... >> >> 1. Imagine two client applications (A and B) accessing an Origin (D) via a Reverse Proxy (C). When a client accesses /index.html on Origin D, the Origin automatically pushes static resources /foo.css, /images/a.jpg and /video/a.mpg to the client. >> >> Basic flow looks something like... >> >> A RP O >> | | | >> | | | >> |==================>| | >> | 1)SYN | | >> |<==================| | >> | 2)SYN_ACK | | >> |==================>| | >> | 3)ACK | | >> |==================>| | >> | 4)SYN_STREAM (1) | | >> | |================>| >> | | 5) SYN | >> | |<================| >> | | 6) SYN_ACK | >> | |================>| >> | | 7) ACK | >> | |================>| >> | | 8) SYN_STREAM(1)| >> | |<================|-- >> | | 9) SYN_STREAM(2)| | >> | | uni=true | | >> |<==================| | | >> | 10) SYN_STREAM(2) | | | >> | uni=true | | | Content Push >> | |<================| | >> | | 11) SYN_REPLY(1)| | >> |<==================| | | >> | 12) SYN_REPLY(1) | | | >> | | | | >> | |<================| | >> |<==================| 13) DATA (2,fin)|-- >> | 14) DATA (2,fin) | | >> | | | >> | | | >> >> My question is: what does this picture look like if Client's A and B concurrently request /index.html? >> >> With HTTP/1.1, static resources can be pushed off to CDN's, stored in caches, distributed around any number of places in order to improve overall performance. Suppose /index.html is cached at the RP. Is the RP expected to also cache the pushed content? Is the RP expected to keep track of the fact that /foo.css, images/a.jpg and /video/a.mpg were pushed before and push those automatically from it's own cache when it returns the cached instance of /index.html? If not, when the caching proxy returns index.html from it's cache, A and B will be forced to issue GETs for the static resources defeating the purpose of pushing those resources in the first place. >> >> In theory, we could introduce new Link rels in the same spirit as http://tools.ietf.org/html/draft-nottingham-linked-cache-inv-03 that tell caches when to push cached content... e.g. >> >> SYN_STREAM >> id=2 >> unidirectional=true >> Content-Location: http://example.org/images/foo.jpg >> Content-Type: image/jpeg >> Cache-Control: public >> Link: </index.html>; rel="cache-push-with" >> >> What does cache validation look like for pushed content? E.g. what happens if the cached /index.html is fresh and served from the cache but the related pushed content also contained in the cache is stale? >> >> I'm sure I can come up with many more questions, but it would appear to me that server push in SPDY is, at least currently, fundamentally incompatible with existing intermediate HTTP caches and RP's, which is definitely a major concern. >> >> As a side note, however, it does open up the possibility for a new type of proxy that can be configured to automatically push static content on the Origin's behalf... e.g. A SPDY Proxy that talks to a backend HTTP/1.1 server and learns that /images/foo.jpg is always served with /index.html so automatically pushes it to the client. Such services would be beneficial in general, but the apparent incompatibility with existing deployed infrastructure is likely to significantly delay adoption. Unless, of course, I'm missing something fundamental :-) >> >> 2. While we on the subject of Reverse Proxies... the SPDY spec currently states: >> >> When a SYN_STREAM and HEADERS frame which contains an >> Associated-To-Stream-ID is received, the client must >> not issue GET requests for the resource in the pushed >> stream, and instead wait for the pushed stream to arrive. >> >> Question is: Does this restriction apply to intermediaries like Reverse Proxies? For instance, suppose the server is currently pushing a rather large resource to client A and Client B comes along and sends a GET request for that specific resource. Assume that the RP ends up routing both requests to the same backend Origin server. A strict reading of the above requirement means that the RP is required to block Client B's get request until the push to Client A is completed. Further, the spec is not clear if this restriction only applies for requests sent over the same TCP connection. Meaning, a strict reading of this requirement means that even if the RP opens a second connection to the Origin server, it is still forbidden to forward Client B's GET request until Client A's push has been completed. >> >> >> - James >> >>
Received on Saturday, 21 July 2012 22:11:31 UTC