Attendees
- Mark Nottingham (Akamai)
- Jun Fujisawa (Canon)
- Herve Ruellan (Canon)
- Patrick McManus (Mozilla)
- Eliot Lear (Cisco)
- Martin Thomson (Microsoft)
- Alexey Melnikov (Isode)
- Barry Leiba (Applications AD)
- Erik Kline (Google)
- Tatsuya Hayashi (Lepidum)
- Kazuki Shimizu (Lepidum)
- Yusuke Kagiwada
- Adam Rice (Google)
- Takashi Toyoshima (Google)
- Emile Stephan (Orange)
- Rajeev Bector (Yahoo!)
- Ben Niven-Jenkins (Velocix)
- Grant Watson (Velocix)
- Hasan Khalil (Google)
- William Chan (Google)
- Roberto Peon (Google)
- Ilya Gregorik (Google)
- Mike Belshe (Twist)
- Jeff Pinner (Twitter)
- Brian Raymor (Microsoft Open Tech)
- Gabriel Montenegro (Microsoft)
- Takeshi Yoshino (Google)
Martin Thomson: one draft - SPDY with name change. new abstract into, addition of some conclusions on upgrade/negotiation. No conclusion on TLS side yet
Still some discussion between authors on framing and how to explain etc.
Mark Nottingham: need to talk about terminology as not clear in document right now.
Martin: current draft is fouled on minimal info to do an implementation but lacking in introduction & "fluff".
Mark: asked TLS WG to work on mechanism to upgrade within TLS which may or may not look like NPN. Have pinged for status but not heard back.
Eliot Lear: see no reason why to normatively block on that which you are in the draft
Martin: that can be removed if you want to weasel around it.
Mark: If we're ready to publish and TLS WG are not ready we can take it out and not block on it so we can take that decision (whether TLS upgrade is normative) later.
Mark: are people comfortable with it? (Text needs some cleanup but not discussing editorial stuff)
Mark would like to see it discussing things in a more fine grained way - e.g. upgrade to HTTP/2.0 but using TLS or HTTPS etc. Might want to put that (large set of constraints)
So could use token as a sort of "profile"
Alexey Melnikov: are you suggesting some structure to token or just more extensive list.
???: What is purpose
Have a HTTP URI and want to use HTTP/2 and TLS to retrieve resource. Client might support 1.1 & 2.0 over TCP and 2.0 over TLS and negotiation between client server and do we want to use the Version ID token to signal that sort of thing?
Token could be used in upgrade, as part of NPN etc.
Eliot: I thought you were referring to the token used in the method/first line.
Mark: that's 1.1 we may or may not do it that way
Gabriel Montenegro: some scope overlap with HSTS?
Mark: this is what are the things I can do next on this connection?
Patrick McManus: this is essentially directory/policy information for what is available.
Eliot: more like capabilities
Mark: see as way to drive convergence - if have named 'profiles' browsers can converge on what they support & server can do the same. People who define 'profiles' define how fine grained they are.
Roberto Peon: discussed internal to Google. We've never seen enough demand to actually do it, so safest thing is describe how one might accomplish it.
Mark: defining those profiles is not something HTTPBIS wants to get into but should we leave that door open?
Eliot: don't want a lot of these things as the more you have the more interop problems you'll have.
Martin: thing that bothers me is we're talking at two levels - top level identification and then things that exist in protocol (e.g. foo transported within HTTP/2). the latter is negotiated feature once decided to use HTTP/2. Whereas what Mark is talking about is "HTTP/2 with this set of options".
Roberto: e.g. client could be browser and may want to indicate whether want/support Websockets. Don't want to have wasted a RTT doing something else
Martin: so Q is have this negotiation upfront, or just negotiate HTTP/2 and let the client try these other things and have them fail like today.
???: is there a difference between upgrade on session and on a per-stream basis
Roberto: can't see this applying per stream, capability is per-session.
Will Chan: talked about it within NPN which just sees opaque strings
Roberto: killer app is HTTP so we haven't seen a need for it up to now.
Mark: overall goal - we want to do this right once.
Roberto: so long as define a portion of this as opaque string which we don't define right now, then we have portion of version string we can put things in for HTTP/3 etc. E.g. can have curly braces in version string and everything in there is opaque.
Mark: reminds me of media types - defines a format but people want to put a scheme etc in it too and we don't do that for media types and there's a reason for that.
Roberto: only concrete thing I can think of right now is whether websockets is available or not.
Martin: doesn't seem to be any kind of registry for these opaque strings.
Mark: sounds like we're talking about an opaque string where we define just a handful. Sounds like we need somebody to start running things down and thinking through the use cases.
No-one volunteers to do that.
Mike Belshe: how about we don't do this at all - opaque string is not interoperable. Unless implementation understand it it won't get used as they don't know what to do with it. So without something concrete specifying it is worse than specifying nothing.
Mark: have number of upgrades mechanisms. However you identify capabilities/tokens they should be the same for the different upgrade mechanisms - e.g. same for NPN & upgrade, etc.
Gabriel?: maybe have some text in HTTP/2 draft and then separate draft(s) e.g. "this is how websockets us it" etc. Need to go through example use case
Mark: sound silk use case are http/1.1 vs 2.0, negotiating TLS in some circumstances, HTTP Vs websockets. As long as mechanism addresses those use cases that's all we need to write down.
???: Little lost, What are we trying to define?
Mark: trying to define shape of strings and place in ecosystem and couple of mechanisms like DNS/NPN/Upgrade that use those strings. Then we'd define a couple of use case for those strings.
Mark: might help if in our discussion HTTP/2 is used to main framing layer & HTTP messages over that framing layer.
???: Need to be careful with websockets as that has it's own 'dance' for negotiating capabilities.
Roberto: one of reasons we haven't done this to date is larger you make the string the more you slow things down.
Mark: Of all upgrade mechanisms, if already know what protocol you want to use (e.g. HTTP vs websockets) do we need to worry about that in this exchange.
Roberto: how does content provider test the waters before deploying anything otherwise they won't. E.g. I'm a bank I might want to say "I want to use HTTP/2 but for X type of transaction I want to use TLS+HTTP/1.1 because it's better known/signed off by my security guys/etc"
Eliot: this is a capability exchange and then you apply local policy.
Mark: we negotiate/advertise versions and transports?!? and then just leave it there?
Martin: easy if we consider two use cases we have today - straight over TCP and TLS. The one that bothers me is use of TLS to request http:// URIs, that implies either use of another label or some other sort of signalling.
Mark: not quite as simple, e.g. if I use Alternate-Protocol.
Roberto: today if using SPDY you have accepted idea you can send both http:// and https:// schemes over same session.
Roberto: some of this string might not be useful because you may already know
- e.g.. if use NPN to negotiate you know you're already over TLS. Someone should go through use cases and have a mechanism that defines "this part of the string you don't interpret unless you know what you're doing" and call it a day.
Roberto: I will try and do it but that is not a promise.
Mark: what does version part in HTTP/2 frame mean?
Roberto: that was put in because thought there may be a proxy and may be receiving frames from different clients talking different versions and would be cute not to have to change the framing, just forward it and the end point can figure out what's going on, and it's turned out not to be very useful, so we should just get rid of it.
Mark: HTTP/1 versioning is hop by hop so don't need to persist in message. If we get rid of version in HTTP/2 message/frame some people will complain it's not self describing. I don't particularly care I'm just commenting there will be some pushback.
Gabriel: maybe have a mechanism so server can tell from first frame (magic number/byte/etc) that allows server to disambiguate the protocol.
Roberto: does this come from hybi approach - if can fast fail that is a good thing.
Mark's suggestion - do something like send a first line of "NULL * HTTP/2.0"
Patrick: if you send anything but HTTP/1 over port 80 you get ???? (a mess??)
Mark: use case as I understand it - if I have a server farm and I've only upgraded half of it I want to fast fail. Is that use case good enough to burn 17 bytes at beginning of every connection for all time?
Mike: proposal is get rid of version in frames and instead put something in the beginning of the connection?
Martin: two things websockets had that protected ???? - non-ASCII and other is look at response to upgrade that shows in the positive that server had not only understood but had acted on the upgrade.
Mike: one option is to always send a SETTINGs frame first
Jeff Pinner: Basically can we kill version in the frame and then fail fast.
Roberto: like sending SETTINGS frame first, don't want to restrict ourselves to whatever framing is guaranteed to make an existing HTTP/1 implementation barf, e.g. if first bytes of SETTINGS frame happen to be "GET".
Will: can we step back so I can understand the use case? Is it only in enterprise setting where you know the out of band negotiation scenario or will this be done out in the open? I don't see browsers doing this out in open.
Mike: declare version upfront without ??? If we're going to have a magic string or whatever upfront it should be constrained to just indicating the version.
???: Another thing we might want is info on user agent, e.g. have a broken version of chrome wrt SPDY and we know Chrome XYZ has this bug and would be nice to say "this version of chrome is broken and so we're going to behave slightly differently"
Mark: one nice thing about that is you would pre-seed the User Agent into your compression context.
Martin: a lot of protocols have a session start message/thing.
Roberto: have 'magic bits' and then after magic bits need to make sure we have mechanism to allow client to send settings/metadata/whatever.
Need some magic we send once per session/connection, question is show much. Need to make reliable enough so it makes things fail but short enough that it doesn't hurt us.
Mark: would be nice if had easier ways to get answers, like a 1 file python script can put on git and have people run it in weird places like airport on way home, etc. to gather information.
Will: Happy in Chrome to do this kind of probing, we have done it in the past.
ACTION: Come up on list of shortlist of reasonable client sequences to test.
Conclusion:
- Get rid of version field in control frames
- Invest in researching 'magic' at the beginning of a connection
- Talk more on list & get proposal for how to identify protocols
- Do some testing
Also agreed (textual) token/string should be opaque and same between all different upgrade mechanisms. Exact format/structure still TBD. How it applies in different situations may differ.
Eliot: registry - HTTP/2 draft explains what's required to get into registry etc. I will volunteer to write that text.
Will: two different directions of negotiation - Upgrade is client side, NPN it's client indicates extension support and server advertises protocols and client selects what it wants to use.
Mark/Eliot - what's the impact there?
Mark: Alternate Protocol is that something folks still want to do or has it fallen by the wayside?
Will: Don't think there is a lot of active interest but keep in the back of mind as alternative option. Happy not to pursue right now, but if I get unhappy with other options I might bring it back up.
Martin: Not going to reduce to a single upgrade mechanism but would like to keep to as few as possible.
Jeff: only thing I can think that Alternative-Protocol gives you the other don't can be solved with a redirect (although A-P is nicer).
Mike: redirect is not equivalent.
Implementation status - Chrome has it, Mozilla have it (but a little buggy).
Mark: Punt on A-P right now, see how things go, if browser or server folks start to feel it is something we need we can talk about it then, but if we don't need it that's nice as one less thing to define.
Will: would like to reserve judgement until we've discussed DNS more as it seems to have similar properties.
Patrick: Bias towards leaving it in for now.
Mark: but we don't have text right now.
Eliot: draft-lear-httpsbis-srvinfo
Eliot: Comparison of different use of DNS, some commonalities, some things we should do no matter what.
Eliot: 1 motivation is to avoid a RTT, so let's avoid just moving that RTT to another protocol.
DNS not a complete replacement for Upgrade as might not have DNS at your disposal. Can not just assume DNS is there.
One question for the end - is discovery of transport protocol important?
Unstated design goal - if going to HTTP/2.0 then goal is for everyone to get there eventually.
SRV as an approach - SIP makes use of it, jabber uses it, MS AD servers make heavy use of SRV. Many folks do a zone cut at _tcp.example.com and so have two sets of authoritative name servers so may need to ask both of them to get the answer you are looking for. Any additional info in response is considered non-authoritative and many clients throw it away to get authoritative info. If don't have additional info, need to do an additional query to get that info.
Mark: summary - because of what SRV is used today, may be zone splits and trying to minimise number of queries may not be easy.
Patrick - what is split DNS?
Mark: one DNS server inside the firewall and another outside.
Eliot: NAPTR and URI as an approach - Builds on SRV, allow transport protocol discovery but not protocol version.
Martin: it can do using the "label" you select on.
Eliot: haven't seen that but you can layer on top of it and end up with some capability to do something like that.
"Running a race" as an option. Browser folks are used to doing that - e.g. Happy Eyeballs. Could do something similar for HTTP/2. Not a bad idea but on its own you still need a capabilities advertisement.
Roberto: likely to be more expensive and end up with two connections.
Martin: depends how far through the connection (like TLS handshake etc) you get.
Eliot: it is something to keep in mind as I think it would be handy.
Martin: what would you be racing?
Eliot: may send multiple queries for multiple info.
New record Eliot put together - SRVINFO
Couple of people raised maybe we combine protocol & version and call it 'profile' or something. There may be some tradeoffs
Because there is no domain name on right hand side have info you need as getting A record - same domain so guaranteed to be the same authority. No risk of required sequential lookups, may need to query in parallel for A/AAAA record.
Are some issues not using underscores, e.g. cnn.com if someone has whole bunch of web servers could get a very long list.
Doing some testing at Cisco, SRV is guaranteed to end up using TCP for DNS queries because of the way Cisco do AD load balancing.
Martin: dropped weight can you explain why?
Eliot: did anyone seriously use it? Don't have opinion, if you think it's useful send something out saying why, but I don't think it's useful operationally.
Martin: I can write something up.
???: something you can do with weight you can't do with more priorities?
Martin: yes as priority is strictly ordered, no way to say these are all equal priority.
Roberto: biggest issue with weight etc is not everyone respects TTL.
Eliot: tell me if you think weight is useful as it didn't seem to be to me and others I spoke to.
Introduced instance ID - idea is have URI you are starting with and implicitly/explicitly have a port you will connect to. That port may be running multiple protocols. But if you want the same service on a different port you need an index to tell you that and that is what instance ID is.
Mark: implies server has to always send something with port 80 in, so I can say here's the 80 port with this instanceID and this other port has the same instance ID so I know it is the same service.
Eliot: yes, need to process whole record. Meets the requirement to run multiple servers off the same host.
Jeff: seems more like a discovery information
Eliot: it is a capabilities discovery mechanism (it is not a negotiation).
Jeff: server has no way of pulling it back once advertised
Martin: DNS TTLs.
Eliot: let's start from the premise that this isn't the first thing you do.
Mark: hopefully when you advertise this and try a HTTP/2 connection it fails fast so client can fallback to HTTP/1
Martin: this might be the first thing you try as a client but may not be the first thing you deploy as a server.
Mark: raises important point - feedback i've gotten is it is important to us to provide guidance to folks who want to deploy HTTP/2 to help hold their hands etc. Anyone interested in helping with such a document (would be separate document) I would love to hear from you.
Mark: Summary: There are a lot of caveats with DNS and new records etc. Not saying we won't do A-P because this is so great, just we're going to run with DNS for the time being.
Eliot: have shown SRVINFO & A as parallel queries. DNS in theory supports multiple questions in same query but don't think anyone has ever deployed it.
Mark: from web standpoint the thing I'm fundamentally connecting to is an origin which is tuple of (scheme, URI, port), whereas this looks like indirection layer
Eliot: actually an equivalence, not an indirection.
Will: browser optimising for latency would issue 3 queries in parallel - SRVINFO, A & AAAA
Eliot: one thing noted in draft - absent DNSSEC it is a bad idea to mix security models with this. Eliot - could include 'profile' (based on token talked about earlier) that indicates that
Mark: instanceID seems convoluted. Can we simplify it?
Roberto: we did, if we put the token/NPN pattern at the end of this then we've normalised what sysadmins have to write to configure, got rid of instance ID as can state multiple things in that pattern?
Eliot: no but could chance instanceID for a mnemonic name.
Jeff: asked questions around how it works with CNAMEs that I missed.
Eliot: this doesn't change how you advertise certificates on hostnames for TLS. IF you want to preserve name across then SRV does that for you.
Eliot: continuing presentation.
All this info is cached and I was viewing as positive thing, but talking to ??? he mentioned it could be negative.
Summary:
- Clients don't know where zone cuts are and you can't make assumption about that
- DNS is one of 3 approaches to provide info prior to connection (other being new URI & HTML).
Questions:
- Is the optimisation worth it? Not a place to start, but a place to go to but a lot of mechanism here. Eliot hasn't formed his own opinion on that yet.
Mark: not just that - increasing deployment flexibility, giving some additional ???
Jeff: and can get to to HTTP/2 only
Eliot: not thought through how this works in a world without HTTP/1.1 I have not thought that through and is something the draft should discuss.
- Are people interested in using HTTP over things other than TCP?
- Couple of people nodded their heads.
Eliot: if talking SRV then doing multiple queries (_tcp, _foo, etc) but those would presumably be done in parallel.
Eliot: I implemented SRVINFO in BIND, so not that hard to do.
Mark: I have a home router with a DNS proxy in, is it going to filter this?
Will: 1) a new record - we see many cases where it completely fails, (2) for substantial portion of user base are using cheap routers, e.g. issue with Comcast where some of their (2wire?) routers will only allow 6 parallel lookups at any time. So if for performance optimisation, it might end up being a performance cost.
Eliot: one design goal is not to impact app performance and this isn't something you introduce 6 months from now. Problem is common to DNS, not this record.
Mike: interesting stuff but when I think about all the things we still need to talk about then this isn't interesting at all. Can we table and look at it again in 6 months.
Eliot: need implementation experience so need to do that before we actually need it, so might want to start earlier.
Mark: hearing doubts on using DNS and whether a new record is a good idea.
(break for lunch)
Any interest in using an existing record type?
Will: We have data on TXT records for latency and success rates, which should improve if we start using it.
Will: there will be a hump that this introduces if this is a serial dependency
- we would have to start with non-blocking [non-serial]. Happy eyeballs style.
Mark: would upgrade mechanisms be interesting for prototypes and testing?
Patrick: upgrade would be one of the first things that I want to test
Patrick: (re: new record types) I'd like to get some data on success rates, etc...
Mark: structured TXT? How is that likely to be a problem.
Eliot: structured TXT might suffer the zone cut issue, but these would be on an underscore prefixed sub-domain, but this doesn't have the same sort of deployment lock-in that _tcp does. DNS directorate loves new record types, but there are some operational concerns.
Will: why do we care about failure rates? We know that there are failures (~4-5%)
Martin: I can find some more stats on DNS success rates from some other IETF work http://www.icann.org/en/groups/ssac/documents/sac-035-en.pdf
Patrick: Are your stats for large requests that required TCP? Will: don't think so
Eliot: What plans do people have for TLSA? Is DNS so hosed that we can never get another RR defined?
Will: Relying on something from the start might not be wise, but once we start trying to use it, as long as there are no serious problems doing so, we can use it and things will improve.
Patrick: I'd be interested in how badly a new RR underperforms TXT.
Mark: we need to get this data and then we'll have the info to choose between new RR and packing stuff into TXT.
Eliot: the SVCINFO to TXT encoding is simple
Will: when are we going to come back to discussing Alternate-Protocol?
Mark: we should try to get the first things done first, unless someone feels strongly about it.
Tour of: https://github.com/http2
Presentation by proxy of https://docs.google.com/presentation/d/1x8GvY-7FJi57DW9vSvjTF1QnTkzM18mXi_LQUUprZeo/view#slide=id.p
Mike: the primary goal for SPDY was latency reduction; getting perfect or near-perfect compression isn't necessary to achieve that end
Goals for header compression include:
- privacy
- CRIME ATTACK
- Latency reduction
- head-of-line blocking?
- Compression efficiency
- CPU overhead
- Memory consumption
- initial
- idle
- under DoS
- receiver control
- implementation complexity
- proxyability
- error handling
- ability to define new headers and maintain properties above
Editors note: we will need to capture this: the header compression algorithm (and many other aspects of the protocol) can only really be changed by also changing the entire protocol - the theory being that the specification and deployment of a feature like this incurs pain roughly equivalent to an entire protocol revision. At a minimum, we can turn off the compressor.
On the question of header types and the registration policies for any registry.
Mark (indiv): we want to limit the number of types to avoid the need for bespoke parsers for every new header field.
Barry Leiba: suggest standards action for new types
Mark/others: don't need that, don't want even the hint of that.
Jeff: what are the implications for proxies
Martin: the upper bound on state at an intermediary can be defined by the server if you use passthrough
Roberto: you can act as a server and advertise a lower limit
Martin: state for headers has implications for servers; e.g. if a request is rejected, the server is required to parse and apply any state changes that the request established
Roberto: yes, it's a trade-off
Jeff: is there any desire to make this non-HTTP?
Mark: I'm not encouraging this
Jeff: Reusable code for the compressor would be good; zlib is easy to integrate
Ilya Grigorik: it would be good to have a good test suite
Mark: modular code would allow for reuse of portions of working implementations in new implementations
Mark: are we interested in binary encoding?
Roberto: yes, it regularises the format, which is valuable
Jeff: SPDY already gets a number of those things, like \r\n processing
Roberto: we get rid of that with anything that we do
Mark: I'm a little concerned about how far we disappear down that rabbit hole
Jeff: key-value compression is going to get us a lot
Martin: Date changes on every request, which makes that a non-gain
James Snell: We can encode dates down to 5-6 bytes. Even if they change every message, that's a significant savings.
Jeff: but Date is so trivially small in comparison to Cookie
Will: James wanted to avoid having state with this
Martin: URIs can be compressed by losing the high bit
Mark: I did that
Jeff: You can't because some clients send UTF-8 text
Roberto: delta is pretty much the same as the simple thing, with 30% extra saving from the huffman coding
Mark: I was surprised at how much could be saved by looking only one message back.
James: Weighing in remotely.. binary encoding offers many benefits at cost of significantly increased complexity... need to be very careful... also, we're going to need to run lots more tests on the delta state management.. very concerned about this in middleboxes
James: Requests are EXTREMELY redundant and wasteful of bits on the wire. Delta does very good with those requests.
James: Looking over header stats, custom unregistered headers tend to take up quite a bit of space. Allowing those to be binary may provide benefits.. on the other hand, base64 and hex compresses very well. So definite tradeoff. I'm not convinced either way completely yet tho.
Jeff: Interested in treating headers as opaque blobs, so that we don't have to look at the contents of the header.
Roberto: looking inside can give us some benefits
Jeff: I don't want to see lots of types, so that the outer layer doesn't have to bother with the content of the headers. Only smaller headers above the compressor, and enumerating types and restricting them, might be. Cookies are better to concentrate on.
Mark: dates were a significant proportion of the compression savings
Roberto: cookies contain significant entropy, so they don't compress well
James: If we allow cookie values to be optionally binary, we can see some savings... but that's a whole can of worms I'm not sure we want to open...
(Martin: James, we did discuss that option and there seemed to be some enthusiasm for it. Keep in mind that binary packing saves only 33% over base64, which is either huge or trivial, depending on perspective.)
James: precisely, right now I'm running tests with binary cookie values and the savings are generally minimal. To realize real savings, we'd likely have to reeducate devs on what "types" of values are best to send... not a good thing imo. Like I said, not sure it's going to be worth it for cookies...
James: However, binary coding the Set-Cookie header itself (not the value) does yield a significant savings (up to 50% in some cases), because of packing the date, changing httponly and secure to single bits, etc
Patrick: cookies aren't a problem everywhere
Jeff: Changing so that we send multiple cookie headers, so that we can compose from the ones that change and the ones that don't, then we can save the space.
Patrick: We can do both.
James: I tend to agree to an extent. It would be beneficial to have some of the data outside of the compressed header block.. path, host, etc. But we cannot predict what other headers may become important later on. So this is hard to optimize.
James: If we take a look at the header stats.. which ones are most redundant, which ones are most variable, and optimize for the extreme cases (dates, cookies, accept-*, p3p, etc) then we make major gains. We don't have to optimize everything.
Roberto: delta can beat gzip on some sites (pinterest)
Jeff: delta alone is great, and we could probably stop there
Mark: you can always just send the text.
Jeff: we shouldn't be spending time on binary coding. it increases complexity, conversions are not hard. it should be simple
James: I agree to an extent. Not convinced we should do binary coding either yet. It does, however, produce significantly fewer bits on the wire, so need to prioritize requirements. Do we want optimum encoding/compression or do we want more simplicity. The potential increased complexity might be too significant to ignore
James: Maybe we should stop at just binary encodings dates and numbers? (date, expires, :status, last-modified, if-modified-since, set-cookie expiration, etc)
Roberto: we can implement the date change and see what that gets us
Mark: date really did move the needle.
Will: does date really move the needle on latency?
Roberto: Date can comprise a significant overhead. Especially when you stop inlining.
Mark: Amazon data moves from 0.47 to 0.39 compression just compressing dates.
Herve Ruellan: there are lots of cache check requests that generate lots of 304 responses, which include lots of dates.
Mark: Maybe we should start with delta and then we can experiment with binary encoding.
Roberto: has a simpler implementation of delta (delta2)
Jeff: I'm going to implement a shim that looks like HTTP/1.1 to its users, simplicity.
James: FWIW, amazon data moves to .20 compression with delta+bohe, which includes dates, cookies, p3p, etc ... that said, delta without bohe is also .20 on requests ... on responses, delta+bohe is .23 and delta is .31
James: Mark, +1.. but we really do need to get a firm grasp on the impact delta has on middleboxes before we commit on that approach. In general (in theory) it looks ok, but there are definite risks.
James: While waiting, just another quick comment re: delta and state management, there definitely is a significant security risk. If a malicious sender purposefully sends junk requests targeted at the maximum storage size specified by the receiver, the receiver could end up being forced to store quite a bit of junk data in memory. Not sure yet how we can deal with this.
headerdiff looks fairly similar to delta in operation - table at decoder or a size the decoder chooses; the encoder sends commands in relation to that table
Name table is pre-populated with common header names; value table keyed by names: as each value appears, three options for the command: add, don't add or replace existing
deflate as optional on the already compressed stream, which makes it faster and more compact than SPDY/3, but no good solutions for the CRIME attack (editorial)
James: I tried many various combinations of stream compressors, every case was subject to CRIME-type attacks. All it required were minor variations in the approach, easily circumvented. Most compelling approach resulted in random compression ratios, but all attacker had to do was average those out over multiple requests. More work, yes, but still feasible
Looks like headerdiff is slightly worse than SPDY/3, without deflate, but adding deflate made it better.
Martin: I like the dictionary of names - from both compression and performance perspective
Mike: the dictionary in SPDY can make a difference initially
Jeff: the dictionary can cause stuff-ups in code
Will: we shouldn't spend a lot of time on this stuff
James: There are many ways we ultimately could end up doing this. bottom line is we need more implementation experience around alternate header encoding and compression. To determine impact on servers, security implications, complexity, and performance, not just ratios. We don't have sufficient data yet to make a solid decision.
Patrick: what do we know about how this resists attacks like CRIME
Roberto: we've had people look at this (AGL) and they are ok with it, you need to guess the entire thing, so if the cookie isn't trivial it's very hard to attack
James: Delta is very good approach re: CRIME ... but carries other risks. We need to determine what the trade offs are. If we stop CRIME, but increase chance of DDOS, did we really help? Consider also that Delta requires state to be kept around in memory for a while. If middlebox is compromised in some other way, an attacker might not have to use CRIME-type attack to get at the data, they could inspect memory and read data from the compression state. More difficult, yes, because it requires direct access to the server, but not impossible. We need to be careful putting cookies and auth headers, etc in stored state
Mark: we'll expect drafts from Roberto and Herve on descriptions of the latest and best algorithms that they can produce and go from there.
Mark: hop-by-hop headers... Not a lot of writing into the header block by intermediaries. Is it OK to drop all the hop-by-hop stuff.
Will: Yes, SPDY doesn't do hop-by-hop.
(Discussion about what happened in SPDY)
Mark: please, can folks changing SPDY make issues in the WG
James: I think dropping them would fine. With SPDY framing, we have other potential options for introducing hop-by-hop stuff later if necessary. But for now, don't see any reason to keep those
Mark: we will drop hop-by-hop headers because all that is moving into the session layer.
Alexey: why?
ISSUE: need a new issue for 100-continue
ISSUE: need a new issue for negotiation of trailers
Roberto: it's not clear when the headers are done
Mark: HTTP/2 currently allows for multiple sets of headers; the receipt of data might be implicit signal for end of headers
Roberto: not in some cases, where data can race headers to the wire
ISSUE: need a clear delineation for the end of the header (or trailer) block
James: I can see a number of compelling theoretical cases where multiple header blocks would be great in applications. However, we should have a distinction between headers in the SYN_STREAM vs. header blocks sent later in the stream
James: Keep headers related to the HTTP Request or Response in the initial block. Allow apps to send other header blocks later, but those serve a separate purpose... trying to come up with a good label for them... But, in theory, I don't think we should rule out sending later header blocks in a stream
Mark: some people want bigger frames
Roberto: smaller, please
Patrick: flow control depends on having gaps to allow pre-emption
Mike: 16k seems small
James: Opt-in for bigger frames? Require receiver to explicitly indicate that they accept large frames as part of flow control mechanism?
Patrick: the argument is a zero-copy argument
Mark: what about a new frame type that allows for pushing massive amounts of bits down the pipe, no flow control, etc...
ISSUE: where do we capture all the important headers (colon headers in SPDY, plus Host) so that they appear first?
Mike: sendfile was added specifically for web servers, let the kernel vendors add some new APIs
James: flow control settings can indicate max frame size or support for large frame type... syn_stream can indicate that sender intends to send large frames?
Mike: you can use smaller frame sizes in practice, even if the maximum size is bigger than what you might normally (or reasonably) use; we did introduce new limits that didn't exist... in theory
Roberto: proposes that control frames have a continuation bit for when you need to send another frame
Mark: I don't want to live in a world that has more than 16Mb of HTTP headers
Jeff: does anyone do more than 8k?
Patrick: 4k. we don't want to get boxed in with a massive frame, I'm more concerned about that than a protocol that stands the test of time
Mark: no one is saying that we need to make it huge
James: just noting... in header stats tests I ran, amazon data, header values accounted for 110,743 bytes across 366 messages... that's uncompressed data.
Mark: what is the optimum size?
Jeff: 16k (for TLS record layer) at which it is too slow
Hasan Khalil: 16k is already too big for TLS
James: Can we just allow receiver to specify a maximum frame size they will accept and go with that? Why should we try to define some arbitrary limit? Allow it to be set as part of the flow control and have the value tweaked over time to reflect requirements
Brian Raymor: there are some issues raised by an MS engineer regarding frame sizes
Roberto: some of these are OK, like settings;
Mike: frame size can affect latency (ed: suspicious); limiting is arbitrary; if we need larger frames later, clamping down might be too early
Mark: we can avoid negotiation by having a goldilocks frame
Mike: same size maximum for control and data (agreement)
Mike: too many times got burned when we placed a limit and discovered that the limit was too low
Roberto: small frames does mean that there is a fair chance that people will have to write continuation code (and test it)
Hasan: we will then need header value continuations for those mega-cookies
Roberto: the continuations would operate at the frame layer, you build all your headers, encode them into a massive buffer, then chunk them up.
James: Header bytes are small in comparison to data... not overly worried about Cookies just yet. Nothing I've seen so far with header data suggests that they would need to be split up among multiple blocks for a while.. at least not in our sample data
Jeff: don't care about size, you can write an implementation that throws stuff away
James: We need to do better at educating developers on more efficient header definitions (look at P3P header as a bad example)... more efficient values, means less wasted space and smaller frames. I'm not overly concerned with control frame sizes tho, those are easily managed.
Mark: a single frame no longer necessarily includes the entire message
ACTION: Roberto to provide a proposal regarding frame size: 16 bit length, +1 bit stolen from flags to indicate continuation.
Flow control principals reviewed; no disagreement.
Stream-level flow control is not controversial.
Session-level flow control is interesting to many (and viewed as necessary to some), but some believe it's not going to be useful, because the response to it being exceeded won't be different (close the connection).
Resolved to spec out a proposal for session-level as well as stream-level flow control.
ACTION: Will Chan to make proposal.
Also some discussion of a new control frame to indicate that flow limit has been reached; optional on both sides. Not much interest, but no disagreement yet. No one offered to make a proposal.
Agreed flow control by a receiver can be disabled by sending a very large / reserved value.
Initial values for flow control - we need some testing.
64k per stream and per-session seems reasonable. Will test and gather feedback.
Patrick: Can we have asymmetric defaults? Big for clients, small for servers.
Will: but client can send settings frame.
Patrick: yes, but defaults for performance.
Martin: Could be part of HTTP/2 magic. Settings are also persisted, but only per origin.
Roberto / Eliot: It may also be interesting to put the default settings frame into DNS.
ACTION: Eliot to scope out a DNS-based "initial options" record (TXT-based?) to avoid wasting a protocol.
- flow control window sizes
- max stream
- pointer to browser hints
Discussion of constant bit rate flow control, e.g., to control use of radio on mobile; felt that this isn't an appropriate use of flow control.
Mark: two axis here how much interleaving Vs waiting to send back response, right now we only have 1 axis
SPDY4 proposal - priority queue, strict between priorities but undefined for requests at same priority.
Eliot: How many queues do you have? Answer: 8
Roberto: tab pri and pri of what is visible is hard/impossible with strict priority. No requirement to finish high pri before sending low pri.
Roberto: rule we execute is send highest pri available thing so don't starve low pri thing waiting to get/generate high pri thing.
Mike: we found same prioritisation approach in browser wasn't optimal for all sites. Room for research here.
Mike: what is use case for strict ordering and why can't be accomplished at app layer?
Roberto: obvious uses cases for priority e.g. switching between tabs
Mark: 3 things - bits for pri, ??? prioritisation, strict priorities
Mike: 3 unique features - priorities, chains, groups
Jeff: if can reorder pri can the dependencies graphs/groups/etc be implemented in browser layer by telling the protocol layer "re prioritise this stream"
Roberto: we discussed that and the implementation gets complex
Mike: tradeoff of complexity, simple approach - 8 priorities. 2 things thinking of adding - dependencies & groups.
Will: one approach is get rid of priorities and just have groups of dependencies
Hasan: a weighted set of ordered dependencies
Mike: what is use case that needs dependencies
Roberto: ???
Martin: What are you proposing we go in protocol?
Roberto: increase size of pri field to size of streamID so can put streamID in there.
Will: and if you are new group of dependency chain that field is used as weight
Mark: priorities is between roots?
Answer: yes
whiteboard discussion of proposal for weights with linked equivalency classes
Mark: 3 things we have here:
- Expanding pri bits
- grouping
- re-prioritisation
Roberto: difficult to split up
Mark: is there something we can put in spec as experiment?
Roberto: can send proposal we have done in SPDY-DEV.
Mark: for me grouping seems attractive, re-pri seems to add tremendous amount of complexity.
Roberto: but we didn't get to talk server push yet as that is a big motivator for it
Mark: Document - complete proposal, use cases and how to use proposal in an experiment?
Will: not happy currently because chrome implementation is buggy so we don't have much data/experience but I think server push has many motivating use cases.
Hasan: server push is something we have already presented & clear on motivating use cases and there is some agreement on the use cases. Now we need to really implement & get some data to figure out how to move forward.
Jeff: we have use cases, reason we haven't rolled out is it is buggy in chrome.
Roberto: everything before SPDY4 will be buggy because server push consumes its own stream. Let's assume no more than 100 concurrent streams. Let's say I want to push 125 resources, I can't. that's stupid because when I'm pushing I'm generally doing it in sequence so max concurrency is 1.
Jeff: for Roberto's use cases want pushes done sequentially and that is broken in spec. parallel pushes works fine.
Mike: that is a pretty small bug, you made it sound like the whole thing is broken.
Mark: I'm concerned because we seem to be slowly adding a whole bunch of new control frames (not just for server push)
Mike: one issue: do we want to reserve streams in advance?
Hasan: reserving request resources.
Jeff: some stuff I missed
Mark: that's clearer don't bother sending that request
Hasan: we would remove Associated-to from SYN_STREAM to PUSH_PROMISE
Jeff: like this idea. Can we take the no-op stream id for it?
Mark: Hasan to write a proposal for PUSH_PROMISE
Mark: I was unsure about push, as were other folks, e.g. web spider may not want to have resources pushed at it, but alternative is often server inlining and client can always squelch the push.
Patrick: some people on client side are worried about it because they pay by the byte.
Hasan: two use cases: 1) client in australia & I don't want any push at all - set max stream limit to 0 for that endpoint. (2) e.g. you're pushing me the yahoo front page image & I already have that cached from a different roaming connection, client gets push_promise and client can cancel that. There may already be things in the socket buffer at that point
Mark: you say 1 RTT but in Australia that can be quite large.
Mark: I was circumspect at the start but you've mostly convinced me. Have google/yahoo guys talked to their web spider guys about this?
Hasan: web spider at google doesn't currently spider over SPDY, secondly it is easy to turn off (send settings frame with zero)
Mark: those sorts of changes need to come across so we can look to incorporate them in the HTTPBIS spec.
ACTION on Will - Double check directionality fo max streams has made it to the HTTP/2 spec.
Mark: I want to optimise further for cacheable content and how to handle if that content is already in cache.
Mark: there are some people who feel web arch is based on resource being the atomic unit of authority, so I might have different resources on server that can't speak authoritatively about each other. There is another whether unit of authority is same origin. You're going for latter model for server push, so we need to fly that around a bit.
Roberto: push into cache, which is temporary cache which only exists for the context of the document.
Mark: I don't think that's spec'd out yet. We need to make sure the security properties are OK and clear. including way it is & is not reused - we need to write that down properly.
Need to Note somewhere (in the spec?): We need some text for security considerations around security model etc. for server push.
discussion on what implementations of server push there are
Conclusion: No one wants to kill server push but some concerns over implementation complexity
Mark: what is minimum bar to get first HTTP2 draft out that folks can implement for experience.
Privacy/settings - missed discussion
ACTION: Mike to make proposal on list
Began with discussion about stablization of the drafts. Need to set expectations that things can still change.
Discussion of SPDYv4 and migrating features into HTTP 2.0 drafts.
Martin made the comment that the base could be used to experiment, unless you're changing frame formats.
Word alignment and sizes need to be adjusted.
Uniform length between control and data frames.
Roberto pointed out that you don't get word alignment with adding type.
Length 16, type/pad 8, flags 8, c/d 1, stream id 31
Magic for the front Martin has picked a random number
Roberto has given Martin a pointer to PUSH PROMISE control frame
Session flow control
How to turn to off push (in either direction)
- Header Compression
- Frame changes
- uniform length / alignment
- remove version
- rearrange/resize - length 16 / type|opcode 8 / flags 8 / C 1 / sid 31
- push promise control frame
- 32bit priority (1st reserved)
- Upgrade
- dance + magic
- npn
- Flow control - session level
- Settings persistence - at risk
- Settings first mandatory
- RST_STREAM / GOWAWAY opaque data?
- 100/continue advice?
Prioritization - just create a 32 bit priority field with MSB reserved.
Rajiv Bector: what sort of changes can we expect beyond this draft?
Mark: there will be more changes
Magic will be sent in all cases, and will include a high order bit set, as well as versioning information.
ACTION: Eliot / Martin to start something on the mailing list. version name to start "http-draft-n/2.0".
Purpose of magic is primarily to fail quickly on middle boxes that improperly respond to upgrade
Mandatory settings in front.
A few weeks to get the proposals shaped up
Decision needed on compression within a month
Late March for implementations
Use github for draft development
Discussion about whether everything on the list of things for the next draft is necessary to test now.
We need to discuss what kind of data we want to collect. What questions do we need to ask?
- Header compression:
- state overhead
- compression efficiency
- split out desktop / mobile
- Use of push
- Latency improvement of push
- Stream utilisation
- Stream concurrency
- Stream max-out
- Variety of settings
- max streams
- flow control - when is sending limited
- error rates
- protocol errors
- unknown control frames
- upgrade errors
- versions
Mark would like a draft or a wiki entry about specific metrics. Will is willing to provide information about what is possible.
Hasan:
- errors specific to http2.0
- reset stream errors
- goaway errors
- is there garbage at the end of the frame?
- improper interpretation of diagnostics that are sent.
Hasan to send a proposal to the list.
What to do about 100/continue?
ACTION: Mark will take this to the list
Clients, servers, and intermediaries all need to be tested.
- adapt wireshark plug-in -- Hasan
- standalone python frame decoder / encoder
- netcat-like tool -- roberto
- catalogue of example sessions and API output (server and client)
- negative testing (e.g., syntax errors, protocol errors)
- stupid push server
Roberto: We need a catalog of example sessions Jeff: we also need error cases
Jeff has some open source java code that we could borrow from that is based on SPDY3. Mark is going to write up some stuff in python.
Intermediaries are hard because you have to simulate clients and servers.
Roberto: would be nice to have a stupid server that just tests PUSH
Discussion of the value of another Interim.
When? September-ish; perhaps June if we can move quickly enough to take advantage of it.
Where? Melbourne/Auckland/London/Toronto/Boston/Berlin/Zurich
- World is very heterogeneous
- Fragmented serving architecture
- All have different software, management, etc. This is a small and big company problem.
- There is very little SSL in many parts of the world.
We're going to need to live with HTTP/1.0/1.1 and 2.0
Operational/administrative/technical reasons drive a separation between base html and CDN content in many sites.
How do we leverage next generation of the web without waiting for people to re-architect for SSL?
Wildcards have problems with 4-level domains; pages having lots and lots of domains.
As a transition mechanism, serve base page over HTTP and then CDN content over SPDY.
Mark: What signalling needs to be available up the stack to indicate which version of HTTP is being used?
People would rather not sniff the UA, because it's getting more difficult
One possibility is to optimise content for SPDY, but then you need a header or something for that.
Roberto: how about a server declaring that a bunch of domains are equivalent.
Will: but you need a bootstrapping mechanism.
Roberto: wildcard dns?
Rajeev: you take a performance hit...
Mark: we need to address this as a transitional problem
Roberto: focus on mixed-mode use case.
Will: Issue is really getting http into the address bar.
One solution: send an HTTP request to a SPDY server and respond with Javascript to rewrite.