@ry reminder about the small feature in the HTTP parser @substack and I asked you about. To support faster proxying we need to expose the total byte length of the HTTP headers parsed.
I'll suggest exposing 'parser.boundary' as an integer. Thoughts?
The proposed boundary integer should also include the framing information for incoming multipart requests (including the final 0\r\n chunk) so that proxies can know when to unpipe themselves so that multiple requests with keep-alive don't bleed into each other.
@substack Agree with you that framing is important, but shouldn't it be exposed though a different var? e.g. parser.frames?
Expose HTTP header boundary
http_parser: expose `parser.boundary`
Requires new `http_parser` (with mmalecki/http-parser@b7136a7 included).
http_parser test: test if parser boundary is exposed
Added some commits. @indexzero @substack can you guys review and make sure this is what you mean? Take a look at the test commit (mmalecki/node@bd06e42) to see how it works.
This is a good start at what needs to happen. The framing boundaries for multipart requests need to be exposed also.
@substack I agree, but we should open a separate issue for that. .boundary should enable most reverse-proxy scenarios.
we need this :)
I'm still +1 on this, but worth mentioning my new approach after speaking with @pgriess. The reason I'm still +1 on this is that the approach using parser.boundary suggests to be more performant because it only invokes a single function on every chunk: ondata instead of both onbody and ondata.
Just a +1 on the general requirement -- I need this for htracr IIRC.
So I'd like to expand a bit on my comment on the joyent/http_parser pull request.
FWIW, I think this is a natural extension of the work that node_http_parser.cc is already doing to avoid extra upcalls by batching headers.
@mikeal @substack This is not blocking any behavior. I thought that it was, but I was wrong:
We should simply buffer data in parser.onBody instead of doing any framing ourselves. I can't speak to the 0\r\n problem that @substack has encountered, but exposing the header boundary would not fix that.
we need to know this:
for a given onBody call, how many actual bytes were parsed?
with just that information we can figure out how many bytes came over the socket from the offset we have for the end of the headers and when onEnd is called we know how much data, if any, the socket has emitted that is part of a new request. i believe that would resolve @substack's issue.
Related issue is mranney/node_pcap#8
I don't think you need to know how many bytes were parsed.
The parser is deigned such that it owns all framing decisions -- you should never be inspecting raw data other than when you know the connection has been upgraded (e.g. via a CONNECT method or an Upgrade: header) and thus you're not dealing with HTTP anymore.
In the scenario you're describing (a single packet contains both the end of the current message and the beginning of a new one), passing the entire buffer into the parser would result in on_body, on_message_complete, on_message_begin, on_header_field, etc ... You can observe the transition from old to new message in the on_message_end/on_message_begin sequence of callbacks. If that doesn't work for you, could you give some more detail on what you're trying to do such that you need to know exactly where at the byte level we transition between messages?
If you want to write a transparent proxy without reframing everything needlessly and just using .pipe() to pump data to a new place, then this boundary information is necessary to compute where to shutdown the pipe to prevent keep-alive connections from bleeding into each other. You really can't do this with the current parser without writing an ad-hoc parser on top of the current parser.
Fair enough; that would require changes in node_pcap (or some really nasty hacks around it), but point taken.
Even if you have header boundary information, you're still faced with figuring out where the end of the message is. This logic is more convoluted than anyone would like: is there a content-length field? is there a transfer-encoding in effect? is this a response? is this eligible for EOF termination? is this the response to a HEAD request? is the status 1xx/204/304? etc. Anyway, all of this is mercifully already implemented in http_parser, but if you don't want to round-trip body data through there, I think you're stuck implementing all this yourself, which it sounds like you don't want to do. Is there another option that I'm missing?
Anyway, if we ignore the boundaries between messages and just focus on the header/body boundary for the time being, you can get at this information by invoking http_parse_pause() in the on_headers_complete callback such that http_parser_execute() stops at the boundary. You can then look at how many bytes it consumed to know where the message body starts.
How is this implemented in other highly performant proxies? Do they deep inspect or just pipe data through (as requested by most in this discussion?
Handling keep-alive in proxies is a bit tricky because, while the proxy may support keep-alive, the upstream server may not.
For both HTTP/1.0 and 1.1, this is a per-hop decision (the Connection: header is per-hop, and is used to indicate keep-alive in HTTP/1.0, which does not implicitly re-use connections like HTTP/1.1 does). As a result, the proxy must be prepared to interrogate the message stream to ensure that it can send a single non-pipelined request to upstream hosts that don't support keep-alive. Even if you knew your upstream host supported keep-alive, it would be bad practice to just pipe the two TCP connections together, again because of the (per-hop) Connection header. The upstream server could send a "Connection: close" header, indicating that it does not wish to re-use the TCP connection with the proxy, but that should not imply that the proxy needs to close its connection with the client.
That said, I'm not sure what the majority of proxies do in practice; @mnot is the real expert there.
I can't help but feel that following the HAProxy implementation pattern would be beneficial beyond trying to re-solve this problem.
defaulting connections to close is not an acceptable solution for performance reasons.
You'll obviously have to implement the proxy in such a manner that upstream limitations are upheld. What I don't understand is why this issue isn't based on the extensive amount of thought put into e.g. HAProxy, rather than being re-explored entirely.
If the rest of you are all intimately familiar with such proxies; please have me excused. Otherwise, that seems like the right thing to spend time digging into at this point.
our goal is not to copy or match the performance of HAProxy.
our goal is to be faster than HAProxy, which will require a fair amount of new thought.
@mikeal; heh, that's a sane goal, but you don't get there by getting stuck on how to parse responses or properly deal with differing upstream servers. Haproxy (and similar) have long since moved beyond that, and onto focusing on optimizing allocators, ebtree based state storage and so on.
This just has the ring of an age-old problem being re-solved.
Closing. Deferring to the work done in nodejs/node#1457