As suggested by Godfrey (off-list), I'm moving this discussion to
Stackoverflow:
http://stackoverflow.com/questions/17797753/why-is-rails-middleware-overhead-so-high-when-nginx-is-used-as-a-proxy
Thanks for all input so far.
Best,
Rodrigo.
Em 22-07-2013 17:14, Rodrigo Rosenfeld Rosas escreveu:
Hi Godfrey, thank you very much for your very helpful input! See more
comments inline below.
Em 22-07-2013 13:55, Godfrey Chan escreveu:
Hi Rodrigo,
For the future, I suspect this question is more appropriate on
Stackoverflow/Quora than rubyonrails-core, but here is my quick take
on this:
On Monday, July 22, 2013 7:40:56 AM UTC-7, Rodrigo Rosenfeld Rosas
wrote:
I'm tuning the performance of some request in my application and
there's
something puzzling me.
Rails reports the action takes 6ms:
Started GET "/templates/2.json" for 127.0.0.1 at 2013-07-22
10:49:15 -0300
Processing by TemplatesController#index as JSON
Parameters: {"id"=>"2"}
Rendered text template (0.1ms)
Completed 200 OK in 6ms (Views: 0.8ms | Sequel: 2.2ms)
Chrome's "waiting" time for a local request (localhost) is 25ms.
Then nginx will report 98ms for the same request when proxying
the Rails
app, but that is a separate issue I'll investigate by other means...
First, I'd like to understand what is this 6ms reported time by
Rails.
Does it take into account the whole request time since it reaches
the
web server or only since it reaches the Rails stack? Does it take
into
account all middlewares? Any hint where those extra 19ms from
Chrome's
reported "waiting" time are being spent on?
The response time reported by Rails is based on a "process_action"
event in ActionController
<https://github.com/rails/rails/blob/master/actionpack/lib/action_controller/metal/instrumentation.rb#L18>,
so it's pretty low on the stack on only includes time spent in your
controllers. There is a Rack::Runtime middleware (included by Rails
by default) that wraps around your entire middleware stack (well,
almost) if you prefer a more accurate measurement. The middleware
appends a X-Runtime response header, which you can probably see from
the Chrome inspector panel. The difference between this value and the
one in your Rails log is probably pretty small though (a few ms) as a
request typically spends the majority of its time in the controller.
Thank you for pointing out to X-Runtime (I've seen it before but I was
unaware of its exact meaning). You're right, the reported Chrome's
timing when I'm accessing the Rails application directly (no nginx
proxy) is very close to the information in X-Runtime:
Chrome "Waiting" time: 27ms
X-Runtime: 0.025152
Reported time in the log: Completed 200 OK in 7ms (Views: 0.9ms |
Sequel: 3.0ms)
So, now I know that the overhead for ActionController::Base is about
18ms and I could try building some actions on top of Metal if I want
to improve it further, but at least now I know where to look at.
Nginx/apache/passenger etc might add 1ms or so to this but in the
grand scheme of things that's pretty much just noise.
You're right, when comparing the X-Runtime with Chrome's "Waiting"
time, the nginx overhead seems to be pretty small (4ms at most), but
still Rails will take longer to process the request when nginx is
involved and I have no clue why is that. Here's what happens when
nginx is involved:
Chrome "Waiting" time: 88ms
X-Runtime: 0.083885
Completed 200 OK in 7ms (Views: 0.8ms | Sequel: 2.9ms)
This is what I'd really like to improve since there's no reason why
Rails should take longer to respond to the request when it is
requested by nginx instead of directly from the browser...
I'm using Puma for development mode (where I'm issuing those tests)
and unicorn under production...
I'd appreciate if you could provide me any hints on why the proxied
request is slower from the middleware perspective. Maybe some settings
I could try on nginx. Of course, I won't be insisting on that on this
mailing list since this is non-related to Rails, but in case you know
the answer I'd be glad to hear your guesses.
The "waiting" time in Chrome is roughly equivalent to the delta of
"responseStart" and "responseEnd" in the navigation timing spec
<https://dvcs.w3.org/hg/webperf/raw-file/tip/specs/NavigationTiming/Overview.html#processing-model>,
which means the time between immediately before browser sent the
request to your sever and when the browser receives the first byte of
the response header. This would include the time it takes to send the
request across the wire (watch your cookie size because they are sent
on every request!) and the time your server takes to generate a response.
The moral of the story is that if you care about the real world
performance of your app, you should probably looking into some kind
of RUM (real user monitoring)
<http://en.wikipedia.org/wiki/Real_user_monitoring> solution in
additional to your sever-side monitoring to get an accurate picture.
If you already using Newrelic you should enable it here
<https://newrelic.com/docs/features/real-user-monitoring>. There are
also other standalone services for doing this. (Shameless plug – if
you're building single page apps, we are working on this at
caliper.io <http://caliper.io>)
I wasn't aware of Caliper before. We're indeed maintaining an SPA and
I'll try to find some time to have a closer look at what Caliper is
and how it could help us. Currently we're indeed monitoring our
application with NewRelic and have RUM enabled for them (we have a set
of applications running actually, Grails, Rails, Solr and a Spring MVC
one). But I already have some hints of which requests I should be
optimizing (almost all JSON requests exclusively) and I'm also
converting some requests handling from the Grails app to the Rails one
and I'm particularly interested in making sure the requests are as
fast or faster when implemented in the Rails side (which has been true
so far, as the ported actions are much faster in the Rails side).
Also, I'd like to take the chance to ask about another thing I've
noticed recently. If I remember correctly, Rails used to fill in the
Etag automatically in every request based on the response content
but
now it will only do that if you use something like fresh_when or
stale?.
Is that correct or has it always been working the way it
currently does?
I understand that the overhead may not be desired, but I'd be
curious if
the automatic Etag header has been removed, what would be the
reason for
that besides the overhead?
The "automatic" etag is generated by the Rack::ETag middleware and I
don't remember seeing this removed from Rails 4. To be sure, you can
always run rake middleware and check if you see something like "use
Rack::ETag" in the list.
Humm... Maybe I've seen then in the past for HTML requests and have
just assumed it worked for JSON format as well? I've confirmed the
middleware is enabled for my application. Maybe Rails will skip JSON
requests since jQuery won't take advantage of Not Modified responses
by default. The ifModified ajax parameter defaults to false and if you
enable it you'll have to take care of caching it in the client-side
yourself, so maybe that was enough reason for disabling the automatic
ETag caching for JSON requests... I've been handling this in jQuery
with an ajax filter:
$.ajaxSetup({ cache: false, traditional: true, ifModified: true }); //
traditional is required by Grails default binding rules
var _ajaxCache = {};
$.ajaxPrefilter('json', function(options, originalOptions, jqXHR) {
options.success = function(data, textStatus, xhr) {
if (xhr.status != 304) {
if (xhr.getResponseHeader('Last-Modified'))
_mhAjaxCache[originalOptions.url] = data;
return originalOptions.success.apply(this, arguments);
}
var args = Array.prototype.slice.call(arguments);
args[0] = _ajaxCache[originalOptions.url];
return originalOptions.success.apply(this, args);
}
});
Thank you so much, Godfrey!
Cheers,
Rodrigo.
--
You received this message because you are subscribed to the Google Groups "Ruby on
Rails: Core" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To post to this group, send email to [email protected].
Visit this group at http://groups.google.com/group/rubyonrails-core.
For more options, visit https://groups.google.com/groups/opt_out.
