-
-
Save igrigorik/3058839 to your computer and use it in GitHub Desktop.
| <!DOCTYPE html> | |
| <html> | |
| <head> | |
| <meta charset=utf-8 /> | |
| <title>Hello</title> | |
| <link rel="stylesheet" href="styles.css"> | |
| </head> | |
| <body> | |
| Hello World | |
| </body> | |
| </html> |
| require 'goliath' | |
| class Delay < Goliath::API | |
| def response(env) | |
| case env['REQUEST_PATH'] | |
| when /css$/ then | |
| # simulate slow (1s delay) stylesheet | |
| EM::Synchrony.sleep(1) | |
| return [200, {'Content-Type' => 'text/css'}, 'body { color: red }'] | |
| else | |
| data = File.open('file.html') | |
| # flush the HTML head (140 bytes) immediately after the headers | |
| # this will allow the preloader to request the stylesheet before | |
| # the rest of the page completes | |
| EM.add_timer(0.01) do | |
| env.chunked_stream_send(data.read(140)) | |
| end | |
| # flush remainder of document after 1.5s | |
| EM.add_timer(1.5) do | |
| env.chunked_stream_send(data.read.to_s) | |
| env.chunked_stream_close | |
| end | |
| # flush 200 response immediately | |
| chunked_streaming_response(200, {'Content-Type' => 'text/html'}) | |
| end | |
| end | |
| end |
Yeah, browsers process nodes on the fly and fire requests for assets as their respective nodes are parsed. I did some research on this back when streaming was added to Rails. Wrote a post here http://weblog.rubyonrails.org/2011/4/18/why-http-streaming. And also have a talk with much more information in Spanish (http://vimeo.com/37688380).
BTW that behavior does not depend on whether the response is streamed, nodes are process on the fly no matter how the bytes arrive to the browser, streamed or not. The point, as you observe, is that if some get earlier, that fires parallel requests so the overall responsiveness is better.
Of course with standard cache practices for assets, the benefit of that is mostly targeted at a first request.
@fxn: Yup. Quick question: did automatic flushing (http://yehudakatz.com/2010/09/07/automatic-flushing-the-rails-3-1-plan/) ever make it into Rails 3.1?
Yes, it is automatic. The layout gets chunked, and the template main content goes in one single chunk. Basically, each time you switch from hard-coded content to dynamic content in ERB you get a chunk. In a template you can still use a kind of content_for helper called provide (the layout communicates with the template via fibers in this case).
Server-side support is also interesting. You need a 1.1 proxy, Rails disables streaming if the client sends a 1.0 request because chunked responses are a 1.1 feature. Also, you obviously do not want your reverse proxy to buffer the response as nginx does by default, at the price of... well not having the response rapidly consumed.
Then, the reverse proxy should be able to compress, since you want compression in a production environment. In order to compress, the reverse proxy must dechunk, compress, and chunk again, on the fly.
It is a very interesting topic with many practical gotchas in production environments. When I researched, from the setups I tested this the only combo that worked was Apache + Unicorn, a combo that is indeed not recommended by Unicorn. By "worked" I mean it didn't buffer and was able to compress on the fly.
Hi there, in terms of the Jekyll site, wondering if is there an equivalent way to do the flush after the header tag as well, something like Flush the Buffer Early. Thank you.
Chrome is able to fetch the stylesheet before the document itself completed, resulting in ~1.5s DCL, instead of a 2.5s DCL if we waited for the document to complete first.
Discussion: https://plus.google.com/114552443805676710515/posts/GTWYbYWP6xP