3 ms·

As others have indicated, it's a proxy server rather than a general purpose webserver. There's no code relation between the two servers; it's simply that the te

by skuhn 11y ago

As others have indicated, it's a proxy server rather than a general purpose webserver. There's no code relation between the two servers; it's simply that the team at Yahoo chose to pursue it as an Apache Foundation project when they open sourced it. ATS scales orders of magnitude better than Apache, due to its process model. Whereas at Yahoo we would budget between 30-200 simultaneous connections per Apache server (prefork), the proxy service which I ran using ATS was budgeted for over 100,000 concurrent connections per machine. It's significantly less featureful than Apache, but it does caching substantially better than any other cache server commonly available (nginx, apache, squid, varnish).

exelius 11y ago
Yeah; I find many people in the modern web services world end up using Apache HTTPD as a caching proxy for availability reasons. It's probably just because it's old as hell so everyone knows the config file format by heart, but I know I went with Apache HTTPD on a recent project because the features we needed were only in the paid version of Nginx and we just didn't want to deal with licensing (the dollar amount was trivial in corporate dollars, but it would have taken us months to get the purchase through procurement and the entire rest of the project used open source software). So it's good to have another open source option to keep in mind.
- therein 11y ago
  What were these features?
  - exelius 11y ago
    Application health checks mostly; though we were using it in a caching server context (static files and non-whitelisted endpoints served directly from Apache vs. hitting the app servers). They're available in Nginx Plus, just not the free version. Again, the pricing wasn't the issue; it was the fact we would have had to go through procurement - which involves a few weeks/months of process at any decently large company. So we ended up using Apache because the team was familiar with it and knew it could support our use case. While Nginx probably would have performed better, in the end it was just easier to use Apache because it was a mature Open Source project and throw a few extra VMs at the caching tier to make up for the performance gap.
jimjag 11y ago
"Apache, due to its process model."... "(prefork)" Seems to me that the above is all based on reflections from decades ago w/ Apache httpd 1.3. Right now, all web servers can handle similar levels of concurrency with the bottleneck being the network pipe itself. ATS is a great platform; using it in combo w/ Apache httpd (2.4) allows a pure open source implementation with all the power, speed, reliability one could want, and protection against Open Core business models.
- skuhn 11y ago
  Uh well, it wasn't "decades ago" but it was some old timey stuff. Yahoo used Apache 1.3 as recently as 2012. They also disabled Keep-Alive on Apache, and most properties would use the hardcoded default number of prefork processes (32). It wasn't the smartest setup. Nevertheless, I don't think that nginx / Apache / Varnish / haproxy / etc. are able to handle similar concurrent connection levels as ATS without significantly impacting 95th percentile latency due to their core architectures.
  - jimjag 11y ago
    There's good info here: http://www.slideshare.net/bryan_call/choosing-a-proxy-server-apachecon-2014
    - skuhn 11y ago
      Agreed, it's a good presentation and does a good job of being fair to the various software considered. (I worked on a team closely related to Bryan's team at Yahoo)
    - jjm 11y ago
      Hey, nice seeing you here! Glad to see the project reach HN (finally). That indeed is a very good slide deck. Some features that I preach are: - good turnkey default values - lua support - config options galore (Bryan labels it a con, but if you want control it's perfect) - good logging - historically proven scalability on large smp, xxlarge memory, multi nic systems Edit: forgot to add one more thing though maybe not worthy a bullet point. If possible, a preference for physical rather than virtual is where I've seen performance with ATS shine. That is one reason why you would want as much config control possible.

by skuhn 11y ago

exelius 11y ago

Yeah; I find many people in the modern web services world end up using Apache HTTPD as a caching proxy for availability reasons. It's probably just because it's old as hell so everyone knows the config file format by heart, but I know I went with Apache HTTPD on a recent project because the features we needed were only in the paid version of Nginx and we just didn't want to deal with licensing (the dollar amount was trivial in corporate dollars, but it would have taken us months to get the purchase through procurement and the entire rest of the project used open source software). So it's good to have another open source option to keep in mind.

therein 11y ago
What were these features?
- exelius 11y ago
  Application health checks mostly; though we were using it in a caching server context (static files and non-whitelisted endpoints served directly from Apache vs. hitting the app servers). They're available in Nginx Plus, just not the free version. Again, the pricing wasn't the issue; it was the fact we would have had to go through procurement - which involves a few weeks/months of process at any decently large company. So we ended up using Apache because the team was familiar with it and knew it could support our use case. While Nginx probably would have performed better, in the end it was just easier to use Apache because it was a mature Open Source project and throw a few extra VMs at the caching tier to make up for the performance gap.

jimjag 11y ago

"Apache, due to its process model."... "(prefork)" Seems to me that the above is all based on reflections from decades ago w/ Apache httpd 1.3. Right now, all web servers can handle similar levels of concurrency with the bottleneck being the network pipe itself. ATS is a great platform; using it in combo w/ Apache httpd (2.4) allows a pure open source implementation with all the power, speed, reliability one could want, and protection against Open Core business models.

skuhn 11y ago
Uh well, it wasn't "decades ago" but it was some old timey stuff. Yahoo used Apache 1.3 as recently as 2012. They also disabled Keep-Alive on Apache, and most properties would use the hardcoded default number of prefork processes (32). It wasn't the smartest setup. Nevertheless, I don't think that nginx / Apache / Varnish / haproxy / etc. are able to handle similar concurrent connection levels as ATS without significantly impacting 95th percentile latency due to their core architectures.
- jimjag 11y ago
  There's good info here: http://www.slideshare.net/bryan_call/choosing-a-proxy-server-apachecon-2014
  - skuhn 11y ago
    Agreed, it's a good presentation and does a good job of being fair to the various software considered. (I worked on a team closely related to Bryan's team at Yahoo)
  - jjm 11y ago
    Hey, nice seeing you here! Glad to see the project reach HN (finally). That indeed is a very good slide deck. Some features that I preach are: - good turnkey default values - lua support - config options galore (Bryan labels it a con, but if you want control it's perfect) - good logging - historically proven scalability on large smp, xxlarge memory, multi nic systems Edit: forgot to add one more thing though maybe not worthy a bullet point. If possible, a preference for physical rather than virtual is where I've seen performance with ATS shine. That is one reason why you would want as much config control possible.