Re: [OxLUG] Web platforms

Top Page
Attachments:
Message as email
+ (text/plain)
+ (text/html)
Delete this message
Reply to this message
Author: Mark Robson
Date:  
To: Oxfordshire Linux User Group Discussion List
Subject: Re: [OxLUG] Web platforms
2010/1/11 Sad Clouds <>

> Hello, I'd like to address people who run Internet sites/portals, or work
> for
> ISPs or hosting companies and administer busy web servers.
>


We run a portal which is a very critical application for a lot of users.
It's not an especially busy site though.

So I'm trying to find out whether people are having any problems with their
> current software stack (Apache, Zeus, etc). What features are lacking and
> what
> they would like to be implemented.
>


No, what happens is, almost nobody (in my experience) has significant
scaling problems with the front-end tier of their web services.

While some servers are slightly more efficient than others, scaling can
still be done by adding hardware. Fancy-pants web servers (lighthttpd, nginx
etc) are all very well, but Apache is well tested and works.

All of the focus of performance and reliability work needs to be at the
middle / back end. This is where all the world of pain begins.

Constructing HTML web pages can be expensive in terms of CPU, but it scales
very, very well. You can simply add hardware until you're blue in the face
(using appropriate load balancers and/or round-robin DNS in front).

On the other hand, anything involving big data and high availability (which
is pretty much crucial, because stuff WILL FAIL) is where the problems lie.

Web front end servers can be state-less (or hold only a small amount of very
unimportant state such as sessions and a little bit of logs which you can
afford to lose). So web servers can be highly available too. When they fail
or need to be shut down, just take them out of the load balancer pool. Or
even just leave the damn things in there, if web browsers make a connection
and they get connection-refused, they just retry again (usually on another
IP if you have RR DNS). The end users won't even notice.

However, things which need to hold on to state, cannot scale as easily as
they need to share data and cooperate in order to provide some level of
consistency.

I feel absolutely that you're barking up the wrong tree.

The World does not need another "super efficient" web server which can serve
static HTML a tiny teeny bit faster. Nobody cares enough. That's not the
bottleneck of performance in real applications, scalability of the web tier
is a solved problem, everyone has moved on.

Mark