Sunday, July 29, 2007

Planning for high traffic

The most worrying project on my list at this time is working out how to achieve as much grunt as possible to withstand the estimated traffic from the first Steve Irwin day tribute website since his passing.

Currently the Zoo has a single primary server hosting all sites and a secondary web server sharing the load on a few sites. It's not the perfect model by far and I intend on tidying it up as follows.

What I plan to do is install MySQL 5.x on the secondary server and ready it for replication. Then I will dump the contents from the existing MySQL 4.x databases and point all the websites at it. Once I'm satisfied that it's functioning as expected (I'm in the process of testing this on the bench). Then I will upgrade the MySQL 4.x to 5.x on the primary server and begin multi-master replication with the other. This completes stage one of the preparations.

With the databases in place and functioning, I will work out how to replicate all vhosts between the two servers. Plesk, the web control panel operating on both servers, makes this more complex than it really needs to be since I will also need to create the client/domain accounts for each of the replicated sites. When I have worked out what is required I will script the synchronization as much as possible and hopefully end up with near 100% automation. I may need to tap into Plesk's API to do this. This will complete stage two.

With replication of both the databases and general structures taking place between the two existing servers I can now introduce more servers and the greater complexity they will bring.

I will be working towards four application servers just for static content/scripts and a single localised database server. I intend to just have raw boxes running either Redhat or FreeBSD, no fancy control panels getting in the way. This will allow me to script everything with simplicity and provide a basic configuration to each server. I will replicate the data from the first of the existing servers to one of the new application servers and from there to each of the remaining three. This is so I can keep the amount of public traffic between the servers to a minimum. I will introduce the new database server into the multi-master replication loop and point the four new application servers at it. This completes stage three.

Once I am satisfied that each application server is connecting to the database server over their private network and that the database server is successfully replicating the databases from the existing two servers I will set up the load balancer to include the four new application servers. We may need to cut over to a new load balancer since the existing one may not support this many servers.

This setup will provide me with six front end servers and three database servers - with a bit of sharing of resources here and there. The following diagram shows what I intend on achieving.

Relevant links:
ONLamp Advanced MySQL Replication Techniques
MySQL 5.0 Manual - Replication
SWSoft Plesk - Upgrading MySQL