rails/mongrel cluster - keeping media in sync across nodes with a CSS/LB

rubyonrails-talk
1

Hi all,

I figured I would see what other people would do and get some ideas on what route I want to take. Let me first give you a brief summary of my network and application.

I have a Cisco PIX FW and a Cisco 11501 CSS/LB, there are 5 network nodes behind them. Three are webservers running apache/mongrel and 2 are mysql servers ( master/slave replication )

The three webservers have limited disk space.

I need a way to keep my data in sync across all nodes. A user would upload artwork and other media after logging in, so since the data is private I can't create a VIP in the LB like http://mydomain.com/images and route it to a specific node.

So essientially, I am trying to figure out the best way to keep my media uploads in sync across all nodes. Right now I am using a mounted NFS mount to a master server on each client, but I am afraid of the bottleneck.

I can't use a master upload server and Rsync hence the disk space issue.

I am open to any suggestions that you guys have here, I don't want to have my media uploads "split-brained". Any thoughts on this would be appreciated.

Regards, -- Mike

2

We use Amazon S3 for storage, which works well across several
servers. It's also instantaneous if you're on EC2 (but it's still
limited beta).

Duane Johnson (canadaduane)

3

You only have 3 options: database, shared file system, replication. If you are not able to increase the disk space on the web servers then replication is out. If you are afraid of a bottleneck then shared files are out. And storing large files in a database is not any better than shared files. So I guess nothing works.

Really, the answer is to look at the requirements. How big are the files, how many hits, etc. In most cases replicated files is the best solution because you only take a hit on upload, not access. But, that requires you add disk. How much will more disk cost? Not much. If you have low enough volume then do not worry about NFS or database blobs being the bottleneck. If you have enough volume then move the performance hit to the upload side, not the access side.

Michael