Currently we are doing a rough load balancing system to improve redundancy and to run some internal tests. When a file is requested you are assigned to a random IPFS node (this distributes the requests across all nodes added to the network) but if the file does not start playing within 10 seconds it will switch to a different node (it assumes that the node being used is either too busy to serve the file at a satisfactory speed or else it is offline for some unforseen reason).

This is not something we plan to do permanently but rather something that allows our site to track node uptime and balance out requests across all nodes added. Moving forward we will be rolling out a much more comprehensive load balancing system which will cache 'known' nodes to forward you to the best node for the file you're requesting to make sure it is served quickly and reliably. Consider the current settings to be a debug mode for us to gather data.

Nodes in IPFS act much like a tracker does with torrents. When you request a hash from a node, it checks for parts of that file within peers (users and other nodes connected to it) and then serves it to you. Pushing all users to a single node would NOT be decentralized and is what we want to avoid. Instead we are running some of our own nodes, adding nodes run by our users and contacting existing ipfs nodes to ask if they are willing to join in our site. So far the response has been very good but we are still focusing on the underlying stability of the system rather than pushing for it to function at optimal speeds. Better to work kinks out now than when the site has grown and has thousands or millions of people trying to view items.