http://bittorrent.org/beps/bep_0005.html#routing-table
Every node maintains a routing table of known good nodes. The nodes in the routing table are used as starting points for queries in the DHT. Nodes from the routing table are returned in response to queries from other nodes.
Not all nodes that we learn about are equal. Some are "good" and some are not. Many nodes using the DHT are able to send queries and receive responses, but are not able to respond to queries from other nodes. It is important that each node's routing table must contain only known good nodes. A good node is a node has responded to one of our queries within the last 15 minutes. A node is also good if it has ever responded to one of our queries and has sent us a query within the last 15 minutes. After 15 minutes of inactivity, a node becomes questionable. Nodes become bad when they fail to respond to multiple queries in a row. Nodes that we know are good are given priority over nodes with unknown status.
The routing table covers the entire node ID space from 0 to 2160. The routing table is subdivided into "buckets" that each cover a portion of the space. An empty table has one bucket with an ID space range of min=0, max=2160. When a node with ID "N" is inserted into the table, it is placed within the bucket that has min <= N < max. An empty table has only one bucket so any node must fit within it. Each bucket can only hold K nodes, currently eight, before becoming "full." When a bucket is full of known good nodes, no more nodes may be added unless our own node ID falls within the range of the bucket. In that case, the bucket is replaced by two new buckets each with half the range of the old bucket and the nodes from the old bucket are distributed among the two new ones. For a new table with only one bucket, the full bucket is always split into two new buckets covering the ranges 0..2159 and 2159..2160.
When the bucket is full of good nodes, the new node is simply discarded. If any nodes in the bucket are known to have become bad, then one is replaced by the new node. If there are any questionable nodes in the bucket have not been seen in the last 15 minutes, the least recently seen node is pinged. If the pinged node responds then the next least recently seen questionable node is pinged until one fails to respond or all of the nodes in the bucket are known to be good. If a node in the bucket fails to respond to a ping, it is suggested to try once more before discarding the node and replacing it with a new good node. In this way, the table fills with stable long running nodes.
Each bucket should maintain a "last changed" property to indicate how "fresh" the contents are. When a node in a bucket is pinged and it responds, or a node is added to a bucket, or a node in a bucket is replaced with another node, the bucket's last changed property should be updated. Buckets that have not been changed in 15 minutes should be "refreshed." This is done by picking a random ID in the range of the bucket and performing a find_nodes search on it. Nodes that are able to receive queries from other nodes usually do not need to refresh buckets often. Nodes that are not able to receive queries from other nodes usually will need to refresh all buckets periodically to ensure there are good nodes in their table when the DHT is needed.
Upon inserting the first node into its routing table and when starting up thereafter, the node should attempt to find the closest nodes in the DHT to itself. It does this by issuing find_node messages to closer and closer nodes until it cannot find any closer. The routing table should be saved between invocations of the client software.