PROBABILISTIC HTTP REQUEST ROUTING

Some embodiments perform probabilistic request routing in addition to or instead of deterministic request routing. The probabilistic request routing is based on probabilistic models that predict the type of content being requested based on commonality in elements between different requests directed to the same type. The probabilistic models accurately route requests that have not been previously encountered and accurately route requests for content whose type is not previously known. The requests are routed across different subsets of servers that are optimized or configured for the predicted type. The probabilistic models can be defined using a decision tree. Machine learning generates and maintains the decision tree. Accuracy predicted by the different branches of the tree is updated through tracking the type of content passed in response to different routed requests. The tree structure is modified based on timestamps associated with the tree elements and based on newly encountered request elements.