The present invention extends to methods, systems, and computer program products for accessing and manipulating data in a data flow graph. A data flow graph includes one more transforms for transforming data from one or more data stores into different formats accessible to clients. Transforms can be configured to join and split data from the same or different data stores. A metadata tree logically represents the data flow graph including locations that provide output from transforms of the data flow graph. The locations can be referenced by clients to access transformed outputs. Clients can register dependencies on data at various locations and can request evaluation of the data in response to data changes at a corresponding data store. Clients can also request data changes be propagated back through the data flow graph to one or more data stores and implemented at the one or more data stores.