An approach for restoring an analytical model in a data streaming application which involves creating a plurality of data buffers for holding data submitted to the data streaming application, wherein the data streaming application comprises a first analytical model, persisting one or more parameters and/or internal state variables associated with the first analytical model, persisting the plurality of data buffers, retrieving a plurality of persisted data buffers and the one or more parameters and/or internal state variables responsive to a failure of the data streaming application, creating a second analytical model associated with the data streaming application and initializing the second analytical model with the one or more parameters and/or more internal state variables, and submitting data, associated with the plurality of persisted data buffers, to the data streaming application for training the second analytical model to be a close approximation of the first analytical model.