Real-time data analysis applications increasingly rely on complex streaming computations over time-series data. We propose StreamQL, a language that facilitates the high-level specification of complex analyses over streaming time series. StreamQL is built upon stream transformations, which allows it to integrate three language-based approaches for data stream processing: relational queries, dataflow composition, and temporal formalisms. The relational constructs are useful for specifying simple transformations, aggregations, and the partitioning of data into key-based groups or windows. The dataflow abstractions enable the modular description of a computation as a pipeline of stages or, more generally, as a directed graph of independent tasks. Finally, temporal constructs can be used to specify complex temporal patterns and time-varying computations. These three constructs can be composed freely to describe complex streaming computations.

StreamQL has a formal denotational semantics, which we use to prove the expressive completeness of the language. We provide an implementation of StreamQL as a lightweight Java library, which we use to experimentally evaluate our approach. The experiments show that the throughput of our implementation is consistently higher than that of state-of-the-art streaming engines.