This talk introduces COMET, a DSL and compiler framework for dense and sparse tensor algebra that employs progressive lowering to generate efficient code for target heterogeneous systems starting from several high-level languages. COMET has been designed to solve some of the challenges in scientific, data analytics, and AI workloads, commonly used at National Laboratories. In addition to common compiler optimizations and code transformations, COMET employs domain-specific and architecture-specific optimizations leveraging the semantics expressed by high-level languages and architectural features. This talk specifically focuses on high-performance data analytics, highlighting challenges and opportunities for domain-specific optimizations at the compiler level and describing the optimizations and code transformation employed by COMET to generate efficient code for graph algorithms.