The Internet of Things (IoT) enables connections of trillions of sensors and data collection for connectivity and analytics. The amount of IoT-generated data has exploded due to the rapid growth of interconnected IoT devices from a wide range of IoT applications. IoT systems need an effective solution to ingest, store, and analyze the exploding scale of IoT data to make the best of the limited resources in IoT systems, such as network, storage, energy, and computation power. Given that IoT data is endless, heterogeneous, and dynamic, the thesis proposes new compression techniques to handle those special data features. MOP handles the endlessness of IoT data by pre-allocating encoding space for the dynamic incoming data and enables in-situ queries in the encoded domain for fast query execution. BUFF addresses the heterogeneity of IoT data by applying decomposed but compact encoding space for the IoT numeric data with different statistics and achieves efficient query execution support according to the host's hardware. In addition, IoT's dynamic data imposes challenges to traditional compression selection strategies as there is no one-size-fits-all solution for IoT systems with varying data statistics, complex workloads, and constrained hardware resources. We introduce AdaEdge as a hardware-conscious encoding selection framework for resource-constrained devices.