Smart systems are increasingly used in a wide range of sectors, addressing environmental, societal, and economic challenges such as limited resources, climate change, aging, and globalization. The Internet of Things (IoT) technology is rapidly spreading in these systems, and the MQTT protocol has become a dominant lightweight publish-subscribe mechanism for them to provide reliable and low-overhead device communication. At the core of MQTT-based systems, message brokers act as intermediaries between IoT devices and application platforms. However, as connection density and message throughput increase, traditional broker deployments often fail to efficiently utilize available hardware resources. In particular, widely used implementations such as Mosquitto exhibit inherent scalability limitations due to single-core connection handling and static deployment models. This paper introduces a controller-orchestrated scaling architecture for the Mosquitto broker to enable dynamic, standards-compliant scalability without requiring client reconnection. The design incorporates a centralized control plane alongside custom broker plugins to facilitate vertical scaling by increasing multi-broker container density on a single host, and horizontal scaling, through elastic distribution of brokers across multiple hosts. A key feature is the decoupling of message ingestion from delivery via a shared state layer, which supports coordinated fan-out across brokers, while fully preserving MQTT semantics, including session persistence and QoS guarantees. We evaluated the framework using a custom benchmarking tool tailored for high-concurrency scenarios with strict Service Level Agreement (SLA) enforcement. Our dataset analysis across six experimental configurations (SP10, SP100, SP1000, CP10, CP100, CP1000) reveals that the controller-orchestrated architecture improves connection success rate by up to 55% at the scalability ceiling of standalone brokers, sustains message delivery ratios of 99.8-99.9% under stress conditions, where single-broker delivery drops to 64%, achieves near-linear throughput scaling up to 11.89 MB/s across 2,500 concurrent subscribers, and maintains stable P95 latency within 400 ms even as subscriber load triples beyond the single-broker failure threshold. These findings position the proposed approach as a practical solution for large-scale MQTT-based IoT telemetry deployments, effectively bridging the divide between basic single-broker configurations and intricate, manually managed clusters. The architecture implements broker-local QoS semantics suitable for sensor data-based monitoring use cases, where operational simplicity is prioritized over distributed exactly-once guarantees.