Embedded & IoT Vocabulary

Bare metal programming means running code directly on the microcontroller without any operating system — no RTOS, no Linux, no scheduler. The firmware is the only software running on the hardware. Bare metal vs RTOS: Bare metal — you write your own main loop and interrupt handlers; full control, lowest latency, no OS overhead. Suitable for simple devices with deterministic timing. RTOS (Real-Time Operating System) — provides task scheduling, timing, IPC (semaphores, mutexes, queues). Examples: FreeRTOS, Zephyr, RTEMS. Suitable for complex firmware with multiple concurrent tasks. Key embedded vocabulary: HAL (Hardware Abstraction Layer) — a library (e.g., STM32 HAL) that wraps register access in functions. BSP (Board Support Package) — hardware-specific code for a particular board. Peripheral — on-chip hardware modules (GPIO, UART, SPI, I2C, timers, ADC). Register — memory-mapped hardware control locations. Bootloader — code that runs at power-on before the main firmware, often handling firmware updates. In conversation: "We kept it bare metal — the device only does one thing and the latency requirements are under 10 microseconds."

MQTT (Message Queuing Telemetry Transport) is a publish-subscribe protocol designed for constrained environments: low bandwidth, high latency, or unreliable networks. Originally developed by IBM for monitoring oil pipelines via satellite. MQTT vocabulary: Broker — the central server that receives and routes messages (Mosquitto, EMQX, HiveMQ, AWS IoT). Publisher — a device that sends messages to topics. Subscriber — a client that receives messages on topics it subscribed to. Topic — a hierarchical string identifier (e.g., "sensors/floor2/temperature"). QoS (Quality of Service) — 0 = at most once (fire and forget), 1 = at least once, 2 = exactly once. Retain — the broker stores the last message on a topic for new subscribers. LWT (Last Will and Testament) — a pre-configured message the broker sends if a device disconnects unexpectedly. Why MQTT over HTTP for IoT: smaller packet overhead (2-byte header vs HTTP headers), persistent connections, push model vs polling, QoS guarantees. CoAP is another IoT protocol (UDP-based, even lighter). In conversation: "We use MQTT QoS 1 for sensor data — some duplicates are acceptable, but we can't afford to lose readings in the database."

volatile in C/C++ instructs the compiler: "do not optimise, cache, or reorder accesses to this variable — always read it from memory." Without volatile, the compiler may cache the variable in a register and never re-read it from memory, missing updates made by an ISR. ISR (Interrupt Service Routine) is a function that runs asynchronously when a hardware interrupt fires — timer overflow, GPIO edge, UART receive complete, etc. ISR best practices: keep ISRs extremely short (set a flag, copy data to a buffer); never do heavy computation, blocking calls, or dynamic memory allocation in an ISR; variables shared between ISR and main code must be declared volatile. Atomic access: on 32-bit systems reading a 32-bit variable is atomic; reading a 64-bit variable in two instructions is not — an ISR could fire between the two reads. Solution: disable interrupts around multi-byte reads, or use atomic types. In conversation: "The bug was a race condition — the main loop read the ring buffer head pointer between the ISR's two writes, seeing a partially-updated index."

Edge computing is processing data at or near the source of generation — on the device, gateway, or a local server — rather than sending everything to the cloud. Benefits: Low latency — local decisions (e.g., stopping a machine) don't depend on round-trip to the cloud. Bandwidth reduction — only relevant/aggregated data is sent to cloud. Offline resilience — the system works even when cloud connectivity is lost. Privacy — sensitive data can be processed locally without leaving the premises. Edge computing vocabulary: Edge device — a sensor, camera, or controller at the data source. Edge gateway — a more powerful device that aggregates and pre-processes data from multiple edge devices. Fog computing — Cisco's term for a hierarchy of edge nodes between devices and cloud. Inference at the edge — running ML models locally (e.g., on-device image recognition). In contrast, cloud computing centralises processing. The architecture: Device (sensor) → Gateway (edge) → Cloud (long-term storage, analytics). In conversation: "We run ONNX models on the gateway for real-time defect detection; only anomalies are uploaded to the cloud for human review."

I2C (Inter-Integrated Circuit) is a synchronous, 2-wire serial protocol: SDA (data) and SCL (clock). A single master can communicate with multiple slaves on the same bus, each identified by a unique 7-bit address. I2C characteristics: Speed — Standard (100 kHz), Fast (400 kHz), Fast-Plus (1 MHz). Multi-master — multiple masters possible but requires arbitration. Open-drain — both lines require pull-up resistors. Common I2C devices: temperature sensors (LM75), IMUs (MPU-6050), OLEDs (SSD1306), EEPROMs, RTC chips. Comparison with other embedded protocols: SPI (Serial Peripheral Interface) — 4-wire, faster (MHz), full-duplex, hardware select per device. Better for high-speed peripherals (displays, SD cards). UART — asynchronous, 2-wire, point-to-point only. Used for GPS, Bluetooth modules, debug output. CAN bus — used in automotive and industrial, robust noise immunity, multi-master, long cables. Note: "master/slave" terminology is being replaced in some communities with "controller/responder" or "host/peripheral." In conversation: "We had I2C address conflicts — two sensors had the same default address, so we had to pull the ADDR pin low on one to change its address to 0x49."