The teams goal for 2026 is to enable this architecture

Responsibilities of services

Preprocesser

• parse raw sensor data/frames
• validate and annotate with mission_id, device, ingest_timestamp, schema_version, producer_id
• smooth or filter eg low pass
• compute derived metrics (velocity etc)
• publish to mqtt topics
• optionally provide rest endpoint for -
  • /health – system heartbeat
  • /schema – current telemetry schema (JSON Schema)
  • /status – CPU/GPU/memory metrics

MQTT Broker

• real time message bus for telemetry and status
• retain latest message for new subscribers (eg after ui disconnect)
• Topics such as rov/{rov_id}/{mission_id}/telemetry/{sensor}
• Allow multiple pub subs
• Bidirectional communication if necessary (might be for float)

Video ingest service

• capture and encode video stream, stream low latency
• Avoid routing frames through mqtt, publish metadata to mqtt
• jetson nano, encode/decode and stream video using ffmpeg
• allow concurrent subscribers eg UI, recorder
• optionally provide rest endpoint for -
  • /health – system heartbeat
  • /schema – current telemetry schema (JSON Schema)
  • /status – CPU/GPU/memory metrics

Schema registry

• json schema with pydantic models accessed by fast api
• retrieve by version and allow ui to request a schema for decoding rather than hard coupling and redefining in all applications

UI

• queries schema registry for data format based on schema present in packets
• subscribe to cleaned topics via async mqtt clients
• decode video on seperate thread
• connect to video ingest service for low latency video
• show mission critical indicators
• display tasks lists, timers, alerts
• potentially allow limited controls input
• be able to render replay streams
• offload analysis to dashboards

Secondary UI

• queries schema registry for data format based on schema present in packets
• Subscribe to telemetry + ML topics
• Display system diagnostics
• Display graphs
• Show ML inference output, species count, iceberg tracking, edna estimates, could be a web dashboard eg grafana

DB writer

• queries schema registry for data format based on schema present in packets
• consumes telemetry stream/meta data from mqtt topic
• writes to postgres and influx or similar

Replay engine

• runs anywhere
• reconstruct past missions from db
• publish to mqtt
• optionally stream timestamped synced video

ML consumers

• subscribe to video data
• run inference tasks
• publish results back to mqtt
• archived processed annotations/images for training

Archive

• s3/minio store for processed video