How Visylix achieves large-scale concurrent streams through microservices, GPU-accelerated transcoding, distributed storage, and load balancing.
Most VMS platforms were designed in an era when 100 cameras was considered a large deployment. Today, smart city projects span tens of thousands of cameras, and global enterprises manage camera networks across hundreds of locations. Supporting estates of that size requires fundamental architectural decisions that cannot be bolted onto legacy monoliths.
Visylix was built cloud-native from day one using a microservices architecture where each function, stream ingestion, AI inference, recording, playback, and user management, runs as an independently scalable service. This allows each component to scale horizontally based on actual demand rather than provisioning for peak load across the entire system.
The ingestion layer handles protocol negotiation (RTSP, RTMP, SRT, ONVIF) and routes streams to processing pipelines. Each ingestion node handles thousands of concurrent streams using async I/O and an optimized memory architecture for buffer passing to minimize CPU overhead.
Distribution uses a tiered SFU (Selective Forwarding Unit) architecture. Origin servers receive one copy of each stream and relay to edge servers positioned close to viewers. This origin-edge topology reduces backbone bandwidth by 90% compared to direct server-to-viewer delivery for high-fanout streams.
Video transcoding and AI inference are the most compute-intensive operations. Visylix uses GPU hardware encoding for transcoding and dedicated hardware-accelerated inference engines for AI model execution. AI inference is GPU-accelerated where an accelerator is present and falls back to CPU where it is not, and how many streams a given GPU carries depends on which models you enable and at what frame rate.
The scheduler dynamically allocates GPU resources between transcoding and inference based on demand. During business hours when more viewers are actively monitoring, transcoding gets priority. During off-hours, GPU cycles shift to batch analytics and forensic search indexing.
At one million streams, even modest retention policies generate petabytes of data. Visylix uses tiered storage: hot storage (NVMe SSDs) for the most recent 24-72 hours of footage, warm storage (HDD arrays or S3-compatible object stores) for 30-90 day retention, and cold archival (glacier-class storage) for compliance-mandated long-term retention.
Intelligent retention policies reduce storage costs by 50-70% by recording at full resolution only when events are detected. Idle cameras store low-resolution keyframes, with full-quality recording triggered automatically by AI detections or manual operator activation.
For mission-critical surveillance, downtime is unacceptable. Visylix offers a contractual 99.9% uptime SLA on Enterprise, built on primary-standby clustering across availability zones with a documented operator-run cutover runbook (target RTO under 60 seconds, RPO under 5 seconds) and continuous data replication with point-in-time restore. Fully automatic failover is on the roadmap and is not something we claim today.
Every component is designed for graceful degradation. If the AI inference cluster goes down, streams keep recording and displaying. Storage node failure and node cutover are handled through the documented DR runbook rather than silently, so an operator always knows a failover happened.
Visylix uses a microservices architecture where ingestion, AI inference, recording, playback, and user management each scale independently. Each ingestion node handles thousands of concurrent streams, and a tiered origin-edge SFU topology reduces backbone bandwidth by 90 percent compared to direct server-to-viewer delivery for high-fanout streams.
It depends on which of the 22 models you enable, at what frame rate, and on the GPU itself, so we size it on your hardware during the proof of concept rather than publishing a single number. The scheduler dynamically splits GPU cycles between transcoding and inference based on demand, prioritizing transcoding during active viewing hours and shifting GPU capacity to batch analytics and forensic indexing during off-hours.
Visylix uses tiered storage: NVMe SSD hot tier for the last 24 to 72 hours, HDD or S3-compatible object storage for 30 to 90 day retention, and glacier-class cold archive for long-term compliance. Intelligent retention records full resolution only when AI detects events, which typically cuts storage cost by 50 to 70 percent.
Enterprise deployments carry a contractual 99.9 percent uptime SLA. Underneath it sits primary-standby clustering across availability zones, a documented operator-run cutover runbook with target RTO under 60 seconds and RPO under 5 seconds, and continuous replication with point-in-time restore. Components degrade gracefully: if AI inference goes down, streams keep recording and displaying. Fully automatic failover is on the roadmap, not a capability we claim today.