Question about deployment strategy

[tuan3w]

Hi team, Thanks for building this framework. We would love to build our product on top of this. But since we're complete new, we would love to hear more about concrete experiences. Our camera AI systems will may have 30-50s cameras. Each will have different config: object detection, object tracking,... What's the common best practices people do when deploy pipelines with Savant:

• Single hybrid pipeline or different pipelines for different config?
• On single machine, is single worker is enough or should we scale it to more than number of workers ?

Additional resources would be valuable. Thanks.

[bwsw]

Hello @tuan3w, we normally design architectures on a paid basis because there is no "one size fits all".

For example, deciding on the number of concurrent instances mostly depends on benchmarking a specific pipeline. If the CPU is a bottleneck and the GPU is underutilized, in such cases, parallel launch can help. In other situations, parallel launch can cause unnecessary resource contention and context switching, leading to a performance decrease.

Extra resources can be found in docs. In addition to documents, you can ask for specific knowledge in the Discord channel.