I am not sure if this post can pass the algorithm of HN this time. Nonetheless, we are eager to learn feedback about the way we build Instill VDP for its no-code and low-code functionalities. Any comment is very appreciated.
One can absolutely lock versions in their Dockerfile. I can see that the design principle of DevBox is to pin the versions. At the end of the day we all need to consider versioning (i.e., the image version) the versions (i.e., package versions) anyway.
6 months after we posted the manifesto (http://go.instill.tech/4bcxuf), we're releasing an Alpha version of VDP under the open-source Apache license 2.0.
VDP is the future for unstructured data ETL, where developers won't need to build their own data connectors, high-maintenance model serving platform or ELT pipeline automation tool.
VDP supports both real-time and on-demand inference.
Blazingly fast speed or super cost efficiency? It’s on your call.
Real-time inference speed bundles with the genuine Vision AI model performance. You can get the fastest inference result ever with VDP from a model serving’s point of view. Thanks to the integrated Triton Inference Server and the high-performant Go backends.
On-demand inference performs batch operation. You can get the most economic inference cost for non-time-critical vision tasks. Schedule your inference tasks and access the structured data results in your data warehouse later.
Self-hosted Vault within a minimum Kubernetes cluster in GCP costs us roughly $35 a month. Maintenance effort can be neglected if not scaling. Vault has its learning curve there but I think it's totally worth it, given its secret management and API-first features integrated with many other DevOps tools.