I did some testing with the Docker API. The only problem is that a container needs a process running to stay active, if it does not exist the instance stop before receiving the command through the API. There is the possibility to maintain an infinite and inexpensive process like: tail -f / dev / null (not a good idea). Another possibility would be refactor the part of the marvin that would remain in the container to become a daemon or leave an instance of Jupyter running (the easy way). Is there any other possibility that I'm not seeing?
Em sex., 13 de mar. de 2020 às 15:46, Daniel Takabayashi < daniel.takabaya...@gmail.com> escreveu: > Take a look at the Docker API https://docs.docker.com/engine/api/v1.24/, I > believe this will simplify the solution. Cause if you add another layer > (Flask application) you going to need to control/manage this layer as well, > creating a chicken and egg problem. > > I believe the Docker API is a simple solution that solves the problem > (communication between toolbox and containers). > > Take a look here > > https://docs.google.com/drawings/d/shajxIpLJHxxMbFgDXiPuhg/image?w=602&h=461&rev=1423&ac=1&parent=1ySERHGBXbHeyCMRookq5UfTuFkzzU0ugtjvR3rF3deY > > > Em sex., 13 de mar. de 2020 às 11:19, Lucas Cardoso Silva < > cardosolucas61....@gmail.com> escreveu: > > > I thought about using a flask application, which can become very fast and > > scalable with a Python WSGI server for production like Gunicorn. We can > > make the communication establish by HTTPS protocol and use a public and > > private key generation system for each engine generated. Another, > althougt > > complex, alternative would be use OAuth 2.0. > > > > Regards, > > Lucas Cardoso > > > > Em sex., 13 de mar. de 2020 às 13:58, Daniel Takabayashi < > > daniel.takabaya...@gmail.com> escreveu: > > > > > @Daniel The biggest challenge is to run commands in a remote container > > > without having access to the whole O.S. Basically the idea is to create > > an > > > interface to make possible, in a secure way the communication > > > between Toolbox and remote engines. > > > > > > Creating this we could start to run our engines in clouds and services > > like > > > Google Run, Kubernetes, Lambda Functions and etc in the same way Marvin > > > runs locally. > > > > > > @Lucas Lets talk about the details of these APIs (interfaces, > technology) > > > > > > Thanks, > > > Taka > > > > > > Em sex., 13 de mar. de 2020 às 07:21, Daniel Lucredio < > > > daniel.lucre...@ufscar.br> escreveu: > > > > > > > Hi Lucas and everyone, > > > > > > > > Couldn't the developer just run the CLI from inside the container, > > > opening > > > > a shell inside it? > > > > > > > > []s > > > > > > > > Daniel > > > > > > > > Em qui., 5 de mar. de 2020 às 13:31, Lucas Cardoso Silva < > > > > cardosolucas61....@gmail.com> escreveu: > > > > > > > > > I think we could define better how we will make the communication > CLI > > > > with > > > > > the Docker development instance in the new architecture. The > > container > > > > > needs a running process to stay active. We could make a > communication > > > via > > > > > API endpoints to keep a web service running on the container > > receiving > > > > > information from the CLI for the execution of tasks in Marvin. I > > > created > > > > a > > > > > use-case scenario to facilitate understanding and further > discussion. > > > > > > > > > > Scenario: A developer created an engine on marvin through the CLI, > he > > > > > configures his engine in order to describe all the dependencies of > > the > > > > > operating system that will be used during development. After the > > engine > > > > > configuration process, the developer uses the CLI to upload a > > > development > > > > > environment on a Docker container, that development environment > will > > > > > contain the refined Marvin toolbox. The CLI then changes its > > interface > > > in > > > > > order to contain the standard commands for using engines. He uses > the > > > CLI > > > > > to communicate with the toolbox endpoints and inform that he wants > to > > > run > > > > > an instance of the notebook. After building the model, the > developer > > > can, > > > > > through the CLI, perform a dryrun and upload the http-server for > > > testing. > > > > > All of these procedures will be done using the communication > through > > > API > > > > > endpoints. > > > > > > > > > > Does that sound like a good strategy? Do you have any suggestions, > or > > > > > something that was already foreseen in the original project? > > > > > > > > > > Best regards, > > > > > Lucas Cardoso > > > > > > > > > > > > > > >
