Node Agent Integration

[mmoreiradj]

Node Agent Integration

gRPC Worklod Service Server Integration

Add a wrapper for containerd-client lib (and ctr cli for creation) to create a container and run a workload.

The wrapper uses gRPC to communicate to the containerd socket. It reconnects to the socket before each request. This is done to avoid potential issues with resource management, isolation, and security and simplicity. It is also more fault-tolerant.

This will however mean more resource usage, overhead, latency, potential socket exhaustion, etc… This is a tradeoff we are willing to make for now, since these issues are not likely to be a problem in the near future.

The ctr cli was used to create the container because the containerd-client we did not manage to create a container with the containerd-client lib. It runs the command ctr run -d <image> <container-id>.

The containerd-client lib was also used to kill the container. We did however have an issue concerning an ambiguous type efinition. Indeed, the Workload Signal enum is ambiguous. It is not clear what signal STOP represents. For now, we made the assumption that stop is SIGINT (2). Both Stop (SIGINT) and Kill (SIGKILL) signals are sent to the workload, then the workload is cleaned up.

Note: It seems @Kuruyia has managed to prepare a rootfs and run a container, you can find it here.

But to be honest, going with runc or a container runtime made in rust might be an easier way to achieve our goals.

gRPC Lifecycle Service Client and Status Update Service Client Integration

When the program starts, it connects to the scheduler and registers itself to it. If it fails, it will retry every 5 seconds. If it fails 3 times (can be modified in args or env, see node-agent/src/args.rs), it will exit. It also starts the gRPC server.

If any of the two processes fail (gRPC of lifecycle), the server will shut down.

Note: We would be glad to know how you think we managed the application's lifecycle. It was a first for us, and it feels clunky.

Cargo Workspaces

We added a Cargo.toml file at the root of the project to manage the dependencies of the different crates. Building, formatting and linting will now be easier.

However, we had an issue when building the node-agent. When you build the binary from the root of the project, a request to the containerd socket fails, saying it didn't find it. We did not manage to fix this issue. We did however manage to make it work only if build from the node-agent directory. We are not sure why this is happening.

Misc

Add ide directory to gitignore.

What's next

Use the network team's CNI implementation.

@sameo we split the commits for convenience. Note that two thirds of the lines we added is a lock file.

@noetarbouriech

[mmoreiradj]

Rebased.