Skip to main content

Running MAC Without the Package

If you are evaluating MAC from a source checkout, you can run the CLI directly from the build tree.

From a Source Checkout​

~$ make cli
~$ ./cli/mac env set MAC_CONFIG_DIR=$PWD/deploy
~$ ./cli/mac start all --local     # builds images from ./services, then runs them

The same env / pull / build / start / stop / status / clean subcommands all work; only the binary path and MAC_CONFIG_DIR differ. The --local flag (or MAC_IMAGE_SOURCE=local) makes start build the images from the checkout instead of pulling from the registry; make build / mac build do the same as a separate step.

If your $HOME is NFS-shared across nodes, export MAC_CONFIG_DIR in the shell instead of mac env set so the checkout path applies only to the current node.

Direct docker compose Usage​

The CLI is a convenience wrapper around docker compose plus the agent registration handshake. If you prefer to drive Compose yourself:

~$ cd /etc/mango/mac          # or your $MAC_CONFIG_DIR
~$ docker compose -f docker-compose.master.yml up -d
~$ docker compose --profile nvidia -f docker-compose.agent.yml up -d   # or --profile amd

The GPU vendor exporters (dcgm-exporter on NVIDIA, rocm-exporter on AMD) are gated behind Compose profiles:, so a bare docker compose -f docker-compose.agent.yml up -d brings up everything except the GPU exporter. Pass --profile nvidia or --profile amd to start it — this is the flag mac start supplies automatically from the detected GPU profile.

Note that bypassing mac start agent also skips the auto-registration step — the agent will only appear on the master once it is successfully scraped by Prometheus.

When This Matters​

  • Reproducing a customer issue from a specific git ref
  • Hot-patching the api-server or web-ui in a lab
  • Running MAC images on Kubernetes via your own manifest