Running a second, isolated stack

You can run two independent OpenProcessor stacks on one host (for example a stable deployment and a development clone). Isolation comes from four things, all set in each clone's own .env.

This is separate from curation projects, which isolate datasets inside one stack. A second stack has its own OpenSearch, so it has its own set of projects. COMPOSE_PROJECT_NAME below names the Compose project, not a curation project.

1. Compose project name

docker-compose.yml sets name: ${COMPOSE_PROJECT_NAME:-openprocessor}, and every container_name is ${COMPOSE_PROJECT_NAME:-openprocessor}-<role>. Give the second clone its own name and it gets its own container names, network (<name>_triton_net) and named volumes automatically:

# second clone's .env
COMPOSE_PROJECT_NAME=opdev

warning

A clone with no COMPOSE_PROJECT_NAME (and no -p) falls back to the project name openprocessor. If the first stack also uses the default, docker compose commands in the second clone act on the first stack's containers. Set the variable, or always pass -p <name>.

2. Host ports

Every published port is interpolated from .env. Remap all of them in the second clone so nothing collides:

API_PORT=4703
TRITON_HTTP_PORT=4700
TRITON_GRPC_PORT=4701
TRITON_METRICS_PORT=4702
OPENSEARCH_PORT=4707
OPENSEARCH_DASHBOARDS_PORT=4708
PROMETHEUS_PORT=4704
GRAFANA_PORT=4705
LOKI_PORT=4706
MLFLOW_PORT=4709
DCGM_PORT=4710
SEGMENTER_PORT=4711
VLM_PORT=4712
OP_CROPWRIGHT_PORT=5284

make reads .env too (-include .env), so make up, make test-detect and friends target the remapped ports without extra flags. The HTTP test scripts read API_PORT, TRITON_HTTP_PORT and OPENSEARCH_PORT from the environment.

3. Image tag

Every built service's image: is ${OP_IMAGE_REPO}/<image>:${OP_IMAGE_TAG:-latest}. If both clones build with the default tag, docker compose build in one retags the :latest the other stack's containers reference, and the next restart of the other stack silently picks up the foreign build. Give each clone its own tag:

OP_IMAGE_TAG=dev

4. GPUs

Pin each GPU-using service of the second stack to the card(s) you want (see GPU sizing):

TRITON_GPU_ID=1
API_GPU_ID=1
EVALUATOR_GPU_ID=1
SEGMENTER_GPU_ID=1
VLM_GPU_ID=1
OP_TRAIN_GPU_ORDER=1

Checklist

cp env.template .env        # in the second clone
# set COMPOSE_PROJECT_NAME, all *_PORT values, OP_IMAGE_TAG, *_GPU_ID
docker compose config | grep -E 'container_name|published'   # sanity check
make up                     # core stack; `make curation-up` adds the workers
curl http://localhost:4703/health

Compose reads COMPOSE_PROJECT_NAME from the clone's .env. From any other directory, pass -p opdev on every docker compose command. The ./openprocessor CLI also reads the name from .env; in a checkout without it the CLI falls back to the project openprocessor, which this checklist avoids by setting COMPOSE_PROJECT_NAME.