Summary

Adds a complete local development workflow for model-engine so developers can iterate on both control plane code and the full endpoint lifecycle without cloud credentials or prod images.

Control-plane-only mode (make dev-server):

• Spins up Postgres + Redis via docker-compose
• LOCAL=true activates fake queue/docker/k8s implementations (mirrors CIRCLECI=true)
• Full gateway API available at :5000 with auth skipped — no k8s cluster needed

Full end-to-end mode (make dev-server-full + make dev-service-builder + make dev-k8s-cacher):

• make kind-up + make kind-image creates a local kind cluster and loads model-engine:local into it
• Service Builder picks up endpoint creation tasks from local Redis and creates real k8s Deployments in kind
• K8s Cacher polls kind and writes endpoint status back to Redis
• Echo server (model-engine:local) used as the inference container — no GPU required

Code fixes included:

• service_builder/celery.py + celery_task_queue_gateway.py: onprem cloud provider now uses redis Celery backend instead of s3 — without this, the Service Builder writes results to Redis but the Gateway looks in S3, leaving endpoints stuck in PENDING
• dependencies.py: LOCAL=true + cloud_provider=onprem falls through to real OnPremQueueEndpointResourceDelegate instead of the fake
• env_vars.py: GIT_TAG defaults to "local" when LOCAL=true so k8s templates reference the correct model-engine:local image

New files:

• docker-compose.local.yml — Postgres 15 + Redis 7 with healthchecks and persistent volume
• service_configs/service_config_local.yaml — HMI config for local services
• model_engine_server/core/configs/local-full.yaml — onprem infra config for kind
• Makefile — all dev targets in one place

Test plan

• make dev-up && make dev-migrate && make dev-server — gateway starts, GET /v1/model-endpoints returns 200
• make kind-up && make kind-image — kind cluster created, model-engine:local loaded
• make dev-server-full + make dev-service-builder + make dev-k8s-cacher — all three processes start cleanly
• POST a sync CPU endpoint with the echo server image → pod appears in kubectl --context kind-llm-engine get pods -n model-engine and endpoint transitions to READY
• Existing unit tests pass: make test

Closes MLI-6681

🤖 Generated with Claude Code

Greptile Summary

• Adds make dev-server (control-plane only, fake k8s/queue) and make dev-server-full / make dev-service-builder / make dev-k8s-cacher (full kind-based end-to-end) workflows; backing services are Postgres 15 + Redis 7 via a new docker-compose.local.yml.
• Fixes a backend/broker mismatch for onprem: both celery_task_queue_gateway.py and service_builder/celery.py now correctly use redis as the Celery result backend for cloud_provider == \"onprem\", and dependencies.py routes LOCAL=true to fake queue delegates for non-onprem configs while still using the real OnPremQueueEndpointResourceDelegate for the full local flow.
• OTLPMetricExporter is imported inside the shared SDK availability guard in correlation.py; this silently expands the OTel requirement to include opentelemetry-exporter-otlp-proto-grpc, which is only listed in vllm-specific requirements and not in the main requirements.txt.

Important Files Changed

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    subgraph Control-plane-only ["Control-plane-only (make dev-server)"]
        A[LOCAL=true\ncloud_provider=aws] --> B[FakeQueueDelegate]
        A --> C[Redis TaskQueueGateway\nlocalhost:6379]
        A --> D[FakeDockerRepository]
    end

    subgraph Full["Full end-to-end (make dev-server-full + builder + cacher)"]
        E[LOCAL=true\ncloud_provider=onprem] --> F[OnPremQueueDelegate]
        E --> G[Redis TaskQueueGateway\nlocalhost:6379]
        G -->|Celery task| H[Service Builder\nredis broker + redis backend]
        H -->|k8s Deployment| I[kind cluster\nmodel-engine:local]
        I -->|status| J[K8s Cacher\nwrites to Redis]
        J --> G
    end

    subgraph Infra
        K[(Postgres\nlocalhost:5432)]
        L[(Redis\nlocalhost:6379)]
    end

    C --- L
    G --- L
    A --- K
    E --- K

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
model-engine/model_engine_server/common/startup_tracing/correlation.py:15-17
**OTLP exporter import tightens OTel availability requirement silently**

`OTLPMetricExporter` is placed in the same `try/except ImportError` block as the core SDK availability check. This means any environment that has `opentelemetry-api` + `opentelemetry-sdk` installed but NOT `opentelemetry-exporter-otlp-proto-grpc` will now get `OTEL_AVAILABLE = False` and all trace correlation will silently be skipped. The exporter is only listed in vllm-specific requirements (`inference/vllm/requirements.txt`), not in the main `requirements.txt`, making this a fragile dependency for a shared utility. The import should be in its own nested `try/except` or removed entirely if `OTLPMetricExporter` isn't actually instantiated in this file.

_{Reviews (10): Last reviewed commit: "Merge branch 'main' into lilyz-ai/mli-66..." | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.