Deployments API

Deployments serve trained models as real-time prediction endpoints. The MLOps lifecycle supports staging and production stages, promotion, rollback, inference logging, and data drift detection.

All deployment endpoints are prefixed with /api/mlops.

Create Deployment 

POST /api/mlops/projects/{project_id}/deployments

Deploy a model to either staging or production. If deploying to production, any existing production deployment in the same project is automatically archived.

Path Parameters

Parameter	Type	Description
`project_id`	UUID	Project that owns the model.

Request Body

Field	Type	Required	Description
`model_id`	string	Yes	UUID of the model to deploy. Must belong to the same project.
`name`	string	Yes	Deployment name.
`stage`	string	Yes	`staging` or `production`.
`traffic_percent`	integer	No	Traffic allocation percentage (1–100, default 100).
`privacy_policy_id`	string	No	Attach a privacy policy for runtime anonymization.
`privacy_anonymize_logs`	boolean	No	Anonymize PII in inference logs.
`privacy_anonymize_response`	boolean	No	Anonymize PII in prediction responses.
`privacy_column_map`	object	No	Map input column names to PII types.
`privacy_threshold`	float	No	Confidence threshold for PII detection (0.0–1.0).

Example

curl -X POST "$BASE_URL/api/mlops/projects/d4e5f6a7-b8c9-0123-def4-567890123456/deployments" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "a7b8c9d0-e1f2-3456-7890-abcdef123456",
    "name": "fraud-detector-staging",
    "stage": "staging",
    "traffic_percent": 100
  }'

import requests

resp = requests.post(
    f"{BASE_URL}/api/mlops/projects/d4e5f6a7-b8c9-0123-def4-567890123456/deployments",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model_id": "a7b8c9d0-e1f2-3456-7890-abcdef123456",
        "name": "fraud-detector-staging",
        "stage": "staging",
        "traffic_percent": 100,
    },
)
deployment_id = resp.json()["deployment_id"]
print("Deployed:", deployment_id)

Response 201 Created

{
  "id": "d0e1f2a3-b4c5-6789-0123-def456789012",
  "deployment_id": "d0e1f2a3-b4c5-6789-0123-def456789012",
  "stage": "staging",
  "is_active": true
}

List Deployments 

GET /api/mlops/projects/{project_id}/deployments

Return all deployments for a project.

Query Parameters

Parameter	Type	Default	Description
`limit`	integer	50	Max items (1–200).
`offset`	integer	0	Pagination offset.

Example

curl "$BASE_URL/api/mlops/projects/d4e5f6a7-b8c9-0123-def4-567890123456/deployments" \
  -H "Authorization: Bearer YOUR_API_KEY"

resp = requests.get(
    f"{BASE_URL}/api/mlops/projects/d4e5f6a7-b8c9-0123-def4-567890123456/deployments",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
)
for dep in resp.json()["items"]:
    print(f"{dep['id']}  stage={dep['stage']}  active={dep['is_active']}")

Response 200 OK

{
  "items": [
    {
      "id": "d0e1f2a3-b4c5-6789-0123-def456789012",
      "model_id": "a7b8c9d0-e1f2-3456-7890-abcdef123456",
      "stage": "staging",
      "is_active": true,
      "config": {"name": "fraud-detector-staging", "traffic_percent": 100},
      "created_at": "2026-02-25T09:00:00Z"
    }
  ],
  "total": 1,
  "limit": 50,
  "offset": 0
}

Get Deployment Detail 

GET /api/mlops/deployments/{deployment_id}

Return deployment metadata and the associated model details.

Example

curl "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012" \
  -H "Authorization: Bearer YOUR_API_KEY"

Response 200 OK

{
  "deployment": {
    "id": "d0e1f2a3-b4c5-6789-0123-def456789012",
    "project_id": "d4e5f6a7-b8c9-0123-def4-567890123456",
    "model_id": "a7b8c9d0-e1f2-3456-7890-abcdef123456",
    "stage": "staging",
    "is_active": true,
    "config": {"name": "fraud-detector-staging", "traffic_percent": 100},
    "created_at": "2026-02-25T09:00:00Z"
  },
  "model": {
    "id": "a7b8c9d0-e1f2-3456-7890-abcdef123456",
    "name": "GBM_1_AutoML",
    "algo": "GBM"
  }
}

Predict via Deployment 

POST /api/mlops/deployments/{deployment_id}/predict

Send input features and receive predictions from the deployed model. This is the primary real-time prediction endpoint. Inference requests are logged automatically for monitoring and drift detection.

Request Body

Field	Type	Required	Description
`inputs`	dict or list[dict]	Yes	Feature values. Single record or batch.
`options`	object	No	Options (see table below).

Predict Options

Key	Type	Description
`include_contributions`	boolean	Include SHAP-style feature contributions (not available for StackedEnsemble models). Default: `false`.
`store_payload`	boolean	Explicitly opt in to storing input and response payloads in inference logs. Default: `false` for data minimization.
`privacy_policy_id`	string	Override the deployment-level privacy policy for this request.
`anonymize_logs`	boolean	Override log anonymization for this request.
`anonymize_response`	boolean	Override response anonymization for this request.

Example

curl -X POST "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/predict" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "inputs": {"amount": 29.99, "merchant_id": 42, "hour": 14},
    "options": {"include_contributions": true, "store_payload": false}
  }'

resp = requests.post(
    f"{BASE_URL}/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/predict",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "inputs": {"amount": 29.99, "merchant_id": 42, "hour": 14},
        "options": {"include_contributions": True, "store_payload": False},
    },
)
result = resp.json()
print("Prediction:", result["predictions"][0]["prediction"])
print("Positive-class risk:", result["predictions"][0]["probability"])
print("Decision threshold:", result["predictions"][0]["decision_threshold"])

Response 200 OK

{
  "predictions": [
    {
      "prediction": "0",
      "probabilities": {"p0": 0.97, "p1": 0.03},
      "probability": 0.03,
      "decision_threshold": 0.08,
      "positive_class": "1",
      "predicted_positive": false
    }
  ],
  "decision_policy": {
    "version": 1,
    "threshold": 0.08,
    "positive_class": "1",
    "source": "model_operating_point"
  },
  "contributions": [
    [
      {"feature": "amount", "value": -0.32},
      {"feature": "merchant_id", "value": -0.15},
      {"feature": "BiasTerm", "value": -1.20}
    ]
  ]
}

probability is the score for positive_class; predicted_positive is the server-side decision obtained from the persisted deployment threshold. A consumer should display both score and threshold instead of treating the class label as an unexplained fraud verdict.

Latency and status are available from inference logs rather than duplicated in this response. The HTTP response also includes Server-Timing with total prediction duration and X-CorePlexML-Model-Cache (hit, miss or not-applicable). H2O deployments use a bounded, process-owned model cache; the JVM is kept warm and only cache-owned models are evicted.

Promote Deployment 

POST /api/mlops/deployments/{deployment_id}/promote

Promote a deployment to production. Any existing production deployment in the same project is automatically archived.

Request Body

Field	Type	Required	Description
`to_stage`	string	Yes	Must be `production`.

Example

curl -X POST "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/promote" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"to_stage": "production"}'

requests.post(
    f"{BASE_URL}/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/promote",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={"to_stage": "production"},
)

Response 200 OK

{
  "ok": true,
  "deployment_id": "d0e1f2a3-b4c5-6789-0123-def456789012",
  "stage": "production"
}

Rollback Deployment 

POST /api/mlops/deployments/{deployment_id}/rollback

Rollback production to a previous deployment or model. Provide either to_deployment_id (reactivate an archived deployment) or to_model_id (create a new production deployment for the given model).

Request Body

Field	Type	Required	Description
`to_deployment_id`	string	No*	UUID of the archived deployment to reactivate.
`to_model_id`	string	No*	UUID of the model to deploy as the new production version.

* At least one of to_deployment_id or to_model_id is required.

Example

curl -X POST "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/rollback" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"to_model_id": "a7b8c9d0-e1f2-3456-7890-abcdef123456"}'

Response 200 OK

{
  "ok": true
}

Deactivate Deployment 

POST /api/mlops/deployments/{deployment_id}/deactivate

Deactivate a deployment without promoting a replacement. The deployment moves to archived stage.

Example

curl -X POST "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/deactivate" \
  -H "Authorization: Bearer YOUR_API_KEY"

Response 200 OK

{
  "ok": true
}

Drift Metrics 

GET /api/mlops/deployments/{deployment_id}/drift

Return drift detection results comparing the training distribution to recent inference data. Returns per-feature PSI scores and overall drift status.

Example

curl "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/drift" \
  -H "Authorization: Bearer YOUR_API_KEY"

resp = requests.get(
    f"{BASE_URL}/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/drift",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
)
drift = resp.json()
for feature, info in drift.get("features", {}).items():
    print(f"{feature}: PSI={info['psi']:.4f} ({info['status']})")

Response 200 OK

{
  "deployment_id": "d0e1f2a3-b4c5-6789-0123-def456789012",
  "overall_status": "low",
  "features": {
    "amount": {"psi": 0.023, "status": "low"},
    "merchant_id": {"psi": 0.15, "status": "moderate"},
    "hour": {"psi": 0.008, "status": "low"}
  },
  "computed_at": "2026-02-28T12:00:00Z",
  "inference_rows": 5000
}

Run Drift Detection 

POST /api/mlops/deployments/{deployment_id}/drift/run

Enqueue a background job to compute drift metrics against recent inference data. The result can be retrieved via the drift metrics endpoint once the job completes.

Example

curl -X POST "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/drift/run" \
  -H "Authorization: Bearer YOUR_API_KEY"

Response 200 OK

{
  "ok": true,
  "job_id": "e1f2a3b4-c5d6-7890-1234-ef5678901234"
}

Inference Logs 

GET /api/mlops/deployments/{deployment_id}/inference-logs

Return a paginated list of inference requests and responses for the deployment. Useful for monitoring, auditing, and debugging.

Query Parameters

Parameter	Type	Default	Description
`limit`	integer	50	Max items (1–500).
`offset`	integer	0	Pagination offset.
`since`	string	–	ISO 8601 timestamp. Only return logs after this time.
`until`	string	–	ISO 8601 timestamp. Only return logs before this time.

Example

curl "$BASE_URL/api/mlops/deployments/d0e1f2a3-b4c5-6789-0123-def456789012/inference-logs?limit=5" \
  -H "Authorization: Bearer YOUR_API_KEY"

Response 200 OK

{
  "items": [
    {
      "id": "f2a3b4c5-d6e7-8901-2345-f67890123456",
      "deployment_id": "d0e1f2a3-b4c5-6789-0123-def456789012",
      "inputs": {"amount": 29.99, "merchant_id": 42, "hour": 14},
      "prediction": "0",
      "latency_ms": 45,
      "status_code": 200,
      "created_at": "2026-02-28T14:30:00Z"
    }
  ],
  "total": 5000,
  "limit": 5,
  "offset": 0
}

Deployment Stages

Stage	Description
`staging`	Pre-production testing. Use this to validate model behavior before going live.
`production`	Live serving. Predictions from this deployment are production traffic.

Deployment Status Values

Status	Description
`active`	Deployment is running and serving predictions.
`inactive`	Deployment has been deactivated. Not serving predictions.
`failed`	Deployment failed to initialize or load the model.

Alert Metrics Reference

Metric	Description
`drift_psi`	Population Stability Index. Measures feature distribution drift between training and production data.
`accuracy_degradation`	Degradation in prediction accuracy over time.
`error_rate`	Percentage of predictions returning errors.
`latency_p99`	99th percentile prediction latency in milliseconds.
`latency_p95`	95th percentile prediction latency in milliseconds.
`request_volume_1h`	Inference requests observed during the last hour.
`worker_job_failures_1h`	Failed background jobs for the deployment project during the last hour.
`model_staleness`	Days since the model was last retrained.
`prediction_anomaly`	Anomalous prediction patterns detected.

Alert Condition Operators

Operator	Symbol	Description
`gt`	>	Greater than threshold.
`lt`	<	Less than threshold.
`gte`	>=	Greater than or equal to threshold.
`lte`	<=	Less than or equal to threshold.
`eq`	==	Equal to threshold.
`neq`	!=	Not equal to threshold.

Alert Severity Levels

Severity	Description
`info`	Informational alert. No action required.
`warning`	Warning. Should be investigated.
`critical`	Critical issue requiring immediate attention.

Notification Channels

Channel	Description
`email`	Send alert to one or more email addresses.
`slack`	Post alert to a Slack channel via webhook.
`webhook`	Send alert payload to a custom HTTP endpoint.

Alert Rule Options

Option	Type	Description
`cooldown_minutes`	integer	Minimum minutes between repeated alerts (1–1440, default 60).
`escalation_minutes`	integer	Minutes before escalating severity (optional, 1–1440).
`escalation_severity`	string	Severity to escalate to: `warning` or `critical`.

Deployment Lifecycle Example 

The following shows the full lifecycle: create a staging deployment, test it, promote to production, and rollback if needed.

import requests

BASE_URL = "http://localhost:8888"
HEADERS = {"Authorization": "Bearer YOUR_API_KEY"}
PROJECT_ID = "d4e5f6a7-b8c9-0123-def4-567890123456"
MODEL_ID = "a7b8c9d0-e1f2-3456-7890-abcdef123456"

# 1. Deploy to staging
resp = requests.post(
    f"{BASE_URL}/api/mlops/projects/{PROJECT_ID}/deployments",
    headers=HEADERS,
    json={"model_id": MODEL_ID, "name": "v2-staging", "stage": "staging"},
)
deployment_id = resp.json()["deployment_id"]

# 2. Test with a sample prediction
resp = requests.post(
    f"{BASE_URL}/api/mlops/deployments/{deployment_id}/predict",
    headers=HEADERS,
    json={"inputs": {"amount": 50.0, "merchant_id": 10, "hour": 12}},
)
assert resp.status_code == 200
print("Staging test passed:", resp.json()["predictions"])

# 3. Promote to production
requests.post(
    f"{BASE_URL}/api/mlops/deployments/{deployment_id}/promote",
    headers=HEADERS,
    json={"to_stage": "production"},
)
print("Promoted to production")

# 4. Later: rollback if needed
# requests.post(
#     f"{BASE_URL}/api/mlops/deployments/{deployment_id}/rollback",
#     headers=HEADERS,
#     json={"to_model_id": OLD_MODEL_ID},
# )