sentinel-mlops 0.1.0

MLOps pipeline sitting between Prometheus and Grafana — pulls metrics, trains models, emits predictions

Sentinel

An MLOps pipeline that sits transparently between Prometheus and Grafana. Sentinel scrapes your metrics, trains lightweight time series models, and emits predictions back as standard Prometheus metrics — so Grafana can display real and predicted data side by side with zero changes to either Prometheus or Grafana.

Prometheus → Sentinel → /metrics → Grafana

How it works

Sentinel has three internal parts:

Data Ingestor pulls metrics from Prometheus on a configurable interval and maintains a sliding window buffer per metric. Once the buffer fills the configured lookback window, cold start training begins automatically.

MLOps Pipeline trains one model per watched metric. After cold start, it monitors prediction drift using a rolling MAE window. When drift crosses a threshold it triggers either a finetune (low severity) or full retrain (high severity). All training runs in a background thread so predictions keep serving during retraining. Every trained model is versioned, persisted to disk, and can be rolled back.

Data Emitter exposes predictions on a /metrics endpoint in standard Prometheus exposition format. Grafana scrapes it the same way it scrapes Prometheus. No plugins, no code changes on either side.

Installation

pip install sentinel-mlops

Quick start

from sentinel import Sentinel, SentinelConfig, WatchConfig
from sentinel.pipeline.models import ExponentialSmoothingModel

config = SentinelConfig(
    prometheus_url="http://localhost:9090",
    emitter_port=8080,
    artifact_store="./sentinel_artifacts",
    watches=[
        WatchConfig(
            metric="http_request_duration_seconds",
            labels={"job": "api"},
            model_class=ExponentialSmoothingModel,
            granularity="1m",
            horizon="5m",
            lookback="30m",
            cron="0 */6 * * *",
        )
    ]
)

Sentinel(config).start()

Point Grafana at http://localhost:8080 as a Prometheus datasource. Done.

Grafana setup

Add a second Prometheus datasource in Grafana pointing to Sentinel's emitter port:

URL: http://localhost:8080

Then on any panel, add a second query using this datasource. Sentinel emits predictions with the original metric name suffixed with _sentinel_predicted:

# original metric
http_request_duration_seconds{job="api"}

# sentinel prediction
http_request_duration_seconds_sentinel_predicted{job="api", sentinel_horizon="5m"}

Both series appear on the same panel as overlapping lines.

Configuration

SentinelConfig

Field	Description	Default
prometheus_url	URL of your Prometheus instance	"http://localhost:9090"
emitter_port	Port Sentinel serves /metrics on	8080
artifact_store	Directory for persisted model artifacts	"./sentinel_artifacts"
max_versions_per_metric	How many model versions to retain per metric	5
scrape_timeout	HTTP timeout for Prometheus requests in seconds	10
emit_confidence_bounds	Whether to emit lower/upper bound gauges	False
watches	List of WatchConfig instances	[]

WatchConfig

Field	Description	Default
metric	Prometheus metric name	Required
labels	Label filters as a dict	{}
model_class	Model class to use	Required
granularity	Resolution of data and predictions e.g. "1m"	"1m"
horizon	How far ahead to predict e.g. "5m"	"5m"
lookback	Feature window size e.g. "30m"	"30m"
cron	Retraining schedule as a cron string	"0 */6 * * *"
drift_finetune_threshold	MAE above this triggers finetune	0.1
drift_retrain_threshold	MAE above this triggers full retrain	0.3

Available models

All models implement the same interface and are interchangeable.

LinearTrendModel

Least squares linear regression. Fastest. Best for metrics with a clear linear trend and no seasonality. Good baseline.

from sentinel.pipeline.models import LinearTrendModel

ExponentialSmoothingModel

Holt's double exponential smoothing. Handles level and trend. Good for metrics with gradual trends but no strong seasonality. Lightweight and interpretable.

from sentinel.pipeline.models import ExponentialSmoothingModel

ARIMAModel

ARIMA(p, d, q) implemented without external dependencies. Suitable for stationary or trend-stationary metrics. Configurable order.

from sentinel.pipeline.models import ARIMAModel

WatchConfig(
    model_class=ARIMAModel,
    ...
)

SGDRegressorModel

Online learning via stochastic gradient descent. Most finetune-friendly — continues training from existing weights without reinitializing. Best for metrics that evolve continuously.

from sentinel.pipeline.models import SGDRegressorModel

Drift detection and retraining

Sentinel continuously compares its predictions against actual values fetched from Prometheus. It maintains a rolling MAE window per metric and classifies drift severity:

Condition	Action
MAE < drift_finetune_threshold	No drift, keep serving current model
drift_finetune_threshold <= MAE < drift_retrain_threshold	Low drift, finetune on recent data
MAE >= drift_retrain_threshold	High drift, full retrain from scratch

Retraining also fires on the cron schedule regardless of drift, so models stay fresh even on stable metrics.

Model versioning and rollback

Every training run produces a versioned artifact stored in artifact_store. To roll back a metric to its previous model version:

sentinel = Sentinel(config)
sentinel.rollback(metric="http_request_duration_seconds", labels={"job": "api"})

On restart, Sentinel automatically restores the last active model from disk so predictions resume without waiting for a cold start.

Checking status

status = sentinel.status()
# returns per-metric dict with:
# ready, active_version, active_version_mae, drift_mae, buffer_fill

Emitted metrics

For each watched metric Sentinel emits:

Metric	Description
{metric}_sentinel_predicted	Predicted values, one gauge per horizon step
{metric}_sentinel_predicted_lower	Lower confidence bound (if emit_confidence_bounds=True)
{metric}_sentinel_predicted_upper	Upper confidence bound (if emit_confidence_bounds=True)

Each gauge carries these labels in addition to the original metric labels:

Label	Description
sentinel_horizon	Prediction horizon e.g. "5m"
sentinel_version	Model version that produced the prediction e.g. "v3"
sentinel_step	Step index within the horizon, 1-indexed

Bringing your own model

Subclass BaseModel and implement five methods:

from sentinel.pipeline.models import BaseModel, PredictionResult, TrainingResult
import numpy as np

class MyModel(BaseModel):

    def fit(self, X: np.ndarray, y: np.ndarray) -> TrainingResult:
        ...

    def partial_fit(self, X: np.ndarray, y: np.ndarray) -> TrainingResult:
        ...

    def predict(self, X: np.ndarray) -> PredictionResult:
        ...

    def save(self, path: str) -> None:
        ...

    def load(self, path: str) -> None:
        ...

Then pass it as model_class in WatchConfig.

X has shape (n_samples, n_features) where features are auto-generated lag values, rolling mean, rolling std, and cyclical time encodings. y has shape (n_samples,). predict() receives X with shape (1, n_features) and must return a PredictionResult with values of length horizon_steps.

Requirements

• Python >= 3.11
• numpy >= 1.26.0
• httpx >= 0.27.0
• joblib >= 1.3.0
• prometheus-client >= 0.20.0
• croniter >= 2.0.0

License

MIT