MLflow

Overview

MLflow is the de‑facto open‑source platform for logging experiments, tracking metrics, storing artifacts and registering models. The MLFlow Core app on Apolo gives you a production‑ready tracking server with:

• A lightweight web UI

• Flexible metadata backend — SQLite or Postgres

• Built‑in artifact storage on Apolo Files

• Optional HTTPS ingress secured by platform auth

Key Features

Feature	How the Apolo App Helps
Experiment Tracking	Logs parameters, metrics & code revisions via REST API.
Artifact Storage	Push model binaries & notebook outputs to an S3‑compatible bucket (storage: URI).
Model Registry	Register & promote models with the same endpoint.
Pluggable Metadata	Choose embedded SQLite (PVC) or external Postgres DSN.
One‑click Scaling	Resize resources by switching the Resource Preset.
Secure Ingress	Auto‑generated HTTPS domain with platform auth.

Apolo Deployment

There are two ways to deploy the app:

1. Web Console UI – click‑through wizard (no YAML).

2. Apolo CLI – declarative app install command that fits CI/CD.

Web Console UI

1 · Open the catalogue

2 · Fill the wizard

Section	Field	Example	Notes
Resource Preset	cpu-large	4 vCPU / 8 GiB	Pick any preset.
Enable HTTP Ingress	auth = true	—	Creates https://mlflow‑<id>.apps.<cluster>.apolo.us.
Metadata Storage	SQLite – PVC mlflow-sqlite-storage Postgres – postgresql://user:pwd@host:5432/mlflow	Choose one backend.
Artifact Store	storage:mlflow-artifacts	Path in Apolo Files.

Click Install. Wait until Status → healthy and copy the External MLFlow URL.

Apolo CLI

The CLI flow mirrors the UI fields but lets you automate everything.

# 1. place this in mlflow.yaml

template_name: "mlflow-core"
template_version: "v25.5.0"
input:
  preset:
    name: "cpu-small"
  ingress_http:
    auth: true          # basic‑auth on
  metadata_storage:
    postgres_uri:
      uri: "postgresql://mlflow:Pa55w0rd@postgres-svc:5432/mlflow"
  artifact_store:
    path: "storage:mlflow-artifacts"

# 2. deploy it
apolo app install -f mlflow.yaml \
  --cluster <CLUSTER> --org <ORG> --project <PROJECT>

Explanation

• preset.name=cpu-small – 2 vCPU / 4 GiB (adjust as needed).

• postgres_uri.uri – Points to an existing Postgres service; omit this block to fall back to SQLite.

• ingress_http.auth=true – Adds basic‑auth to the public URL.

---

Inputs / Outputs (schema v1)

Inputs

JSON Path	Default	Description
preset.name	–	Resource preset.
ingress_http.auth	true	Basic‑auth on external URL.
metadata_storage.pvc_name	mlflow-sqlite-storage	For SQLite backend.
metadata_storage.postgres_uri.uri	null	Full Postgres DSN.
artifact_store.path	storage:mlflow-artifacts	Artifact bucket path.

Outputs (seen in Details ▸ Output)

Key	Purpose
external_web_app_url.*	Public UI & REST endpoint.
external_server_url.*	Tracking server URL for clients.

Usage

Use the CLI to set the MLFLOW_TRACKING_TOKEN with the token from your session (only if you enabled platform auth)

export MLFLOW_TRACKING_TOKEN=$(apolo config show-token)

The only requirement is to expose MLFLOW_TRACKING_URI and a basic‑auth header.

import mlflow
from mlflow.tracking import MlflowClient

TRACKING_URI = "https://apolo-taddeus-mlflow-core-d660e438.apps.apolo.us"
mlflow.set_tracking_uri(TRACKING_URI)
mlflow.set_registry_uri(TRACKING_URI)  # optional

client = MlflowClient()
print("Current experiments:")
for exp in client.list_experiments():
    print(exp.name)

References

• MLflow Official Documentation – Comprehensive guides covering Tracking, Projects, Models, and Registry

• MLflow GitHub Repository – Source code, issue tracker, and community discussions