---
title: Deployment
description: Run Avra in Avra's cloud, or in your own environment, with the data layer in your Kubernetes cluster and Avra RFM in your AWS account.
sidebar:
  label: Overview
---

Avra runs in one of two places. In Avra's cloud, Avra operates everything for you. In your own environment, Avra's software runs next to your data: the data layer in your Kubernetes cluster, and Avra RFM in your AWS account on Amazon SageMaker. Avra's control plane manages the data layer from outside, and nothing connects into your network. Your data is protected in both.

This page explains what runs where in each option, what crosses the boundary between Avra and your environment, and how to choose.

**Avra's cloud**

Avra runs everything for you. Your workspace is a tenant of its own, isolated from every other customer, and your data, your models, and their records stay inside it.

**Your environment**

Avra's data layer runs in your Kubernetes cluster, installed with Helm. Avra RFM runs in your AWS account on Amazon SageMaker, through AWS Marketplace. Avra's control plane manages the data layer from outside, with no inbound connection.

## Deploying in your environment

A deployment in your environment follows five steps. You provide the infrastructure and run one install command, and your Avra team configures the rest with you.

1. **Plan with your Avra team**

    Agree on what you provide: a Kubernetes cluster, a bucket, a PostgreSQL database, outbound network access, and an AWS account. See [Before you start](/deployment/data-layer#before-you-start).

2. **Install the data layer**

    Run the install command Avra sends you. A deployment agent installs the data layer in your cluster, and the Avra agent starts polling Avra's control plane. See [Install the data layer](/deployment/data-layer).

3. **Bring your data**

    Agree on the data contract, and Avra sets up your pipelines with you. Each run that passes the contract's checks publishes a new version in your bucket. See [Data ingestion](/data-ingestion).

4. **Subscribe to Avra RFM**

    Subscribe to the Avra RFM listing on AWS Marketplace from the account where your data lives, then prepare S3 and an execution role. See [Avra RFM on AWS SageMaker](/deployment/aws-sagemaker).

5. **Predict**

    Run inference jobs to score records, and training jobs to adapt or fine-tune a copy of the model, all in your AWS account.

## What runs where

<svg class="fig" viewBox="0 0 640 444" role="img" aria-label="Two areas. Avra's cloud holds the control plane, releases, and the telemetry endpoint. Your environment holds your Kubernetes cluster, with the Avra agent, the ingestion pipeline, the deployment agent, and the telemetry collector; your bucket and database; and your AWS account, where Avra RFM runs as SageMaker jobs. The Avra agent, the deployment agent, and the telemetry collector each open an outbound connection to Avra's cloud. The ingestion pipeline writes to your bucket and database, and Avra RFM reads the published data.">
  <defs>
    <marker id="fig-deploy-arrow" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
      <path d="M0,0 L10,5 L0,10 z" class="arrow" />
    </marker>
    <marker id="fig-deploy-arrow-accent" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
      <path d="M0,0 L10,5 L0,10 z" class="arrow-accent" />
    </marker>
  </defs>
  <rect class="soft" x="16" y="16" width="608" height="96" rx="12" />
  <text class="small title accent" x="32" y="38">Avra's cloud</text>
  <text class="small label end" x="608" y="38">Every connection opens from your side</text>
  <rect class="card" x="32" y="50" width="184" height="48" rx="8" />
  <text class="small title" x="48" y="70">Control plane</text>
  <text class="small label" x="48" y="87">dashboard, API, registry</text>
  <rect class="card" x="300" y="50" width="148" height="48" rx="8" />
  <text class="small title" x="316" y="70">Releases</text>
  <text class="small label" x="316" y="87">charts and versions</text>
  <rect class="card" x="464" y="50" width="144" height="48" rx="8" />
  <text class="small title" x="478" y="70">Telemetry endpoint</text>
  <text class="small label" x="478" y="87">operational data</text>
  <line class="edge-accent" x1="98" y1="206" x2="98" y2="102" marker-end="url(#fig-deploy-arrow-accent)" />
  <text class="small label" x="106" y="146">polls and reports</text>
  <line class="edge" x1="374" y1="206" x2="374" y2="102" marker-end="url(#fig-deploy-arrow)" />
  <text class="small label" x="382" y="146">pulls releases</text>
  <line class="edge" x1="520" y1="206" x2="520" y2="102" marker-end="url(#fig-deploy-arrow)" />
  <text class="small label" x="528" y="146">sends telemetry</text>
  <rect class="rule dash" x="16" y="160" width="608" height="268" rx="12" />
  <text class="small title label" x="32" y="416">Your environment</text>
  <rect class="card" x="32" y="192" width="576" height="104" rx="10" />
  <text class="small title label end" x="592" y="284">Your Kubernetes cluster · data layer</text>
  <rect class="soft" x="48" y="208" width="100" height="50" rx="8" />
  <text class="small title mid" x="98" y="229">Avra agent</text>
  <text class="small label mid" x="98" y="246">runs commands</text>
  <rect class="head" x="156" y="208" width="144" height="50" rx="8" />
  <text class="small title mid" x="228" y="229">Ingestion pipeline</text>
  <text class="small label mid" x="228" y="246">and Spark</text>
  <rect class="head" x="308" y="208" width="132" height="50" rx="8" />
  <text class="small title mid" x="374" y="229">Deployment agent</text>
  <text class="small label mid" x="374" y="246">installs layers</text>
  <rect class="head" x="448" y="208" width="144" height="50" rx="8" />
  <text class="small title mid" x="520" y="229">Telemetry collector</text>
  <text class="small label mid" x="520" y="246">metrics and logs</text>
  <line class="edge" x1="200" y1="260" x2="200" y2="314" marker-end="url(#fig-deploy-arrow)" />
  <rect class="card" x="48" y="318" width="192" height="48" rx="8" />
  <text class="small title" x="64" y="338">Your bucket and database</text>
  <text class="small label" x="64" y="355">tables and run state</text>
  <rect class="card" x="344" y="310" width="264" height="88" rx="10" />
  <text class="small title" x="360" y="330">Your AWS account</text>
  <rect class="solid" x="360" y="340" width="232" height="46" rx="8" />
  <text class="small title on-solid mid" x="476" y="359">Avra RFM</text>
  <text class="small on-solid mid" x="476" y="376">SageMaker jobs, network-isolated</text>
  <path class="edge" d="M240,342 C298,342 300,363 356,363" marker-end="url(#fig-deploy-arrow)" />
  <text class="small label mid" x="296" y="334">published data</text>
</svg>

Avra's control plane stays in Avra's cloud. In your environment, the data layer runs in your Kubernetes cluster, and Avra RFM runs as network-isolated SageMaker jobs in your AWS account. Every connection that crosses the boundary starts on your side.

| Component | Runs in | What it does |
|-----------|---------|--------------|
| **Control plane** | Avra's cloud | The dashboard, the API, and the model registry. It queues commands for your environment and receives their results |
| **Avra agent** | Your Kubernetes cluster | Polls the control plane over outbound HTTPS, runs the commands it receives, and reports back |
| **Ingestion pipeline** | Your Kubernetes cluster | Validates and versions your data, with Spark doing the processing. Its tables live in your bucket, and its run state in your database. See [Data ingestion](/data-ingestion) |
| **Deployment agent** | Your Kubernetes cluster | Installs and upgrades the data layer, and reports the status of each release |
| **Telemetry collector** | Your Kubernetes cluster | Forwards operational telemetry to Avra |
| **Avra RFM** | Your AWS account | Runs on Amazon SageMaker: training jobs for adaptation and fine-tuning, inference jobs for predictions |
| **Avra GFM** | Avra's cloud only | Never deployed to your environment. Models built on its embeddings can run in your environment with a snapshot of the embeddings they use |

What returns to Avra is execution metadata, such as each command's status and result, and operational telemetry. Your data, the models adapted to it, and their outputs stay in your environment.

## What crosses the boundary

Every connection between your environment and Avra opens from your side, over HTTPS. The Avra agent asks the control plane for work, the commands come back in the reply, and the agent sends each result in a request of its own. Nothing in your environment listens for connections from Avra, so your network needs no inbound rule.

```mermaid
---
config:
  sequence:
    width: 110
    actorMargin: 36
---
sequenceDiagram
    autonumber
    box Avra's cloud
        participant TE as Telemetry<br/>endpoint
        participant CP as Control<br/>plane
    end
    box Your environment
        participant AG as Avra<br/>agent
        participant TC as Telemetry<br/>collector
    end
    AG->>CP: Poll, with health
    CP-->>AG: Pending commands
    AG->>AG: Run them locally
    AG->>CP: Report each result
    TC->>TE: Metrics, logs, traces
    Note over TE,TC: Every request opens from your side. Nothing connects in.
```

| Flow | Opened by | What it carries |
|------|-----------|-----------------|
| **Polls** | Avra agent | The agent's version, your cluster name, and the health of its layer. The reply carries pending commands, cancellations, and when to poll next |
| **Reports** | Avra agent | Each command's status, its progress, and its result or error |
| **Telemetry** | Telemetry collector | Logs, metrics, and traces from Avra's components, and the resource metrics the kubelet reports for the node the collector runs on and the pods on it, labeled with your cluster name |
| **Releases** | Deployment agent | Fetches your deployment's versions and values, pulls the data layer's charts, and reports the status of each release and its own logs. Depending on its settings, it also sends the logs and resource usage of the layers' containers, and the cluster's CPU and memory totals |
| **Images** | Your nodes' container runtime | Pulls the images the layers run: Avra's from Avra's registries, and those of the open-source components from their public registries |

Your data, the versions the ingestion pipeline publishes, its run state, the copies of Avra RFM adapted to your data, and the scores they produce stay in your environment. Avra RFM's jobs on SageMaker run network-isolated: they can't open a connection to Avra or anywhere else.

## Avra's cloud or your environment

| | Avra's cloud | Your environment |
|---|---|---|
| **Control plane** | Avra | Avra |
| **Data ingestion** | Avra's cloud, inside your isolated workspace | Your Kubernetes cluster |
| **Where your data stays** | Isolated by tenant and workspace in Avra's cloud | Your bucket and your database |
| **Avra RFM** | Avra's cloud | Your AWS account, on Amazon SageMaker |
| **Predictions** | Live through the API, or in batches | Inference jobs that write scores to your S3 |
| **Adapted and fine-tuned copies** | Inside your workspace | Your S3, encrypted |
| **Avra GFM** | Avra's cloud | Avra's cloud only |
| **What returns to Avra** | Not applicable, since everything runs in Avra's cloud | Execution metadata, and operational telemetry |
| **Who runs the infrastructure** | Avra | You run the cluster, the bucket, the database, and the AWS account. Avra installs and upgrades its software |

## Who does what

In your environment, the work splits between you and Avra:

| | You | Avra |
|---|---|---|
| **Cluster, namespace, and network access** | Provide | Specify what's needed |
| **Bucket, database, and identities** | Provide | Specify what's needed |
| **Charts, images, and their versions** | | Build and publish |
| **Installing and upgrading the layers** | Run the install command | Configure the deployment and roll out versions |
| **Pipelines** | Agree on the data contract | Write and run them with you |
| **AWS account, S3, and execution role** | Provide, and subscribe on AWS Marketplace | Specify what's needed |
| **The RFM algorithm and its releases** | | Build, govern, and release |

## Choosing

Avra's cloud asks the least of you. There's nothing to install or operate, and Avra RFM serves live predictions through the API as well as batches. Your workspace is isolated from every other customer's, and nothing about your business trains another customer's model.

Choose your environment when regulation, data sovereignty, or your own policy requires your data to stay in infrastructure you control. It takes a Kubernetes cluster for the data layer and an AWS account for Avra RFM, and a team to run them. In a customer's environment, Avra RFM runs only on Amazon SageMaker, so if yours isn't on AWS, talk to your Avra team before you plan. Avra GFM stays in Avra's cloud either way, and its outputs reach you through [Avra Core](/core).

Contact your Avra representative to plan a deployment in your environment.

**[Install the data layer](/deployment/data-layer)**

What you provide, how the installation runs, and how to verify and upgrade it

**[Avra RFM on AWS SageMaker](/deployment/aws-sagemaker)**

Subscribe on AWS Marketplace, then run training and inference jobs in your account

**[Data ingestion](/data-ingestion)**

How your data becomes validated, versioned inputs for Avra's models

**[Data privacy & compliance](/security/data-privacy-and-compliance)**

How Avra keeps your data isolated, wherever it runs
