Deployment Guide

Introduction

Item	Content
Applicable Role	Pre-sales solution engineer, delivery owner, implementation engineer, customer architect
Navigation Path	Deployment > Deployment Guide
Function Description	Helps users select deployment mode, request resources, and plan networking, middleware, and compute node onboarding before installation

This document answers "how should we deploy" rather than "which command should I run now". For a first read, follow the path below before entering the detailed installation documents.

Beginner Explanation

Use this document before touching the installer. It helps you decide the deployment shape first: PoC or production, single-node or multi-node, cloud-managed middleware or self-managed middleware, and which ports must be opened.

Reading Path

Stage	What You Decide	Output
Step 1: Confirm deployment goal	PoC, public-cloud production, or private-cloud / IDC	Deployment mode is selected
Step 2: Confirm middleware approach	Use cloud-provider managed middleware or self-managed middleware deployed by the installer	managed-middleware, hybrid, or self-managed is selected
Step 3: Confirm resource sizing	Required machine count, CPU, memory, and disk	Resource request list is ready
Step 4: Confirm network and ports	How users, business nodes, middleware, and compute clusters communicate	Security group / firewall allowlist is ready
Step 5: Enter execution docs	Choose single-node, multi-node, or compute node installation docs based on the decision	Ready for precheck and installation

Terminology Quick Reference

Term	Plain Explanation
PoC	Proof-of-concept or demo environment focused on quickly validating functionality, not high availability
Production deployment	Deployment for formal business use, with high availability, backup, monitoring, permissions, and operations requirements
Management plane	Area where AGIOne control plane and business services run; users mainly access this part
Compute node onboarding	Connecting GPU / NPU nodes to the platform for training, inference, IDE, and similar workloads
Public-cloud managed middleware	Cloud-provider database, Redis, Nacos, Kafka, and object-storage services; formal delivery may use only the products and versions in this document's compatibility matrix
Compatibility baseline	A combination of cloud provider, product, version, and access mode that has passed installation, initialization, runtime, and cleanup regression
Private cloud / IDC	Customer-owned data center or private cloud environment, usually requiring more self-managed middleware
VPC	Cloud private network. Business nodes and middleware should usually be in the same VPC.
ELB / ALB	Load balancer entry that distributes user requests to multiple App / Edge nodes

1. Overview

An AGIOne platform deployment is logically divided into two relatively independent parts:

• Management plane: Hosts the AGIOne control plane, business services, databases, and middleware, and serves as the entry point for users and the platform.
• Compute node onboarding: Integrates accelerator nodes such as GPU / NPU nodes into unified scheduling and hosts compute workloads such as training, inference, and IDE sessions.

The two parts communicate over the internal network. The management plane manages compute clusters and collects observability data through the Kubernetes API and monitoring interfaces.

2. Overall Architecture

The following diagram shows the overall logical architecture of the AGIOne platform, including the relationship between the user access entry point, the management plane, and the compute node onboarding layer.

Architecture highlights:

• The access layer can expose management-plane services through a load-balancing service accepted for the project. In production, a domain name + HTTPS (443) is recommended.
• The management plane consists of at least 2 business nodes + 2 database/middleware nodes, and business nodes can scale horizontally.
• Compute clusters are deployed independently by regional compute pool. Each region deploys a near-node image service to improve image pull performance.
• The management plane schedules and accesses compute clusters through the Kubernetes API (6443) and extended ports (32761-32765).

---

3. Deployment Mode Selection

Deployment Mode	Applicable Scenario	Nodes	High Availability	Recommended Environment
PoC Deployment (All in One)	Proof of concept, demo, internal testing	1	No	Single virtual machine
Production Deployment - Public Cloud	Formal production and external service delivery	Multi-node	Yes	Public cloud (Recommended)
Production Deployment - Private Cloud / IDC	Data compliance and internal network isolation	>= 4	Yes	Customer-owned private cloud or IDC

Selection recommendation: If there are no mandatory data compliance or network isolation requirements, prioritize public cloud deployment to benefit from cloud-provider managed middleware and operational convenience.

The current installer supports only the Alibaba Cloud and Huawei Cloud managed-middleware combinations that have completed verification. "Supported" applies only when the product, version, and access constraints in 5.1.2 Databases and Middleware all match; it does not cover every similar product offered by that provider.

Cloud Provider	Support Status	Covered Managed Middleware	Applicable Installer Mode	Notes
Alibaba Cloud	Conditionally supported	ApsaraDB RDS for MariaDB, Tair / ApsaraDB for Redis, MSE Nacos, ApsaraMQ for Kafka, and OSS	managed-middleware / hybrid	Only the tested versions and access modes in 5.1.2 are supported. Confirm RAM permissions, service-linked roles, VPC connectivity, and allowlist policies before delivery.
Huawei Cloud	Conditionally supported	RDS for MariaDB, DCS for Redis, CSE Nacos, DMS for Kafka, and OBS	managed-middleware / hybrid	Only the tested versions and access modes in 5.1.2 are supported. App / Edge nodes must access managed middleware endpoints through private networking.
Other public clouds	Unsupported	Not verified	Not applicable	Tencent Cloud, AWS, Azure, Google Cloud, and other providers have not completed end-to-end compatibility testing with the current installer and must not be used for formal delivery.

---

4. Management Plane - PoC Deployment (All in One)

4.1 Resource Requirements

Item	Minimum Requirement
Number of nodes	1
CPU	>= 8 cores
Memory	>= 24 GB
Disk	>= 200 GB
Network	Internet access required
Operating system	Linux (Ubuntu 22.04 / CentOS 7+ recommended)

4.2 Architecture Diagram

Item	Description
Scope	AGIOne full-stack deployment solution design, pre-sales support, PoC assessment, and production delivery
Constraint Level	This document is a planning reference. Official delivery is governed by the Release Note and compatibility matrix distributed with agione-release-v1.0-XXX.tar.gz.

• Business services, databases, and middleware are all deployed on the same node.
• Services are exposed externally through HTTP port 18090 by default.
• Because images and dependencies need to be pulled, the deployment node must be able to access the Internet.
• Not recommended for production because it has no high availability or data redundancy.

---

5. Management Plane - Production Deployment (Public Cloud)

Public cloud is the recommended production deployment mode. It leverages cloud-provider managed database, cache, messaging, registry, and object-storage capabilities. Entry load balancing requires separate project acceptance.

5.1 Resource Requirements

5.1.1 Management Nodes (Business Nodes)

Item	Requirement
Number of nodes	>= 2
CPU	>= 8 vCPU
Memory	>= 16 GiB
Disk	>= 200 GiB
Private network	All management nodes are in the same VPC
Public network	Internet access available, bandwidth >= 100 Mbps

Optional: Shared storage, such as block storage, of >= 1024 GB can be mounted for shared use by management nodes.

5.1.2 Databases and Middleware

Generic managed middleware baseline requirements:

Component	Purpose	CPU	Memory	Disk	Nodes	Network Requirement
RDS (MySQL-family relational database)	Stores primary AGIOne platform data	>= 4 vCPU	>= 16 GiB	>= 100 GiB	Follow the verified product shape	Same VPC as management nodes
Nacos	Service registration and discovery	Basic specification	-	-	1	Same VPC as management nodes
Redis (cache)	Caches data	Basic specification	-	-	1	Same VPC as management nodes
Kafka (messaging)	Core service message bus	Cluster node specification	-	>= 100 GiB	>= 3	Same VPC as management nodes
Object storage	Stores images and other static resources	-	-	-	-	Network access via AK/SK
Entry load balancing (project acceptance)	AGIOne API load balancing	-	-	-	1	Private connectivity to management nodes; public access, >= 100 Mbps

The following compatibility matrix is the only cloud-native middleware scope currently approved for formal delivery with the installer. Before purchasing resources, also confirm available specifications, availability zones, billing mode, and account permissions in the customer's target region.

Hard compatibility boundary: The database layer supports only the MySQL protocol and SQL ecosystem, but that does not certify every MySQL-compatible database. Current validation covers only Alibaba Cloud RDS MariaDB 10.3 and Huawei Cloud RDS MariaDB 10.11 listed below. Other standard MySQL versions, other MariaDB versions, and Dameng (DM), Kingbase, GaussDB, openGauss, OceanBase (including MySQL-compatible mode), TiDB, PostgreSQL, Oracle, and SQL Server are unsupported. Any unlisted cloud provider, product, version, instance shape, or access protocol is unsupported until the same version completes end-to-end compatibility testing and this matrix is updated. Port reachability or a successful login alone is not compatibility evidence.

Alibaba Cloud Managed Middleware List

AGIOne Component	Verified Alibaba Cloud Product	Verified Version / Shape	Required Delivery Constraints	Support Status
Database	ApsaraDB RDS for MariaDB	MariaDB 10.3; High-availability edition; cloud_essd; tested class mariadb.n2.small.2c	Use a VPC internal endpoint; allow only App nodes in security groups / allowlists; the initialization account needs database, table, schema-change, and read/write permissions	Supported only for this baseline
Redis	Tair / ApsaraDB for Redis	Redis 5.0; Local standard primary/replica; tested class redis.master.small.default	Use a VPC internal endpoint and add App nodes to the allowlist. Redis 4.0, 6.0, 7.0, cluster editions, and other product shapes are outside the current support scope	Supported only for this baseline
Nacos	Microservices Engine (MSE) Nacos	Nacos 2.0 (NACOS_2_0_0); mse_dev; tested shape MSE_SC_1_2_60_c, one instance	Use the native Nacos API; the runtime account must be able to publish configuration and manage services in namespace agione-prod; other versions, Professional edition, or different instance counts require new testing	Supported only for this baseline
Kafka	ApsaraMQ for Kafka	Service version 2.2.0; Normal edition; alikafka.hw.2xlarge; 3 brokers; VPC PLAINTEXT	Use the VPC advertised brokers returned by the service; enable broker-side automatic topic creation and create required consumer groups. SASL / SSL, other versions, and Serverless shapes are outside the current support scope	Supported only for this baseline
Object storage	Object Storage Service (OSS)	Cloud service with no separate engine version; Standard private bucket; internal S3-compatible endpoint	Use virtual-host-style access; prohibit public writes; grant runtime credentials only the required upload, download, list, and delete actions for the target bucket / prefix	Supported only for this access mode

Alibaba Cloud permission focus: When the optional cloud-resource helper creates resources, its account / RAM role must be able to create and query the selected products, manage allowlists, and authorize required service-linked roles. Keep this cloud-resource identity separate from the database, Redis, Nacos, Kafka, and OSS runtime credentials. The main installer consumes only final endpoints, ports, and least-privilege runtime credentials; do not store cloud-account AK/SK in the main installation YAML. Do not start formal installation while RAM policy propagation, a service-linked role, or an endpoint allowlist is incomplete.

Huawei Cloud Managed Middleware List

AGIOne Component	Verified Huawei Cloud Product	Verified Version / Shape	Required Delivery Constraints	Support Status
Database	RDS for MariaDB	MariaDB 10.11; tested as a single instance with CLOUDSSD, class mariadb.n1.large.2	Use a VPC internal endpoint; allow only App nodes through the security group; grant schema-initialization permissions. Primary/standby HA has not completed same-version end-to-end regression	Supported only for the tested single-instance baseline
Redis	Distributed Cache Service (DCS) for Redis	Redis 7.0; single-node shape; tested class redis.single.au1.large.1, 1 GiB	Use a VPC internal endpoint with password and access policy. Primary/standby, Cluster, and other Redis versions have not completed end-to-end regression	Supported only for the tested single-node baseline
Nacos	Cloud Service Engine (CSE) Nacos	Nacos2 service version 2.1.0.24; shape cse.nacos2.c1.large.10; RBAC	Use the native Nacos API; the runtime account must have namespace and config-publish permissions for agione-prod; other CSE versions / shapes are outside the support scope	Supported only for this baseline
Kafka	Distributed Message Service (DMS) for Kafka	Kafka 2.7; 3 brokers; product s6.2u4g.cluster.small; 300 GiB general storage; PLAINTEXT	Use private bootstrap and advertised brokers. Other versions, SASL / SSL, or different product shapes require new compatibility testing	Supported only for this baseline
Object storage	Object Storage Service (OBS)	Cloud service with no separate engine version; private bucket	Use virtual-host-style access and least-privilege AK/SK or temporary authorization; avoid public read/write	Supported only for this access mode

Entry-service boundary: Alibaba Cloud ALB, MSE Cloud-native Gateway, and Huawei Cloud ELB are not part of the current end-to-end managed-middleware compatibility matrix. Before use, the project must accept HTTPS, health checks, sessions, upload size, and long-lived connections; treat these services as unsupported until that acceptance is complete. A generic TCP load balancer cannot replace Kafka advertised broker addresses.

5.1.3 Capacity and Scalability

• AGIOne management nodes can scale horizontally.
• Baseline capacity per node (8 vCPU / 16 GB RAM) is approximately 2,000 requests/minute.
• In scenarios with many long-lived connections or time-consuming requests, each additional node can add approximately 500 requests/minute. Actual values vary by business scenario.

5.2 Architecture Diagram

5.3 Deployment Notes

• Use cloud-provider managed database, Redis, Nacos, Kafka, and object storage only within the compatibility matrix to reduce operational complexity.
• Business nodes can expose an entry through a load-balancing service accepted for the project. After a domain name is configured, use 443 (HTTPS) and 80 (HTTP redirect).
• All internal components are located in the same VPC. Do not expose internal ports across VPCs.
• Recommended public bandwidth is >= 100 Mbps, adjustable according to business volume.

---

6. Management Plane - Production Deployment (Private Cloud / IDC)

Applicable to scenarios where public cloud cannot be used and data must be fully self-managed.

6.1 Resource Requirements

6.1.1 Required Resources (Nodes)

Role	Nodes	CPU	Memory	Disk	Network	Description
Business nodes	>= 2	>= 8 cores	>= 16 GB	>= 200 GB	LAN; Need access Internet, bandwidth >= 100M	Deploy AGIOne business services
Database / middleware nodes	>= 2	>= 8 cores	>= 16 GB	>= 200 GB	LAN	Deploy MariaDB (primary/replica), Nacos, Redis, Kafka, and MinIO
Total	>= 4	-	-	-	-	-

6.1.2 Optional Resources

Resource	Required / Optional	Recommended Configuration	Purpose
Load Balancer (LB)	Optional	Hardware LB such as F5, or software LB such as Nginx / HAProxy + Keepalived	Provides a unified entry point, traffic distribution, and health checks. Without an LB, DNS round-robin can connect directly to business nodes, but it does not provide health checks or automatic removal of failed nodes.
NAS Shared Storage	Optional	Capacity >= 1024 GB; mounted to all nodes through NFS / CIFS	Stores shared logs, service configuration, shared files, and temporary data that multiple nodes must access. Without NAS, use local disks with a log collector such as Filebeat or Fluent Bit.

6.2 Architecture Diagram

6.3 Deployment Notes

• Start with at least 4 nodes: 2 business nodes + 2 database/middleware nodes.
• The database uses primary/replica mode. It is recommended to deploy databases and middleware separately, and split them into more nodes as business scale grows.
• MinIO provides object storage capabilities as an alternative to public cloud OSS.
• Load balancing is optional: Deploy an LB to provide a unified entry point and high availability. Without an LB, use DNS round-robin to connect directly to business nodes.
• NAS shared storage is optional: Use it for shared logs, service configuration, and other cross-node files. Without NAS, use local disks with a log collection solution.

---

7. Compute Node Onboarding

7.1 Design Principles

• Each independent regional compute pool is deployed as a logical unit to avoid network jitter caused by cross-region scheduling.
• Each regional compute pool deploys an image service close to compute nodes and can reuse already-onboarded node resources.
• Deploy a Kubernetes cluster and supporting plugins for scheduling, monitoring, and related capabilities.

7.2 Kubernetes Control Plane Scale

Node requirements:

Total Compute Pool Nodes	Kubernetes Control Plane Nodes	Description
< 3 nodes	1 node	Single control plane, suitable for small compute pools
>= 3 nodes	3 nodes	Highly available control plane with multiple etcd replicas

Storage requirements:

• At least 2 TB NAS storage is required.
• Optional: S3-compatible object storage larger than 2 TB is supported.

Network requirements:

• The management network must be in the same LAN.
• Inter-node network bandwidth must be >= 1000 Mbps.
• Optional internet access: all nodes can access the internet.

7.3 Architecture Diagram

7.4 Externally Exposed Ports

Each compute cluster exposes the following ports to the management plane through NodePort:

Port	Purpose
6443	Kubernetes API Server
32761	Monitoring interface
32762	Model and IDE invocation port
32763	Reserved extension port
32764	Reserved extension port
32765	Reserved extension port

7.5 Deployment Notes

• Compute nodes should have GPU / NPU drivers, container runtime such as containerd, and the corresponding device plugin installed and validated in advance.
• The image service should be located in the same Layer 2 or low-latency network as the compute nodes to accelerate pulling large model images.
• For multi-region deployment, each region should maintain an independent near-node image service to avoid cross-region image pulls.

---

8. Network Planning

8.1 Network Zoning

8.2 Network Requirements Overview

Traffic Direction	Requirement
Users -> ELB	Public network; 443/HTTPS is recommended in production, with DNS resolving the domain name
ELB -> business nodes	Internal network, same VPC
Business nodes <-> DB / middleware	Internal network, same VPC, low latency
Business nodes -> object storage	AK/SK authentication; VPC internal endpoint can be used
Business nodes -> compute cluster	Through ports 6443 and 32761-32765
Compute nodes -> near-node image service	Compute-pool local LAN, gigabit or higher recommended
Management nodes -> Internet	Required for PoC; outbound access is recommended in production for pulling images and upgrades, bandwidth >= 100 Mbps

8.3 VPC / Subnet Recommendations

• Management VPC: All management nodes, RDS, Nacos, Redis, Kafka, and MinIO are located in the same VPC. Subnets can be divided into business-node subnets and data-node subnets.
• Compute VPC: Each regional compute pool uses an independent VPC or subnet and connects to the management VPC through VPC peering or a dedicated line.
• Security groups: Deny by default and allow only the ports listed in the following "Port List" section.

---

9. Port List

9.1 Management Plane Ports

Port	Protocol	Source	Purpose
18090	TCP / HTTP	Users / internal	Default HTTP service port (used by default in PoC)
80	TCP / HTTP	Public users	Production environment HTTP after domain configuration, usually redirects to 443 with 301
443	TCP / HTTPS	Public users	Production environment HTTPS after domain configuration

9.2 Compute Cluster Ports

Port	Protocol	Source	Purpose
6443	TCP	Management plane	Kubernetes API Server
32761	TCP	Management plane	Monitoring interface
32762	TCP	Management plane	Model and IDE invocation
32763	TCP	Management plane	Reserved extension port
32764	TCP	Management plane	Reserved extension port
32765	TCP	Management plane	Reserved extension port

9.3 Internal Service Ports (Reference)

Databases, middleware, and internal application services should be exposed only within the VPC / LAN and not to the public internet. The actual check scope is adjusted automatically based on self-managed middleware, managed middleware, and optional service groups:

Component / Service Group	Default Port (Reference)
RDS (MySQL family)	3306
Nacos	8848 / 9848 / 9849
Redis	6379
Kafka / Kafka UI	9092 / 9093 / 18091
MinIO / MinStore	8080 / 9000 / 9001
AGIOne base application services	3000 / 4000 / 5007 / 7002 / 7003 / 8031 / 8032 / 8033 / 8080 / 8089
kubem optional service group	8021 / 8022 / 18088
cloud optional service group	8011 / 8017 / 8012 / 8001 / 8002
core_isync optional service group	7091 / 18181 / 18082

Actual ports depend on the version and configuration used during deployment.

---

10. Pre-Deployment Checklist

Before deployment, confirm each item to ensure a smooth rollout:

Base Environment

• [ ] Deployment mode has been selected according to the scenario (PoC / public cloud / private cloud IDC)
• [ ] Node quantity and specifications meet resource requirements
• [ ] Operating system and kernel version meet requirements
• [ ] Time is synchronized (NTP), and all nodes use a consistent time zone

Fixed download page: Download link

After opening the page, copy Download URL and MD5 URL. MD5 detects download or transfer corruption only and does not authenticate the package publisher. For production delivery, independently obtain the outer .tar.gz SHA-256 digest through an access-controlled channel and verify it before extraction.

# 1. Download the bundle and verify MD5
ssh root@<target>
AGIONE_RELEASE_PAGE="https://agione.pro/release/download/agione-release-latest"
AGIONE_RELEASE_URL="<copy-the-Download-URL-from-the-page>"
AGIONE_RELEASE_MD5_URL="<copy-the-MD5-URL-from-the-page>"
AGIONE_RELEASE_ARCHIVE="${AGIONE_RELEASE_URL##*/}"

mkdir -p /opt/hyperone && \
cd /opt/hyperone && \
curl -fL -o "$AGIONE_RELEASE_ARCHIVE" "$AGIONE_RELEASE_URL" && \
curl -fL -o "$AGIONE_RELEASE_ARCHIVE.md5" "$AGIONE_RELEASE_MD5_URL" && \
echo "$(awk '{print $1}' "$AGIONE_RELEASE_ARCHIVE.md5")  $AGIONE_RELEASE_ARCHIVE" | md5sum -c -

# 2. Production delivery: verify the outer archive SHA-256
AGIONE_RELEASE_SHA256="<outer-archive-SHA-256-from-a-trusted-delivery-channel>"
echo "$AGIONE_RELEASE_SHA256  $AGIONE_RELEASE_ARCHIVE" | sha256sum -c -

# 3. Extract after verification, then verify the split bundle
AGIONE_RELEASE_DIR="$(tar -tzf "$AGIONE_RELEASE_ARCHIVE" | head -1 | cut -d/ -f1)"
tar -zxvf "$AGIONE_RELEASE_ARCHIVE"
cd "/opt/hyperone/$AGIONE_RELEASE_DIR"
chmod +x ./agione
./agione verify-bundle

verify-bundle validates extracted split-bundle files against SHA256SUMS. Do not use AGIONE_SKIP_BUNDLE_VERIFY=1 to bypass this check in formal delivery.

---

Network

• [ ] Management nodes are located in the same VPC
• [ ] Management nodes can access the Internet, or offline images have been prepared
• [ ] Security group / firewall rules have opened the ports in the port list
• [ ] The compute cluster and management plane have network connectivity

Domain Name and Certificate (Production)

• [ ] Domain name has been applied for and resolved
• [ ] HTTPS certificate has been prepared
• [ ] ELB listeners for 443 / 80 have been planned

Compute Cluster

• [ ] The number of nodes in each regional compute pool has been confirmed, and the K8s control plane scale has been determined (1 or 3)
• [ ] GPU / NPU drivers have been installed and validated
• [ ] Near-node image service nodes have been planned
• [ ] NodePort ports 6443 and 32761-32765 have been allowed for the management plane

---

11. Appendix

11.1 Resource Specification Quick Reference

Deployment Mode	Minimum Nodes	Minimum Per-Node Specification	Total Resource Reference
PoC All in One	1	8C / 24G / 200G	8C / 24G / 200G
Public Cloud (business nodes)	2	8C / 16G / 200G	16C / 32G / 1 TB+
Private Cloud IDC	4	8C / 16G / 200G	32C / 64G / 800G+

11.2 Capacity Estimation Reference

• Baseline: each business node (8C / 16G) supports approximately 2,000 requests/minute
• Scaling: each additional business node adds approximately +500 requests/minute in long-lived connection / complex request scenarios
• Actual capacity should be evaluated based on request complexity, model inference duration, number of concurrent sessions, and other factors.

11.3 Terminology Quick Reference

Term	Description
AGIOne	Name of this platform
All in One	Simplified mode that deploys all components on a single node
VPC	Virtual Private Cloud
ELB	Elastic Load Balancer
RDS	Relational Database Service
AK/SK	Access Key / Secret Key
NodePort	One of the Kubernetes service port exposure methods
Regional compute pool	A group of compute nodes physically located in the same geographic region and connected over the network