Skip to content

Chapter 22 — ZTNA Connector vs Service Connections & Hosting Environments

This chapter compares ZTNA Connector and Service Connections to clarify when to use each, and documents the supported platforms and hardware requirements for deploying Connector VMs.


ZTNA Connector vs Service Connections

The two technologies provide private application access through different models. Neither is strictly a replacement for the other — they address different scenarios.

Capability Service Connection ZTNA Connector
Overlapped private networks (same RFC 1918 space) No — requires NAT Yes
Access to all ports and protocols Yes Limited — app-specific targets
Automatic tunnel establishment to Prisma Access No — manual IPSec/BGP config Yes
Automatic Prisma Access location discovery No Yes
Autoscaling bandwidth (up to 6 Gbps per Connector Group) No Yes
Access via IP address and subnet Yes Yes (from PA 5.0)
Server-initiated traffic (e.g. remote help desk) Yes Limited (from PA 5.0 with add-on)
Unique client IP per session (AD, SMB) Yes No
Dynamic port protocols (FTP active, VoIP) Yes No
Hybrid with on-prem NGFW (User-ID policies) Yes No

Decision guide:

flowchart TD
    Q1{"Overlapped RFC 1918<br/>across DCs?"}
    Q2{"Need full network<br/>access (all ports)?"}
    Q3{"Need AD/SMB or<br/>dynamic ports?"}
    ZTNA["Use ZTNA Connector"]
    SC["Use Service Connection"]

    Q1 -->|Yes| ZTNA
    Q1 -->|No| Q2
    Q2 -->|Yes| Q3
    Q2 -->|No| ZTNA
    Q3 -->|Yes| SC
    Q3 -->|No| ZTNA

📷 PaloAlto diagram — ZTNA Connector vs Service Connections comparison

PaloAlto diagram — ZTNA Connector vs Service Connections comparison

Use Case ZTNA Connector Support Service Connection Support
Access to applications in overlapped networks √ —
Access to all ports and protocols for client-initiated application traffic √ √
Automatic tunnel establishment to Prisma Access √ —
Automatic Prisma Access location discovery √ —
Access to private apps using IP addresses and subnets √ √
Server-initiated traffic (such as a remote help desk) reaching out to a managed device (GlobalProtect mobile user or remote network). √ √
Applications or services that require a unique client IP address √ √
Access to applications with dynamic ports (such as FTP Active mode or VoIP) — √

|Hybrid deployments with on-premises next-generation firewalls where policy rules based on User-ID are applied|√|√|

Supported Hosting Environments

Connector VMs can be deployed on:

Platform Image Format Notes
Amazon AWS Cloud Marketplace or CSP m5.xlarge per region
Google Cloud Platform Cloud Marketplace or CSP n2-standard-4 per region
Microsoft Azure Cloud Marketplace or CSP Standard_D4_v3 per region
Oracle Cloud (OCI) CSP VM.Standard.E5.Flex
VMware ESXi OVA image Thick Provisioned Lazy Zeroed disk
KVM qcow2 image Two-arm recommended
Microsoft Hyper-V VHD image Two-arm recommended

Hardware Specifications

All Connector VMs require the same minimum resources regardless of platform:

Resource Requirement
vCPU 4
Memory 16 GB (8 GB acceptable for ESXi/Hyper-V)
Disk 40 GB

Deployment Topologies

Topology Interfaces Supported On
One-arm Single interface (WAN + app traffic combined) AWS, Azure, GCP, OCI, ESXi
Two-arm Separate WAN-facing and app-facing interfaces KVM, Hyper-V (recommended)

One-arm is simpler; two-arm provides granular separation between internet and DC traffic.


Operational Constraints

  • vMotion, snapshots, and live VM migration are not supported — Connector VMs must remain static
  • IPv6 connectivity to the Prisma Access controller is not supported
  • RFC 6598 shared-address space is not supported in a default deployment — requires Palo Alto Networks to activate it (see Chapter 23 for IP planning detail)
  • Service Connection subnets must not overlap with ZTNA IP blocks (Application IP Block or Connector IP Block)
  • Do not mix NGFW (Service Connection) and ZTNA Connectors in the same Connector Group
  • Limit a single FQDN to 4 Connectors per region

Key Takeaways

  • Service Connection: full network access, unique client IPs, dynamic ports, NGFW integration — but no overlap support and manual tunnel config
  • ZTNA Connector: automated tunnels, overlap support, auto-scaling, app-granular — but no dynamic ports or unique client IPs
  • Connector VMs run on all major cloud platforms plus on-premises (ESXi, KVM, Hyper-V)
  • Hardware minimum: 4 vCPU, 16 GB RAM, 40 GB disk on all platforms
  • vMotion and snapshots not supported; ZTNA and SC IP blocks must not overlap

Previous: Chapter 21 — ZTNA Use Cases & Packet Flow · Next: Chapter 23 — Network Requirements & Prerequisites