Proxy vs Passthrough Network Load Balancers in Google Cloud

GCP Study Hub
November 18, 2026

Google Cloud's Network Load Balancers come in two designs. A proxy Network Load Balancer ends the client's TCP connection and opens a new connection to the backend. A passthrough Network Load Balancer does not end the client connection, so packets arrive at the backends unchanged, with the original client source IP address and the original destination port. Many of the differences between them follow from that design choice, including which protocols they carry, whether they can map ports and what source address the backends see. For the Professional Cloud Network Engineer exam, it helps to sort a requirement into one of these two designs before choosing the specific load balancer.

What a proxy Network Load Balancer can do

Because it terminates the client connection and opens its own, a proxy Network Load Balancer has abilities a passthrough design lacks:

  • Port mapping. It can accept traffic on one port and send it to a different port on the VMs, without any host-level redirection.
  • Edge termination. The global external proxy Network Load Balancer ends client connections at a Google Front End close to the user, which shortens connection setup. It then carries the traffic over Google's network to a backend region chosen by proximity, backend health and available capacity, all behind one anycast IP address.
  • IPv6 frontends. It supports IPv6 frontends alongside IPv4, including one anycast IPv4 address and one anycast IPv6 address for a global service.

TCP proxy and SSL proxy modes

A proxy Network Load Balancer runs in one of two modes, and the choice depends on where TLS should end.

In TCP proxy mode, the load balancer only forwards the TCP stream. A TLS session inside that stream passes through untouched, the private keys stay on the backends, and the backends can perform mutual TLS themselves. No certificate is uploaded to Google Cloud in this mode, and TLS stays end to end between the client and the backend.

In SSL proxy mode, the load balancer decrypts traffic at the edge using a certificate attached to the load balancer. Because the load balancer holds the certificate and decrypts the session, TLS ends at Google's edge rather than at the backends.

Limits of the proxy design

The same termination that enables port mapping also sets the limits of a proxy Network Load Balancer:

  • Both modes are TCP only and do not support UDP.
  • Backends see the proxy's address as the source, not the client's. PROXY protocol can pass the client address along in an extra header at the start of the stream, but that only helps when the backend software can parse it.
  • It works at the TCP layer, so it cannot route by HTTP hostname or request path.
  • It does not accept Cloud Run service URLs as backends.

The internal proxy variants behave the same way on source addresses. A cross-region internal proxy Network Load Balancer and regional internal proxy Network Load Balancers also proxy traffic, so their backends see a proxy-only subnet address as the source.

How a passthrough Network Load Balancer behaves

A passthrough Network Load Balancer forwards packets rather than connections. The backends receive the client's own source IP address and the destination port the client used. It is regional, and it handles TCP on any port as well as UDP. It comes in external and internal forms.

The external passthrough Network Load Balancer

The external passthrough Network Load Balancer has several properties that decide when it fits:

  • It is the only external load balancer that accepts UDP traffic from internet clients, so it fits syslog on UDP port 514, SIP and RTP audio, and similar protocols.
  • It preserves the client IP address. Backend software that inspects the source address of each packet, such as a VoIP application, can rely on that. It also means ordinary VPC firewall rules on the backends can filter traffic by caller.
  • It cannot map ports, so the VMs receive traffic on the same port the clients used.
  • Its recommended design uses a regional backend service, which can point to managed or unmanaged instance groups. A backend service-based design supports IPv6 forwarding rules, but those need backends that have IPv6 addresses.
  • Legacy target pools are an older backend type that cannot serve IPv6 forwarding rules.
  • It works at layer 4, so it does not support Cloud CDN. Because it is regional, it cannot provide a global anycast address.

The internal passthrough Network Load Balancer

The internal passthrough Network Load Balancer keeps the client IP address and handles non-HTTP TCP and UDP traffic on multiple ports. It gets only an internal address, so devices on the public internet cannot reach it. Internet-facing UDP or client-IP requirements point to the external passthrough design instead.

Choosing between proxy and passthrough

A few requirements settle the choice quickly. UDP points to passthrough, because neither proxy mode supports it. A backend that reads the client address from the IP header points to passthrough, unless the software can parse PROXY protocol. Different frontend and backend ports point to proxy. A single global anycast address for TCP traffic points to the global external proxy Network Load Balancer, and within that design, where TLS is allowed to end decides between SSL proxy and TCP proxy mode. We would check those signals before looking at anything else, because each one rules out a group of load balancers on the Professional Cloud Network Engineer exam.

Our Professional Cloud Network Engineer practice exams cover proxy and passthrough Network Load Balancers alongside load balancer selection and health checks, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow