Ping-Pong¶
Argo Rollouts version
Ping-pong support for plugin-based traffic routers requires Argo Rollouts v1.10 or later.
Overview¶
The ping-pong feature enables zero-downtime rollouts for long-lived TCP and gRPC connections. It is the recommended approach for workloads where dropped connections during promotion are not acceptable — for example, services behind an AWS Network Load Balancer (NLB) with weighted target groups.
How it differs from the default canary strategy¶
In a normal canary rollout, canaryService and stableService use selector swaps at promotion time: the stable service starts pointing at the new ReplicaSet. For AWS NLB (and similar load balancers that track connections at the target-group level), this swap triggers target deregistration and re-registration, which drops all in-flight connections.
Ping-pong avoids the swap entirely. Two persistent services — pingService and pongService — are defined upfront. Their selectors never change. Instead, the rollout controller alternates which one is "stable" via status.canary.stablePingPong, and the plugin shifts traffic weight between them. No selector change occurs at promotion, so no connections are dropped.
Configuration¶
Instead of canaryService / stableService, define a pingPong block:
spec:
strategy:
canary:
pingPong:
pingService: ping-service
pongService: pong-service
pingService and pongService must exist as Kubernetes Services before the rollout starts. Both services should initially select the same pods as the stable deployment.
The TCPRoute (or HTTPRoute / GRPCRoute / TLSRoute) must reference both services as backendRefs.
Example: TCPRoute with ping-pong¶
This is the primary use case — zero-downtime rollouts for long-lived TCP or gRPC connections.
1. Create ping and pong services¶
apiVersion: v1
kind: Service
metadata:
name: ping-service
namespace: default
spec:
selector:
app: rollouts-demo
ports:
- port: 80
targetPort: 8080
---
apiVersion: v1
kind: Service
metadata:
name: pong-service
namespace: default
spec:
selector:
app: rollouts-demo
ports:
- port: 80
targetPort: 8080
2. Create TCPRoute with both services as backendRefs¶
apiVersion: gateway.networking.k8s.io/v1alpha2
kind: TCPRoute
metadata:
name: rollouts-demo-tcproute
namespace: default
spec:
parentRefs:
- name: my-gateway
sectionName: tcp
namespace: default
rules:
- backendRefs:
- name: ping-service
port: 80
- name: pong-service
port: 80
3. Create the Rollout¶
apiVersion: argoproj.io/v1alpha1
kind: Rollout
metadata:
name: rollouts-demo
namespace: default
spec:
replicas: 2
strategy:
canary:
pingPong:
pingService: ping-service
pongService: pong-service
trafficRouting:
plugins:
argoproj-labs/gatewayAPI:
tcpRoute: rollouts-demo-tcproute
namespace: default
steps:
- setWeight: 20
- pause: { duration: 10 }
- setWeight: 50
- pause: { duration: 10 }
- setWeight: 80
- pause: { duration: 10 }
revisionHistoryLimit: 2
selector:
matchLabels:
app: rollouts-demo
template:
metadata:
labels:
app: rollouts-demo
spec:
containers:
- name: rollouts-demo
image: argoproj/rollouts-demo:blue
ports:
- name: tcp
containerPort: 8080
protocol: TCP
Example: HTTPRoute with ping-pong¶
The same approach works for HTTPRoute when you want to avoid selector swaps for HTTP workloads.
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
name: rollouts-demo-httproute
namespace: default
spec:
parentRefs:
- name: my-gateway
rules:
- backendRefs:
- name: ping-service
port: 80
- name: pong-service
port: 80
---
apiVersion: argoproj.io/v1alpha1
kind: Rollout
metadata:
name: rollouts-demo
namespace: default
spec:
replicas: 2
strategy:
canary:
pingPong:
pingService: ping-service
pongService: pong-service
trafficRouting:
plugins:
argoproj-labs/gatewayAPI:
httpRoute: rollouts-demo-httproute
namespace: default
steps:
- setWeight: 20
- pause: { duration: 10 }
- setWeight: 80
- pause: { duration: 10 }
revisionHistoryLimit: 2
selector:
matchLabels:
app: rollouts-demo
template:
metadata:
labels:
app: rollouts-demo
spec:
containers:
- name: rollouts-demo
image: argoproj/rollouts-demo:blue
ports:
- containerPort: 8080
How it works¶
At each rollout step:
- The controller reads
status.canary.stablePingPongto determine which service is currently stable (pingorpong). - The plugin calls
GetStableAndCanaryServicesfrom the Argo Rollouts core library, which returns the correct(stable, canary)pair based on that status field. - The plugin updates the
backendRefweights in the route accordingly.
At promotion, the controller flips status.canary.stablePingPong to point at the new stable service. No Kubernetes Service selector is modified, so no connections are dropped.