When utilizing SRS Edge to forward traffic to the origin cluster, the traffic distribution is highly uneven.

[nighttidesy]

Describe the bug
When utilizing SRS Edge and SRS Origin clusters, it has been observed that the majority of the stream pushing from SRS Edge is directed to the first node of the SRS Origin cluster, resulting in significantly less traffic to the other nodes and thus causing an imbalance in traffic distribution.

Version
srs4

To Reproduce

1. SRS Edge Configuration
   [root@k8s-ops srs-cluster]# cat srs-edge.yaml
   apiVersion: v1
   kind: ConfigMap
   metadata:
   name: srs-edge-config
   data:
   srs.conf: |-
   listen 1935;
   max_connections 1000;
   daemon off;
   Enable smooth exit feature.
   grace_start_wait 700;
   grace_final_wait 800;
   force_grace_quit on;
   http_api {
   enabled on;
   listen 1985;
   }
   http_server {
   enabled on;
   listen 8080;
   }
   vhost defaultVhost {
   cluster {
   mode remote;
   origin srs-origin-0.socs srs-origin-1.socs srs-origin-2.socs;
   token_traverse on;
   }
   http_remux {
   enabled on;
   mount [vhost]/[app]/[stream].flv;
   hstrs on;
   }
   }

---

apiVersion: apps/v1
kind: Deployment
metadata:
name: srs-edge-deploy
labels:
app: srs-edge
spec:
replicas: 3
revisionHistoryLimit: 10
selector:
matchLabels:
app: srs-edge
template:
metadata:
annotations:
version/config: "20231228201317"
labels:
app: srs-edge
spec:
volumes:
- name: config-volume
configMap:
name: srs-edge-config
- name: av-cfs
persistentVolumeClaim:
claimName: av-cfs-pvc
containers:
- name: srs
image: ossrs/srs:4
imagePullPolicy: IfNotPresent
ports:
- containerPort: 1935
- containerPort: 1985
- containerPort: 8080

    resources:
      limits:
        memory: 12Gi
        cpu: 2000m
      requests:
        memory: 4Gi
        cpu: 500m

    volumeMounts:
    - name: config-volume
      mountPath: /usr/local/srs/conf
    - name: av-cfs
      mountPath: /app/nfs/

    lifecycle:
      preStop:
        exec:
          command: ["/usr/local/srs/etc/init.d/srs", "grace"]
  terminationGracePeriodSeconds: 120

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
          - matchExpressions:
            - key: fadada.com/avs2.domain
              operator: In
              values:
              - standard
    podAntiAffinity:
      preferredDuringSchedulingIgnoredDuringExecution:
      - weight: 50
        podAffinityTerm:
          labelSelector:
            matchExpressions:
            - key: app
              operator: In
              values:
              - srs-edge
          topologyKey: "kubernetes.io/hostname"

---

apiVersion: v1
kind: Service
metadata:
name: srs-edge-service
spec:
type: NodePort
selector:
app: srs-edge
ports:

• name: srs-edge-service-1935-1935
  port: 1935
  protocol: TCP
  targetPort: 1935
  nodePort: 30228
• name: srs-edge-service-1985-1985
  port: 1985
  protocol: TCP
  targetPort: 1985
  nodePort: 30229
• name: srs-edge-service-8080-8080
  port: 8080
  protocol: TCP
  targetPort: 8080

Configure SRS (Simple Real-time Streaming) origin settings.
[root@k8s-ops srs-cluster]# cat srs-origin.yaml
apiVersion: v1
kind: ConfigMap
metadata:
name: srs-origin-config
data:
srs.conf: |-
listen 1935 {
idle_timeout 600;
}
max_connections 1000;
daemon off;
srs_log_tank console;
srs_log_file ./objs/srs.log;
http_api {
enabled on;
listen 1985;
}
http_server {
enabled on;
listen 8080;
dir ./objs/nginx/html;
}
vhost defaultVhost {
cluster {
origin_cluster on;
coworkers srs-origin-0.socs:1985 srs-origin-1.socs:1985 srs-origin-2.socs:1985;
}

    play {
        time_jitter             full;
        mix_correct             on;
    }

Configure HTTP-FLV settings.
http_remux{
enabled on;
mount [vhost]/[app]/[stream].flv;
hstrs on;
}

    hls {
        enabled    off;
    }

Configure DVR settings.
dvr {
enabled on;
dvr_path /app/nfs/livevideo/srs/[app]/[stream]_[timestamp].flv;
dvr_plan segment;
Split video files by time.
dvr_duration 600;
dvr_wait_keyframe on;
}

HTTP callback.
http_hooks {
enabled on;

AVS management side token verification interface.
on_publish http://avs-admin:9102/v1/token/check;

When the client starts streaming, the AVS management side token verification interface.
on_play http://avs-admin:9102/v1/token/check;
}

}

---

apiVersion: v1
kind: Service
metadata:
name: socs
spec:
clusterIP: None
selector:
app: srs-origin
ports:

• name: socs-1935-1935
  port: 1935
  protocol: TCP
  targetPort: 1935

---

apiVersion: apps/v1
kind: StatefulSet
metadata:
name: srs-origin
labels:
app: srs-origin
spec:
serviceName: "socs"
replicas: 3
selector:
matchLabels:
app: srs-origin
template:
metadata:
labels:
app: srs-origin
spec:
volumes:
- name: cache-volume
persistentVolumeClaim:
claimName: av2-srs-pvc
- name: config-volume
configMap:
name: srs-origin-config
- name: av-cfs
persistentVolumeClaim:
claimName: av-cfs-pvc
containers:
- name: srs
image: ossrs/srs:4
imagePullPolicy: IfNotPresent
ports:
- containerPort: 1935
- containerPort: 1985
- containerPort: 8080
volumeMounts:
- name: cache-volume
mountPath: /usr/local/srs/objs/nginx/html
readOnly: false
- name: config-volume
mountPath: /usr/local/srs/conf
- name: av-cfs
mountPath: /app/nfs/

  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
          - matchExpressions:
            - key: fadada.com/avsrs.node
              operator: In
              values:
              - standard
    podAntiAffinity:
      preferredDuringSchedulingIgnoredDuringExecution:
      - weight: 50
        podAffinityTerm:
          labelSelector:
            matchExpressions:
            - key: app
              operator: In
              values:
              - srs-origin
          topologyKey: "kubernetes.io/hostname"

---

SRS currently does not support cluster-level APIs. To retrieve streaming live broadcast information from all origin servers, it is necessary to iteratively call the API of each origin server and aggregate the data.

apiVersion: v1
kind: Service
metadata:
name: srs-api-service-0
spec:
type: NodePort
selector:
#app: srs-origin
statefulset.kubernetes.io/pod-name: srs-origin-0
ports:

• name: srs-origin-service-1985-1985
  port: 1985
  protocol: TCP
  targetPort: 1985

---

apiVersion: v1
kind: Service
metadata:
name: srs-api-service-1
spec:
type: NodePort
selector:
#app: srs-origin
statefulset.kubernetes.io/pod-name: srs-origin-1
ports:

• name: srs-origin-service-1985-1985
  port: 1985
  protocol: TCP
  targetPort: 1985

---

apiVersion: v1
kind: Service
metadata:
name: srs-api-service-2
spec:
type: NodePort
selector:
#app: srs-origin
statefulset.kubernetes.io/pod-name: srs-origin-2
ports:

• name: srs-origin-service-1985-1985
  port: 1985
  protocol: TCP
  targetPort: 1985

3. Use a stress testing tool to push 30 streams. By accessing the origin API interface, it is observed that all streams are located on the first origin node, named srs-origin-0.socs.
   

Could you please explain the reason for this? Is it because the edge does not support round-robin forwarding of traffic to the origin by design?

TRANS_BY_GPT4

[nighttidesy]

Is @winlinvip designed this way intentionally? If the traffic is not balanced, the origin server can easily reach its performance limit. Additionally, can I add a load balancer to the origin server cluster? By using the load balancer to distribute traffic to various nodes in the origin cluster, and then configuring the edge cluster to forward traffic to the address of this load balancer?

TRANS_BY_GPT4

[winlinvip]

Please use new origin server, the SRS proxy, see https://ossrs.io/lts/en-us/docs/v7/doc/origin-cluster