Is there a way to set upstream connection max lifetime?

[WEIZIBIN]

Description

I’m encountering an issue in a Kubernetes environment where Kong (accessing upstream services via Kubernetes Service) returns intermittent 502 errors during rolling updates of upstream pods.

Problem

During pod rolling updates, upstream pods are terminated with a preStop hook that sleeps for 15 seconds:

lifecycle:
  preStop:
    exec:
      command:
        - /bin/sh
        - '-c'
        - sleep 15

We are using the K8s Service address for our downstream endpoints

services:
  - name: service-smc
    url: http://service-smc.dev-go-1.svc.cluster.local:8080
    routes:
      - name:service-smc
        paths:
          - /api/service-smc
        strip_path: false

however, it appears that Kong is still routing traffic to connections associated with pods that are in the process of terminating.

120.229.251.78 - - [11/May/2026:13:25:22 +0000] "POST /api/service-smc/grayapi/city/shop-page HTTP/1.1" 200 req_time: "0.010" ups_addr: "10.101.241.164:8080" ups_status: "502" ups_rsp_time: "0.001" "https://servicewechat.com/wx696a42df4f2456d3/1317/page-frame.html" "Mozilla/5.0 (iPhone; CPU iPhone OS 18_7 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Mobile/15E148 MicroMessenger/8.0.73(0x18004925) NetType/WIFI Language/zh_CN" kong_request_id: "42136306e1bc3932d151a33be0c78f69" trace_id: "5f03eb3f00a07cc0ecc3ec01635f2088"
    
2026/05/11 13:25:22 [error] 1406#0: *2290317 recv() failed (104: Connection reset by peer) while reading response header from upstream, client: 120.229.251.78, server: kong, request: "POST /api/service-smc/grayapi/city/shop-page HTTP/1.1", upstream: "http://10.101.241.164:8080/api/service-smc/grayapi/city/shop-page", host: "go.heytea.com", referrer: "https://servicewechat.com/wx696a42df4f2456d3/1317/page-frame.html", request_id: "42136306e1bc3932d151a33be0c78f69"

Suspicion

We suspect the issue is caused by Kong reusing stale keepalive connections from its upstream connection pool after pods have started terminating.
The expectation is that setting a maximum connection lifetime lower than the pod termination delay would avoid this issue.

Attempted Configuration

We tried configuring upstream keepalive behavior, but it did not achieve the expected effect:

KONG_NGINX_UPSTREAM_KEEPALIVE_TIME: 10s
KONG_UPSTREAM_KEEPALIVE_IDLE_TIMEOUT: 10
KONG_UPSTREAM_KEEPALIVE_MAX_REQUESTS: 10000

Expected Behavior

We expect Kong to:

• Respect a configurable maximum upstream connection lifetime

Actual Behavior

• Results in intermittent 502 errors (connection reset by peer)

Question

Is there a supported way in Kong to:

• Enforce maximum upstream connection lifetime
• Ensure upstream keepalive connections are invalidated when Kubernetes endpoints change
• Prevent reuse of connections to terminating pods during rolling updates

If not, what is the recommended approach to handle this scenario in Kubernetes environments?