如何实现nginx 优雅的reload

[duancanran]

目前的版本reload nginx，都需要重启进程，对于线上环境就需要断流，影响很大。该如何实现不断流的nginx reload

[jfb8856606]

F-Stack nginx does not support graceful reload out of the box. This is explicitly documented in doc/F-Stack_Nginx_APP_Guide.md:

Command-line reload: the reload is not graceful, service will still be unavailable during the process of reloading.

Why it cannot be made graceful easily

In F-Stack's multi-process model, each worker binds exclusively to a NIC hardware queue (via RSS). During reload, when old workers exit and new workers have not yet initialized their DPDK/f-stack stack, there is an unavoidable window where no process owns the queue — packets are dropped.

This is fundamentally different from vanilla nginx, where workers share a kernel socket and the master can hand off connections between old and new workers transparently.

Recommended approach: community patch based on DPDK 18.11

The closest working solution comes from community member @orange30 (#547), who implemented a modified version that achieves reload without packet loss on f-stack-1.20 + DPDK 18.11. The key ideas:

1. Designate specific cores as dedicated receive cores (rcv cores) in f-stack.conf and nginx.conf — these cores keep running across reloads and never drop packets.
2. During the transition period (old workers exiting, new workers starting), dynamically adjust process priority via renice based on connection count to control traffic handoff.
3. Old and new workers coexist briefly; once the new worker is fully initialized and connections have drained from the old one, the old worker exits cleanly.

⚠️ Important limitation: this patch relies on the DPDK 18.x timer library, which used a static global variable for the timer state. Starting from DPDK 19.x, the timer library switched to rte_memzone_reserve for shared memory, breaking cross-process timer sharing. If you need to use this approach, you will need to downgrade to DPDK 18.11 or adapt the patch to the new timer API.

Other workarounds

• External load balancer: Place a layer-4 LB (e.g. DPVS/LVS) in front of your F-Stack nginx instances. During reload, drain traffic from the instance at the LB level before reloading, then restore. Zero packet loss at the cost of additional infrastructure.
• Blue-green deployment: Run two F-Stack nginx instances bound to different RSS queues. Use ethtool or DPDK flow rules to shift traffic between them during config updates.

Closing this issue as it duplicates #547. Please refer to that thread for the community patch details and discussion.