Performance issues with dynamic/static multi process management

[rguennichi]

Recently I faced issues with significant latency and delays during periods of scaling or when new requests arrive after periods of processes idleness.

Environment:

• Unit version: 1.31.1 (Alpine 3.18 - php82)
• Application type: PHP (Symfony)
• Traffic: ~5K requests/minute
• Resources: 5 pods - 1 core / 512mb memory per instance

Details:

1. Dynamic Process Configuration:

First I configured Unit to dynamically manage processes (scaling up from idle), I noticed considerable latency in processing incoming requests. This is particularly evident after deployments, processes scaling operations or when new requests hit a pod with a process that has been idle. The process latency manifests as delays and lags in request processing, leading to sporadic timeout errors (more than 3 seconds).

    "applications": {
      "symfony": {
        "type": "php",
        "processes":{
          "max":10,
          "spare":5,
          "idle_timeout":500
        },
        "limits": {
          "timeout": 10,
          "requests": 10000
        },
        "targets": {
          "direct": {
            "root": "/app/public/"
          },
  
          "index": {
            "root": "/app/public/",
            "script": "index.php"
          }
        }
      }
    }

2. Static Process Configuration:

Adjusting the configuration to use a static number of processes reduces the frequency of these issues but does not eliminate them. The problem persists, particularly during deployment and scaling operations, where the initial requests to a newly started pod experience significant delays.

  "applications": {
    "symfony": {
      "type": "php",
      "processes": 10,
      "limits": {
        "timeout": 10
      },
      "targets": {
        "direct": {
          "root": "/app/public/"
        },

        "index": {
          "root": "/app/public/",
          "script": "index.php"
        }
      }
    }
  }

3. Single Process Per Pod:

Configuring it to run only one process per pod seems to resolve these latency issues. However, this approach is really a resource-intensive especially considering amount of pods to create / memory consumption I need to allocate per pod.

Steps to Reproduce:

1. Deploy a a very simple PHP application on Docker/Kubernetes with nginx Unit, configured for dynamic process management (2 or more)
2. Simulate traffic (10 req/sec)
3. Observe the latency in processing the initial batch of requests after running the container
4. Stop traffic (and wait till you reach the processes.idle_timeout) and try again

Please let me know if you need more details or I miss something in the configuration.

[tippexs]

Thanks for reporting this to us. We are currently working on some enhancements to our PHP implementation and I am more than happy to put this one to the list of things we are looking into.

Would you mind sharing the Dockerfile you are using?

Unit version: 1.31.1 (Alpine 3.18 - php82)