Skip to main content

Metrics for Scaling Workers

Workers are responsible for executing business logic in the workflow applications. Scaling and performance tuning workers are dependent on the following metrics:

  1. No. of pending requests in the task queue
  2. Throughput of a single worker
  3. Total number of worker processes running

Conductor servers publish metrics that allow you to monitor the health of your workers. Here is a guide to using these metrics: PromQL

tip

Each of the below metrics contains taskType as a tag that can be used to monitor the metrics for a specific task.

Pending requests (Gauge)​

max(task_queue_depth{taskType=})

How to use the metric:

  1. The goal should be to keep the queue depth constant (it may not be zero if the tasks are long running)
  2. Configure alerts and autoscaling policies for your workers based on the increase/decrease in the queue depth in a given time period.

No. of tasks completed per second (Counter)​

task_completed_seconds_count is published as a counter with taskType as a tag.

rate(task_completed_seconds_count{taskType=}[$__rate_interval])

How to use the metric:

  1. The metric shows the throughput. The goal is to keep the throughput at a threshold depending on the application's needs.
  2. Configure alerts and autoscaling policies for your workers based on the increase/decrease in the throughput.

Amount of time the task was in the queue (Timer)​

This metric measures the time the task was sitting in the queue before getting picked up by a worker.

max(task_queue_wait_time_seconds{quantile=, taskType=})

How to use the metric:

If the value is very large (more than a few seconds), check the following:

  1. No. of workers and if they are busy processing the tasks. If so, increase the no. of workers.
  2. Check the polling interval for the worker and reduce the polling interval if needed.
    note

    Reducing the polling interval could increase the no of API requests to the server triggering system limits.