Skip to content

Retries for functions

OpenFaaS Pro offers an enterprise-ready queue-worker which retries failed invocations using an exponential back-off algorithm, supports multiple named queues, and provides detailed monitoring.

Note: This feature is included for OpenFaaS Standard & For Enterprises customers.

There are two primary use-cases for retrying asynchronous messages:

  • When a downstream API is likely to fail some of the time, and cannot be fixed easily, retrying can work around this problem
  • When max_inflight has been set on a function such as a web scraper, or memory-intensive function and you do not want to overload the function. When a function cannot accept any more connections due to the inflight setting, it will return a 429 error, which indicates the message can be retried at a later time.

Two types of retries

  1. Retries for errors from Core Components Whenever a Core Component of OpenFaaS emits a HTTP error status code along with a X-OpenFaaS-Internal header, then the queue worker will retry the request automatically.

  2. Retries for statuses from functions When a function emits a HTTP error status code, the queue-worker may retry that request, if the code is specified in the queue-worker's configuration, or via an annotation on the function.

In both cases, the initial retry delay, maximum retry attempts, and maximum wait between attempts are controlled through a priority order. If there is an annotation set on the function, then this will take precedence. If there is no annotation set on the function, then the default values set on the queue worker will be used.


You can install the Pro version of the queue worker by editing the values.yaml file of the OpenFaaS chart.

You must also set:

openfaasPro: true

  maxRetryAttempts: "10"
  maxRetryWait: "120s"
  initialRetryWait: "10s"
  httpRetryCodes: "429,502,500,504,408"

Timings are specified in Golang durations such as 10s or 1m.

  • initialRetryWait is the amount of time to wait for the first retry, this can be set 0s if desired, or a higher value.
  • maxRetryWait is the maximum amount of time to wait between retries.
  • maxRetryAttempts is the amount of times to try sending a message to a function before discarding it.
  • httpRetryCodes is a comma-separated list of HTTP status codes which the queue worker will retry when received from a function.

Configure retries per function

The retry configuration can be overridden on a per function basis using annotations on the function:

    image: alexellis2/chaos-fn:0.1.1
    skip_build: true
      com.openfaas.retry.attempts: "30" "429"
      com.openfaas.retry.min_wait: "5s"
      com.openfaas.retry.max_wait: "1m"
faas-cli store deploy chaos \
  --annotation com.openfaas.retry.attempts=30 \
  --annotation \
  --annotation com.openfaas.retry.min_wait=5s \
  --annotation com.openfaas.retry.max_wait=1m

If no annotation is specified, or if the value is invalid, then the default for the queue-worker will be used.

Random jitter for retries

The queue-worker supports adding random jitter into the exponential backoff delay. The aim of random jitter is to reduce competing requests to busy functions. In testing this produced better results for fast running functions, for very slow running functions it is less advantageous.

The jitter mode can be configured by setting queueWorkerPro.backoff in the values.yaml file for the OpenFaaS deployment.

  backoff: full

Supported backoff values:

  • exponential - no jitter is added to the random backoff delay (default).
  • full - wait between 0-100% of the exponential backoff delay.
  • equal - wait between 50-100% of the exponential backoff delay.

The feature was influenced by Marc Brooker's article on Exponential Backoff And Jitter.


To test the retry functionality, you can use our chaos function, which allows a function to be configured to return a canned response, or to timeout with a given duration.

We will configure a testing function to output a 429 HTTP response, which indicates that the function is too busy to process the request. This is one of the core use-cases for retrying an invocation. Once we've observed the retrying functionality, we will update the function to return a 200 response, then we'll see the request complete.

You can also add a long timeout or a different HTTP code, see the testing function's notes on GitHub: alexellis/chaos-fn

faas-cli store deploy chaos

Open a separate terminal to observe the retrying mechanism with the logs of the queue-worker.

kubectl logs deploy/queue-worker -n openfaas -f

Configure the chaos-fn to start failing with a 429 error code:

export CODE=429
curl -i --data-binary '
{   "status": '$CODE',
    "delay": "1s",
    "body": "1 second delay, then '$CODE'"}
' --header "Content-type: application/json"

You will always see a 202 Accepted from the /set path, if you see an error then check the format of your request. The status code must be an integer.

Invoke the function synchronously, to see that it's returning the expected error code:

curl -i

HTTP/1.1 429 Conflict
Content-Length: 24
Content-Type: text/plain; charset=utf-8
Date: Tue, 08 Mar 2022 11:37:01 GMT
X-Call-Id: 24ae517c-ad0d-4654-8160-d2a7f7d68fab
X-Duration-Seconds: 1.001162
X-Start-Time: 1646739420463153513

1 second delay, then 429

Now invoke the function asynchronously.

curl -i -d "input"

The invocation will fail, returning a 429 message. This will be received by the queue-worker which will then retry the request. Check the logs of the queue-worker in the other terminal to see this happening.

Allow it to fail several times, seeing the retry time back off exponentially.

For undelivered messages that reach maximum retries the response is posted to the callback url.

Then, whenever you like, fix the error by changing the canned HTTP response to "200 OK". This will allow the next retry to complete:

curl -i --data-binary '
{"status": 200,
    "delay": "1ms",
    "body": "1ms second delay, then 200"}
' --header "Content-type: application/json"

Would you like a demo?

Feel free to reach out to us for a demo or to ask any questions you may have.