Automatic http retries for intermittent failures, with exponential backoff, based on https://github.com/cenkalti/backoff.
The reason for a separate library, is that this library handles http status codes to know whether to retry or not. HTTP codes in range 500-599, and code 429 are retried. Connection failures are also retried. Status codes 400-499 (except for 429) are considered permanent errors and are not retried.
The Retry function performs the http request and retries if temporary errors
occur. It takes a single parameter as its input - a function to perform the
http request. This function must return (resp *http.Response, tempError error, permError error)
where tempError
must be non-nil if a temporary error occurs
(e.g. dropped connection), and permError
must be non-nil if an error occurs
that does not warrant retrying the request (e.g. badly formed url).
For example, the following code that is not using retries:
res, err := http.Get("http://www.google.com/robots.txt")
can be rewritten as:
res, attempts, err := httpbackoff.Retry(func() (*http.Response, error, error) {
resp, err := http.Get("http://www.google.com/robots.txt")
// assume all errors are temporary
return resp, err, nil
})
Please note the additional return value attempts
is an int
specifying how
many http calls were made (i.e. = 1 if no retries, otherwise > 1).
The go http package has 9 functions that return (*http.Reponse, error)
that
can potentially be retried:
- http://golang.org/pkg/net/http/#Client.Do
- http://golang.org/pkg/net/http/#Client.Get
- http://golang.org/pkg/net/http/#Client.Head
- http://golang.org/pkg/net/http/#Client.Post
- http://golang.org/pkg/net/http/#Client.PostForm
- http://golang.org/pkg/net/http/#Get
- http://golang.org/pkg/net/http/#Head
- http://golang.org/pkg/net/http/#Post
- http://golang.org/pkg/net/http/#PostForm
To simplify using these functions, 9 utility functions have been written that wrap these. Therefore you can simplify this example above further with:
res, _, err := httpbackoff.Get("http://www.google.com/robots.txt")
To use cusom back off settings (for example, in testing, you might want to fail quickly), instead of calling the package functions, you can call methods of HTTPRetryClient with the same name:
package main
import (
"log"
"net/http/httputil"
"time"
"github.com/cenkalti/backoff"
"github.com/taskcluster/httpbackoff"
)
func main() {
// Note, you only need to create a client if you want to customise
// the back off settings. In this example, we want to, but if you
// wish to use the reasonable default settings, no need to do this.
retryClient := httpbackoff.Client{
BackOffSettings: &backoff.ExponentialBackOff{
InitialInterval: 1 * time.Millisecond,
RandomizationFactor: 0.2,
Multiplier: 1.2,
MaxInterval: 5 * time.Millisecond,
MaxElapsedTime: 20 * time.Millisecond,
Clock: backoff.SystemClock,
},
}
res, _, err := retryClient.Get("http://www.google.com/robots.txt")
if err != nil {
log.Fatalf("%s", err)
}
data, err := httputil.DumpResponse(res, true)
if err != nil {
log.Fatalf("%s", err)
}
log.Print(string(data))
}
The package has tests, which run in travis. See http://travis-ci.org/taskcluster/httpbackoff.
As far as I am aware, there is nothing in this library that prevents it from being used concurrently. Please let me know if you find any problems!
Contributions are welcome. Please fork, and issue a Pull Request back with an explanation of your changes. Also please include tests for any functional changes.