A RetryPolicyM is a function that takes an RetryStatus and possibly returns a delay in microseconds. Iteration numbers start at zero and increase by one on each retry. A *Nothing* return value from the function implies we have reached the retry limit.

Please note that RetryPolicyM is a Monoid. You can collapse multiple strategies into one using mappend or <>. The semantics of this combination are as follows:

1. If either policy returns Nothing, the combined policy returns Nothing. This can be used to inhibit after a number of retries, for example.
2. If both policies return a delay, the larger delay will be used. This is quite natural when combining multiple policies to achieve a certain effect.

Example:

One can easily define an exponential backoff policy with a limited number of retries:

> limitedBackoff = exponentialBackoff 50 <> limitRetries 5

Naturally, mempty will retry immediately (delay 0) for an unlimited number of retries, forming the identity for the Monoid.

The default retry policy retryPolicyDefault implements a constant 50ms delay, up to 5 times:

> retryPolicyDefault = constantDelay 50000 <> limitRetries 5

For anything more complex, just define your own RetryPolicyM:

> myPolicy = retryPolicy $ \ rs -> if rsIterNumber n > 10 then Just 1000 else Just 10000

Since 0.7.

Simplified RetryPolicyM without any use of the monadic context in determining policy. Mostly maintains backwards compatitibility with type signatures pre-0.7.

Helper for making simplified policies that don't use the monadic context.

Default retry policy

Applies a natural transformation to a policy to run a RetryPolicy meant for the monad m in the monad n provided a transformation from m to n is available. A common case is if you have a pure policy, RetryPolicyM Identity and want to use it to govern an IO computation you could write:

  purePolicyInIO :: RetryPolicyM Identity -> RetryPolicyM IO
  purePolicyInIO = natTransformRetryPolicy (pure . runIdentity)

How to handle a failed action.

Convert a boolean answer to the question "Should we retry?" into a RetryAction.

Datatype with stats about retries made thus far. The constructor is deliberately not exported to make additional fields easier to add in a backward-compatible manner. To read or modify fields in RetryStatus, use the accessors or lenses below. Note that if you don't want to use lenses, the exported field names can be used for updates:

> retryStatus { rsIterNumber = newIterNumber }
> retryStatus & rsIterNumberL .~ newIterNumber

Initial, default retry status. Exported mostly to allow user code to test their handlers and retry policies. Use fields or lenses to update.

Apply policy on status to see what the decision would be. Nothing implies no retry, Just returns updated status.

Apply policy and delay by its amount if it results in a retry. Return updated status.

Retry combinator for actions that don't raise exceptions, but signal in their type the outcome has failed. Examples are the Maybe, Either and EitherT monads.

Let's write a function that always fails and watch this combinator retry it 5 additional times following the initial run:

>>> import Data.Maybe
>>> let f _ = putStrLn "Running action" >> return Nothing
>>> retrying retryPolicyDefault (const $ return . isNothing) f
Running action
Running action
Running action
Running action
Running action
Running action
Nothing

Note how the latest failing result is returned after all retries have been exhausted.

Same as retrying, but with the ability to override the delay of the retry policy based on information obtained after initiation.

For example, if the action to run is a HTTP request that turns out to fail with a status code 429 ("too many requests"), the response may contain a "Retry-After" HTTP header which specifies the number of seconds the client should wait until performing the next request. This function allows overriding the delay calculated by the given retry policy with the delay extracted from this header value.

In other words, given an arbitrary RetryPolicyM rp, the following invocation will always delay by 1000 microseconds:

retryingDynamic rp (\_ _ -> return $ ConsultPolicyOverrideDelay 1000) f

Note that a RetryPolicys decision to not perform a retry cannot be overridden. Ie. when to stop retrying is always decided by the retry policy, regardless of the returned RetryAction value.

Run an action and recover from a raised exception by potentially retrying the action a number of times. Note that if you're going to use a handler for SomeException, you should add explicit cases *earlier* in the list of handlers to reject AsyncException and SomeAsyncException, as catching these can cause thread and program hangs. recoverAll already does this for you so if you just plan on catching SomeException, you may as well ues recoverAll

The difference between this and recovering is the same as the difference between retryingDynamic and retrying.

A version of recovering that tries to run the action only a single time. The control will return immediately upon both success and failure. Useful for implementing retry logic in distributed queues and similar external-interfacing systems.

Retry ALL exceptions that may be raised. To be used with caution; this matches the exception on SomeException. Note that this handler explicitly does not handle AsyncException nor SomeAsyncException (for versions of base >= 4.7). It is not a good idea to catch async exceptions as it can result in hanging threads and programs. Note that if you just throw an exception to this thread that does not descend from SomeException, recoverAll will not catch it.

See how the action below is run once and retried 5 more times before finally failing for good:

>>> let f _ = putStrLn "Running action" >> error "this is an error"
>>> recoverAll retryPolicyDefault f
Running action
Running action
Running action
Running action
Running action
Running action
*** Exception: this is an error

List of pre-made handlers that will skip retries on AsyncException and SomeAsyncException. Append your handlers to this list as a convenient way to make sure you're not catching async exceptions like user interrupt.

Helper function for constructing handler functions of the form required by recovering.

For use with logRetries.

Implement a constant delay with unlimited retries.

Grow delay exponentially each iteration. Each delay will increase by a factor of two.

FullJitter exponential backoff as explained in AWS Architecture Blog article.

http://www.awsarchitectureblog.com/2015/03/backoff.html

temp = min(cap, base * 2 ** attempt)

sleep = temp / 2 + random_between(0, temp / 2)

Implement Fibonacci backoff.

Retry immediately, but only up to n times.

Add an upperbound to a policy such that once the given time-delay amount *per try* has been reached or exceeded, the policy will stop retrying and fail. If you need to stop retrying once *cumulative* delay reaches a time-delay amount, use limitRetriesByCumulativeDelay

Add an upperbound to a policy such that once the cumulative delay over all retries has reached or exceeded the given limit, the policy will stop retrying and fail.

Set a time-upperbound for any delays that may be directed by the given policy. This function does not terminate the retrying. The policy `capDelay maxDelay (exponentialBackoff n)` will never stop retrying. It will reach a state where it retries forever with a delay of maxDelay between each one. To get termination you need to use one of the limitRetries function variants.

Run given policy up to N iterations and gather results. In the pair, the Int is the iteration number and the Maybe Int is the delay in microseconds.

Run given policy up to N iterations and pretty print results on the console.