justinmchase

AWS Step Functions has some built-in features for catching and handling errors but, surprisingly, it doesn’t have semantics for the usually accompanying “finally” concept.

In my scenario I am creating an ephemeral Kinesis stream in my State Machine, which I then stream a large number of records into while executing one lambda. I then process those records more slowly in a series of subsequent lambda functions. Once completed I then delete the ephemeral kinesis stream.

The problem with this approach is that if there is an unexpected error anywhere in one of my steps it can cause the whole Step Function to fail and end up orphaning the kinesis stream. Therefore I needed a way to reduce the likelihood of this problem with a try/finally pattern.

To accomplish this, first imagine we have this step function:

StartAt: ConfigureIterator
States:
  ConfigureIterator:
    Type: Pass
    Result:
      limit: 500
    ResultPath: $.iterator
    Next: InitializeIterator
  InitializeIterator:
    Type: Task
    Resource: iterator
    InputPath: $.iterator
    ResultPath: $.iterator
    Next: ConfigureXmlStream
  ConfigureXmlStream:
    Type: Pass
    Result:
      gz: true
      root: item
    ResultPath: $.options
    Next: XmlStream
  XmlStream:
    Type: Task
    Resource: xmlstream
    ResultPath: $.xml
    Next: SendItemsToApi
  SendItemsToApi:
    Type: Task
    Resource: items2api
    ResultPath: $.iterator
    Next: IterateNext
  IterateNext:
    Type: Choice
    Choices:
      - Variable: $.iterator.state
        StringEquals: done
        Next: Cleanup
    Default: SendItemsToApi
  Cleanup:
    Type: Pass
    Result: done
    ResultPath: $.iterator.state
    Next: IteratorDone
  IteratorDone:
    Type: Task
    Resource: iterator
    InputPath: $.iterator
    ResultPath: $.iterator
    Next: Finally
  Done:
    Type: Pass
    End: true

In the InitializeIterator step we are creating our ephemeral Kinesis stream. In the XmlStream step we are streaming items from a large xml document into JSON objects which are then written to the stream. Next, in the SendItemsToApi we are reading items out of the kinesis stream, doing some formatting and validation on those items, and then sending each item to a REST endpoint for storage and/or other actions. Finally in the IteratorDone step we are destroying the Kinesis stream.

You could imagine a variety of other possible scenarios where one would need to cleanup resources allocated in a previous Step. In this particular scenario we need to ensure that the IteratorDone step is called regardless of errors that may happen between it and the InitializeIterator step.

To do this we first will wrap then XmlStream and SendItemsToApi steps in a Parallel block with a single branch. The reason we want to do this is so that these steps can be treated like a single block where any errors in any state can be caught and handled in a single Catch clause.

The three steps wrapped in a Parallel block now look like this:

  Main:
    Type: Parallel
    Branches:
      - StartAt: XmlStream
        States:
          XmlStream:
            Type: Task
            Resource: xmlstream
            ResultPath: $.xml
            Next: SendItemsToApi
          SendItemsToApi:
            Type: Task
            Resource: items2api
            ResultPath: $.iterator
            Next: IterateNext
          IterateNext:
            Type: Choice
            Choices:
              - Variable: $.iterator.state
                StringEquals: done
                Next: Cleanup
            Default: SendItemsToApi
    Next: Cleanup
    ResultPath: $.main
    Retry:
      - ErrorEquals: [ 'States.ALL' ]
        MaxAttempts: 3
    Catch:
      - ErrorEquals: [ 'States.ALL' ]
        ResultPath: $.error
        Next: Cleanup

It’s important to note here that the result of the block is an array of results where each index in the array is the result object from the last step of each branch. So in this case we will have an array with a single object in it [ { iterator: ... } ]. If you don’t specify a ResultPath it will replace the entire context object $, which is undesirable in this case since we need to still access the iterator object in a later step.

It’s also important to note that we are storing the caught exception into the $.error field, which we will rethrow later, after cleanup.

  Cleanup:
    Type: Pass
    Result: done
    ResultPath: $.iterator.state
    Next: IteratorDone

  IteratorDone:
    Type: Task
    Resource: iterator
    InputPath: $.iterator
    ResultPath: $.iterator
    Next: Finally

  Finally:
    Type: Task
    Resource: throwOnError
    Next: Done

  Done:
    Type: Pass
    End: true

So now if an error occurs while processing our xml file or sending items to the api it will retry a couple of times and then ultimately capture the error and move to the Cleanup ​phase. We’ve added a new Finally Step, which will throw an exception if there is a value stored in $.error, which will allow the Step Function to complete in an Error state rather than a Success state so we can further trigger alarms through Cloud Watch.

Here is the code for the throwOnError lambda:

import { log, parse, handler } from 'mya-input-shared'

function RehydratedError (message, name, stack) {
  const tmp = Error.apply(this, arguments)
  this.name = tmp.name = name
  this.message = tmp.message = message
  Object.defineProperty(this, 'stack', {
    get: () => [`${this.name}: ${this.message}`].concat(stack).join('\n    at ')
  })
  return this
}

RehydratedError.prototype = Object.create(Error.prototype, {
  constructor: {
    value: RehydratedError,
    writable: true,
    configurable: true
  }
})

export const throwOnError = handler((event, context, callback) => {
  const { feed, error } = event
  if (error) {
    const Cause = error.Cause || '{}'
    parse(Cause, (err, cause) => {
      if (err) return callback(err)
      const { errorMessage, errorType, stackTrace } = cause
      err = new RehydratedError(
        errorMessage || 'An unknown error occurred.',
        errorType || 'UnknownError',
        stackTrace || '')
      log.error('feed_error', err, { feed }, callback)
    })
  } else {
    callback(null, event)
  }
})