Step Functions limitations

Where Overcast’s Step Functions behaves differently from AWS’s. Anything not listed here is meant to behave as AWS documents it; a difference that is not listed is a bug.

Integrations

An aws-sdk: Task reaches its service over the protocol that service speaks — AWS JSON, Query, EC2 Query, REST-JSON, REST-XML or Smithy RPC v2 CBOR — with AWS’s conventions: PascalCase parameter and result names, map keys left as written, timestamps as ISO-8601 strings, and errors named like S3.NoSuchKeyException or, for an error the service does not model, Ec2.Ec2Exception. A service Overcast does not implement is accepted by CreateStateMachine and fails the execution with States.Runtime.

Where it still differs:

• Two shared paths. AppSync’s Events API (/v2/apis) and AppConfig’s /applications actions reach API Gateway v2 and Service Catalog AppRegistry instead, because Overcast tells those services apart by the request’s signature, which an internal call does not carry.
• Blob parameters other than a payload (Body, Payload) are read as base64, and as plain text when they are not valid base64.
• Fields outside the action’s members are dropped when they come from the state input of a Task with no Parameters; a static Parameters field is rejected when the state machine is created, as on AWS.
• An error’s Cause ends with the wire error code, which AWS leaves out.

Query languages

JSONata is evaluated by a JSONata 2.2 engine; AWS runs JSONata 2.0.6. The whole 2.x function library and syntax behave as on AWS, AWS’s own functions ($partition, $range, $hash, $random, $uuid, $parse) are provided, and $eval fails the state with States.QueryEvaluationError, as AWS does not offer it. $now() and $millis() read Overcast’s clock. The ?? and ?: operators arrived in JSONata 2.1, so Overcast accepts them where AWS may not. Regular expressions use Go’s syntax, which has no lookahead, lookbehind or backreferences inside the pattern.

JSONPath filters support comparisons, &&, ||, ! and existence tests, not the regex (=~) or in/nin operators. A multi-name union ($['a','b']) returns an array of the values rather than an object. States.JsonToString sorts object keys, where AWS keeps the input’s order.

An unknown intrinsic function fails at run time with States.IntrinsicFailure; AWS rejects it when the state machine is created.

Map

Executions

A redrive resumes inside a failed Parallel or Map as AWS does: branches and iterations that succeeded keep their outputs and are not run again, and the rest resume at the state they stopped in. An inline Map works out its items again from the input it was entered with; if they now number differently — an Items expression using $random or $uuid, say — the whole Map runs again. A distributed Map’s child executions are redriven by redriving the parent; RedriveExecution on a child itself is refused with ExecutionNotRedrivable.

With a persistent store (OVERCAST_STATE other than memory), an execution survives an Overcast restart when it is parked where AWS itself waits on the outside world: a top-level Task waiting on its .waitForTaskToken token or on an activity worker, or a top-level Wait of a second or more. Its token answers SendTaskSuccess, SendTaskFailure and SendTaskHeartbeat again, an activity task no worker had taken is handed out by GetActivityTask again (one already taken stays with its worker), and the execution continues from that state with its history, Retry position, Catch and ResultPath intact. Heartbeat, TimeoutSeconds and Wait deadlines keep counting while Overcast is down, as they would on AWS. The first Step Functions request after the restart resumes them, so a deadline that passed while Overcast was down fires then.

An execution caught anywhere else is not resumed. That covers a synchronous Task in flight, a token waited on inside a Parallel branch or Map iteration, a distributed Map, an EXPRESS workflow, and the child a .sync parent is blocked on. Such an execution ends FAILED with States.Runtime and a cause naming the restart, and RedriveExecution runs the interrupted state again. If Overcast was killed rather than shut down, this happens the first time the execution is read afterwards. Only the in-memory store loses executions outright on a restart.

Versions, aliases and validation

GetExecutionHistory for a Standard execution caps at AWS’s 25,000 events, and an execution also stops at OVERCAST_STEPFUNCTIONS_EXECUTION_TIMEOUT (default 15 minutes) — see the landing page’s gotchas.

Related

• Step Functions
• Step Functions execution history
• Step Functions operations