Interpreter for Cadence/Temporal

Workflow Engine

Quanzheng Long

Cadence/Temporal is HARD to learn!

• It’s powerful, but really complicated
• Knowing how to write a sample is far from feeling comfortable to write and maintain production code

Proposal: better interface on top of C/T

• Keep almost SAME POWER
• Hide/Abstract the hard concepts:
• Separation of Workflow vs Activity
• Replay & non-determinism & versioning
• Weird exception handling
• Workflow threads
• Sticky Cache
• TestEnv
• …

Comparison summary

• Separation of Workflow vs Activity
• Unit test
• Replay & non-determinism & Versioning
• Other interface contract
• Workflow threads
• Operation

​

Potential benefits:

• Every dev will become experts of using Cadence easily
• Deliver same products >2 times faster than using raw Cadence
• Much easier to maintain

Separation of Workflow vs Activity

Cadence workflow:

• Schedule timer/activities/etc
• Deciding flows/parallel

Cadence activity:

• Communicate with externals

​

iWF framework:

• No distinction of workflow vs activity.
• No restriction in iWF workflow code

​

* Exception for long running activity

Replay & non-determinism & Versioning

Cadence workflow:

• Must be replay-safe
• Must use versioning if not safe
• Must use replay test to safeguard the determinism
• Code gets really messy after several iterations

​

iWF framework:

• None!!!

​

Unit tests

Cadence workflow:

• Super hard to write & maintain
• TestEnv is really tricky
• Super hard to debug when things go wrong
• Error messages are not obvious to fix the issues. It requires you to understand the internal details

​

iWF framework:

• Only need the well-known unit test library:
• E.g. For Java only Junit is enough!

​

Other interface contract

Implicit contracts in Cadence:

• Signal cannot return data or throw back exception
• Signal method is in a different thread
• Query cannot modify states
• Regular exception will block workflow
• Checked exception is prohibited
• Exception from activity is wrapped

​

iWF framework:

• All is explicitly defined in APIs
• Uses standard exception handling models

​

Workflow threads

Cadence multi-threading:

• Super complicated to understand
• Consuming lots of resources in Java SDK, not easily scale up(each instance can only run thousands of threads)

​

iWF framework:

• Super straightforward
• Super efficient using lightweight threading, support millions of threads per instance

​

Operation

Cadence operation

• Lots of complicated metrics to monitor/understand that is not related to business. Especially for stickiness

​

iWF framework:

• Operation is like operating a microservice (in fact it is).
• No stickiness concept exposed

​

Setup projects using Cadence/Temporal

Cadence workflow:

• Need to configure for every project that Start/Query/Signal workflows

​

iWF framework:

• For workflow workers, or any project to Start/Signal/Query workflow, only need a URL of iWF server. Nothing else

​

​

Other benefits of iWF

• No need to do continueAsNew. Potentially, iWF can do this underneath
• Super easy to build SDKs for different languages
• Easy to migrate between Cadence/Temporal/Others

​

iWF cons & counter arguments

• Programming model change.
• In raw Cadence, the workflow is programmed as a single function. In iWF, the workflow is required to modeled into explicit states of different Java classes.
• In certain use cases, iWF may lose some “readibility” after dividing into multiple state files.
• “Normal looking” workflow code is a “trap” in Cadence/Temporal.
• Counter argument
• In most cases, users usually don’t need too many states if workflow is designed correctly.
• Multiple states will get some visibility like this if we build some visualization tool in IDE for iWF states which is not possible to build for Cadence workflow code.
• In practice, forcing design in multiple states make it easier for multiple developers working in parallel
• This iWF programming model change is very similar to AWS SWF-> AWS Step Functions. Step Functions has become very successful.

iWF vs C/T vs SWF vs Step Functions