Concerns about Medusa 2.x workflows when running large scale applications

[khanh-to-niteco]

Hi Medusa team,

We have workflows in many places in Medusa 2 core framework. This raises several questions:

1. Workflow compensation:

• Workflows seem to not roll back by default when there is an error, we have to write a compensation action for each step which increases development effort and more code to maintain.
• In order to be able to compensate for update actions, we need to load the previous version of the object before updating, this could have a little performance hit. If the system runs at scale, this issue would compound.
• Testing compensation actions tends to be more complicated which could increase testing effort.
• In case the compensation itself fails, maybe due to coding bugs or database unavailability, we may need to monitor and resolve the issue manually which increases operational effort.

2. Concurrency issues:

• With each step, we can use the same transaction to deal with concurrency issues, but there is no transaction across many steps to prevent concurrency issues across steps.

Example: cart workflow with inventory check. When we have only few items in inventory, and many requests to add items to cart run at the same time, all these requests could get confirmation that there is enough stock, but actually only the first request should get that confirmation.
This may not be a real issue in a particular application, but just to illustrate the point.

Workflows do have benefits when used in the right context:

• When we have to coordinate with external systems where transactions across systems do not work anyway. Workflows really shine in this case when the execution is persisted and can be rerun to retry in case it failed before.
• When business requirements in a particular case don't need strong consistency, and can accept some concurrency issues in exchange for more scalable applications.

But since workflows in Medusa 2 framework core don't have our business requirement context, it may not work well under load.

Please give your thoughts, thanks.