[Feature Request] Simplify proper usage of AsyncLocalStorage in Workflow context

[mjameswh]

Description

AsyncLocalStorage are frequently used inside of Workflow context, for example to implement correlation/tracing/context propagation mechanisms. However, most users don't know (or forget) that ALS must be "disabled" when they are no longer required (i.e. when the workflow execution gets evicted from cache). Otherwise, they do retain heap and still get called for every single AsyncResource (e.g. mostly Promise in this context) creation.

Given that we already control access to the AsyncLocalStorage in Workflow context, it would be easy for the SDK to track instantiation of AsyncLocalStorage, and automatically dispose them when the execution context is about to be destroyed.

We also need to investigate either usage of ALS inside of Workflow Context is efficient. Since ALS hooks work at the Worker Thread level, each ALS instance “tags” (i.e. adds a [kResourceStore] property on) every single AsyncResource that gets created in that Worker Thread, without regard to which VM the ALS or the AsyncResource was created in.

This could possibly mean that, even when ALS get disposed properly, using ALS inside workflows may currently add a memory and performance penalty that grows exponentially with the number of currently cached workflows. Based on memory snapshots provided by a customer, the constant multiplier here would be relatively small (48 bytes per “tag” in >99% of cases), and therefore, this problem may not be immediately apparent, but could still have sizable impact on Workflow Worker where maxCachedWorkflow is high.

Again, given that we already control access to the AsyncLocalStorage in Workflow context, it would be possible for the SDK to provide a custom implementation of ALS that is workflow-aware and/or that scales better.