process instance events lose task info when tasks are deleted

[jasquat]

Originally found in #1909 (comment)

When we delete tasks from the database such as during process instance migration and reset to task, the task information can no longer appear with its corresponding event. This means that the events table on the process instance show page only shows "task_completed" without showing what task it is.

A couple potential fixes include:

• delete events when we delete related task
  • events are ideally static though so this feels wrong
• cache the task info on the event table
  • feels wasteful due to the duplication
• revamp how we store tasks and events to avoid this problem

[essweine]

If we consider revamping how tasks and events are stored, we should consider relying on the event stream that @jbirddog and I are working on. If we considered this as more of a data warehouse problem than a running process problem, we could get them out of the db entirely:

• they'll be duplicated there for historical reasons anyway
• application db can focus solely on holding the current state, which would ultimately make things simpler

It would require more stuff being redirected to the event store than our initial effort, but we've talked about potentially doing that anyway.

[burnettk]

another thing we were brainstorming today, like instead of getting process_instance_event out of the app, we could kill task and rely on a persisted version of the event stream (always insert, never update) for transactional purposes. it's definitely convenient to have one record per thing with a task guid, but it seems possible that we could write a query that would work to put things together to hand to spiff lib.

i like your idea of killing the existing process_instance_event stuff better, but it's good to have options. :D

[jbirddog]

If we consider revamping how tasks and events are stored, we should consider relying on the event stream that @jbirddog and I are working on. If we considered this as more of a data warehouse problem than a running process problem, we could get them out of the db entirely:

* they'll be duplicated there for historical reasons anyway

* application db can focus solely on holding the current state, which would ultimately make things simpler

It would require more stuff being redirected to the event store than our initial effort, but we've talked about potentially doing that anyway.

This sounds like a nice approach/first use case for querying the event stream server. If the backend can't find the data then it could ask the event stream for historical data. If it can't be found there we can fall back to what we do today. Over time more of these "non current state" types of queries could shift to leveraging the historical data from the event stream server.