You're asking exactly the right questions, but you're already outlining the failure scenario in your last bullet point. The hierarchical model *is* the bottleneck, because it's also the permission model. You'll spend more time debating folder taxonomy and access rights than you ever did on Jira's clunky workflows.
On API limitations, the deal-breaker isn't a missing endpoint, it's the random latency. A 30-second P99 on a status update webhook response means your CD pipeline is held hostage by their infrastructure's mood. You don't integrate with GitLab, you build a queuing system with dead-letter queues and retry logic that pretends to integrate.
And no, you can't migrate active sprints. You'll manually recreate them while your team works in two systems for a month, which defeats the entire purpose.
prove it to me
> you build a queuing system with dead-letter queues and retry logic that pretends to integrate.
This is the precise architectural tax people miss. Once your integration requires a persistent queue, you're no longer just handling webhooks; you're operating a stateful service with its own failure modes. We had to add monitoring for queue depth, consumer lag, and poison pill detection, all because the upstream API's SLA was essentially nonexistent.
The latency variance means your backoff logic can't be sane. If the P99 is 30 seconds but the median is 800ms, your exponential backoff either gives up too quickly on true outliers or wastes minutes retrying what is just a normal, albeit slow, response. You're forced to choose between dropping legitimate updates or creating unacceptable lag in your own pipeline.
--perf