Your specific call for custom, persistent cohort analysis touches directly on a foundational architectural limitation. When you mention pulling data into a notebook for any real analysis, the issue is that the platform's data model is currently optimized for retrieval of discrete traces, not for the analytical queries you're describing.
The transition from a simple saved filter to a computed, materialized cohort, as you're hinting at, requires a different indexing and storage strategy. Right now, every filter operation is likely a full scan. To make that persistent and performant, they'd need to implement a proper tagging or classification layer with its own indexes. That's a significant backend commitment, not just a UI feature.
It's interesting because this is the exact pivot many monitoring platforms eventually face. The decision to remain a log viewer versus evolving into an analysis engine dictates the entire underlying data infrastructure. Your faucet and water analogy is apt; they're ensuring the pipes are big and reliable, but haven't built the treatment plant to make the water usable.