We currently do not have a good understanding of typical dwell times in the transaction system. We should consider adding metrics that give us more visibility into

• The distribution of durations of transactions.
• The amount of time transactions spend in the prepared state.
• The oldest/slowest unresolved transactions in the system (current and max over process lifetime).

We should also take time to consider if there are any additional metrics not described here that would help us better understand the performance and usage patterns for the transaction subsystem.