How do you diagnose and improve NoSQL query performance?

Start with the execution plan and the actual metrics: rows or documents examined, time spent, partitions contacted and bytes returned. A slow query is usually reading far more data than it returns or waiting on a hot partition.

db.orders.find({ customerId: "cus_7" })
  .sort({ createdAt: -1 })
  .limit(20)
  .explain("executionStats");

const plan = db.orders.find({ customerId: "cus_7" })
  .sort({ createdAt: -1 })
  .limit(20)
  .explain("queryPlanner");

The usual fixes

• Add or fix an index so the filter and sort are supported
• Project only the needed fields to reduce network and memory
• Paginate with a stable cursor instead of large offsets
• Avoid scans and unbounded ranges on large collections
• Cache repeated reads when staleness is acceptable
• Fix the key design if one partition receives most of the traffic

In distributed stores, latency is often dominated by the slowest replica or by cross-partition fan-out. Measure p99, not just the average, and separate query time from queueing and network time. A query that is fast on average can still be unacceptable for a small fraction of users.