Application Performance Monitoring
also called APM
Instrumentation inside the application that attributes latency and errors to specific code paths, queries and dependencies.
Where infrastructure monitoring says a service is slow, APM says which endpoint, which database query, which external call and often which line. In practice most of the value comes from three views: slow transaction traces, the database query breakdown (which is where N+1 patterns become obvious), and error grouping by stack trace.
The architectural considerations: agents add overhead (usually a few percent, occasionally much more), auto-instrumentation covers common frameworks but misses custom code, and vendor agents are a form of lock-in that OpenTelemetry exists to remove — instrument once with an open standard, export to whichever backend you are currently paying.
Worth adding to the standard set: continuous profiling, which attributes CPU and memory to functions in production and answers questions APM traces cannot.