Latency Analysis Workbench
What is end-to-end latency?
End-to-end latency is the time a user waits while a request crosses network, application, queue, database, and dependency work. The critical path matters: serial stages add, while parallel branches are gated by the slower branch.
Step 1
Model
Set the P95 target, stage budgets, median service times, execution shape, arrival rate, worker capacity, dependency capacity, failure rate, retries, and timeout.
Step 2
Observe
Inspect modeled P50, P95, and P99 user waits, the critical path, queue pressure, retry amplification, timeout rate, and successful-request rate.
Step 3
Challenge
Compare serial and parallel work, then inject a slow dependency, queue buildup, retry amplification, timeout squeeze, or partial failure. Use production traces and load tests to replace the illustrative assumptions.
Trace the request, not an average
Allocate a latency budget, compose serial or parallel work, then pressure the same path with queueing, retries, deadlines, and dependency failures.
P50 observed
188 ms
Median user wait
P95 observed
298 ms
Inside target
P99 observed
361 ms
Includes waits capped by timeout
Successful requests
100%
0% timed out
Challenge mode
Inject pressure into the same request path
Selected scenarios override only the values they name.
Feedback loop 1
Allocate budget, then shape each stage
Path allocation 320 ms
130 ms unallocated
Edge and network
Modeled P95 72 ms
Application service
Modeled P95 93 ms
Database
Modeled P95 114 ms
Remote dependency
Modeled P95 176 ms
Request topology
Critical path under parallel execution
Database and dependency work start together. The slower branch gates the response; their latencies are not added.
Edge
P95 72 ms
P99 98 ms · budget 70 ms
Application
P95 93 ms
P99 126 ms · budget 90 ms
Database
P95 114 ms
P99 154 ms · budget 110 ms
Dependency
DominantP95 176 ms
P99 238 ms · budget 160 ms
Response
298 ms P95
Edge → Application → Dependency branch
Feedback loop 2
See queueing and retry pressure accumulate
Worker pool
69.3 concurrent demand · 96 slots capacity
Remote dependency
428 attempts/s demand · 700 attempts/s capacity
Queue P99
2.7 ms
Attempt multiplier
1.02×
Failure after retries
0%
User-visible result
Latency envelope
Within the modeled envelope
P95 stays inside the 450 ms target with capacity reserve.
Applied scenario
Planned peak
- No scenario override
- Configured values remain unchanged
Modeling assumptions
- 1,600 deterministic request samples; values are illustrative, not production telemetry.
- Each stage is log-normal and calibrated from its P50 and selected P99 multiplier.
- Parallel branches use the slower sampled branch; percentiles are measured after composition.
- Queue delay is a bounded capacity approximation; observed waits stop at the client timeout.
- Timeout and final-failure rates can overlap; a request succeeds only when neither occurs.
Use real traces to calibrate stage distributions and correlation before making a production capacity decision. This workbench teaches path composition and pressure behavior; it does not replace load testing.