Skip to main contentSkip to user menuSkip to navigation
Interactive design lab

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.

Latency analysis workbench

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

Dominant

P95 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

72%

Remote dependency

428 attempts/s demand · 700 attempts/s capacity

61%

Queue P99

2.7 ms

Attempt multiplier

1.02×

Failure after retries

0%

User-visible result

Latency envelope

P50
188 ms
P95
298 ms
P99
361 ms
P95 target450 ms
Effective timeout650 ms

Within the modeled envelope

P95 stays inside the 450 ms target with capacity reserve.

0% above target0% time out0% fail after retries

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.

No quiz questions available
Could not load questions file
Spotted an issue or have a better explanation? This page is open source.Edit on GitHub·Suggest an improvement