A faster coding step does not necessarily get a feature to users sooner. A bottleneck is a necessary stage in a value stream where demand persistently exceeds capacity, limiting the output of the whole stream. It is not simply the step that takes longest to perform. Look for a growing queue, longer waits, aging work items, or a low activity ratio. section 6.4 section 6.1.2
The pull-request queue
Six developers open about 15 pull requests each week, but reviewers close about 10. Five more requests join the queue each week, so lead time rises even if coding or deployment gets faster. Increasing coding capacity feeds the queue; improving review capacity addresses the constraint. Open branches, context switching, and merge conflicts can grow while completed value does not. section 6.4
Map the work end to end before choosing an improvement. A bug fix can take three hours to code but wait three days for triage, review, and deployment. In another stream, technical deployment takes about 30 minutes while an approval gate takes a day. Cutting two minutes from a build will barely affect that lead time. Work-in-progress limits can make such waits visible on a pull-request dashboard, CI queue, or ticket board. section 6.1.2 section 6.3
Table 11.5 Checkout request latency breakdown by service
|
Service
|
Avg response (normal)
|
Avg response (peak)
|
Change
|
|
Restaurant Service
|
45ms
|
380ms
|
8.4x
|
|
Order Service
|
60ms
|
520ms
|
8.7x
|
|
Driver Service
|
30ms
|
290ms
|
9.7x
|
|
Payment Service
|
200ms
|
210ms
|
1.05x
|
Compare normal and peak service response times to locate a shared constraint: the three services using the shared database slow sharply, while the payment service barely changes. section 11.4
The same test applies to running software. If requests through a shared database slow under peak load while a service that bypasses it remains steady, investigate the database rather than tuning every service. Identify the constraint, get more from its existing capacity, reduce other components’ demands on it, and only then add capacity. Recheck afterward: the limiting stage may have moved. Optimize the stream, not whichever step is easiest to speed up. section 11.4 section 11.10

