The original job
The Last One Back / The completion desk
Ready enough?
To find a warehouse that can fill an order, staff can start with replies already in. An early report leaves each missing warehouse’s stock unknown. Wait for all replies when you need the company-wide total.
Set the share of warehouses that must reply and how often the report can miss its deadline.
Adjust either limit to find a rule for this comparison.
The calculation is loading. Controls and setup links become available when it is ready. If they remain unavailable, reload this page; your linked setup stays in the address.
01 / Set your two limits
What is enough?
At least 16 of 20 warehouse replies. We never lower this limit to find an answer.
No more than 5 in 100 reports past 200 ms, under the chosen assumption.
Use the worst risk allowed by each warehouse’s slow chance. This protects against an unknown pattern within this model.
The warehouse brief
Invented inputs. Every request starts together and takes either 40 ms or 800 ms. “Late” means after 200 ms. No service is being measured.
02 / Fullest qualifying rule
Require 17 replies.
85% required coverage. A late risk of at most 5%, even with an unknown slowdown pattern.
Requiring 18 raises that upper risk to 6.67%, above your limit.
03 / Same replies, two completion rules
When can each report leave?
One deliberately chosen configuration, not a sampled workload. No frequency is assigned while the pattern is unknown.
The rule you are inspecting
Require 20 replies
Inside the inspected report
- Replies used
- Missing at issue
A report includes every reply available when it first becomes ready. Its contents stay fixed after issue. Later arrivals are not a live update.
04 / Inspect the tradeoff
Require more. Wait on more.
Inspect any rule. The marked answer is the greatest required reply count meeting both of your limits.
· Below your coverage floor× Too much lateness risk✓ Meets both limits◆ Fullest qualifying rule
20 of 20 required
How much does the slowdown assumption change the answer?
These are alternative mathematical descriptions of the same per-warehouse slow chance. They are not recommendations to synchronize real failures. A real system would need evidence for its chosen description.
| Slowdown assumption | Required replies | Risk used |
|---|
Where this matters
Speed has an information cost.
A search page might show results from the indexes that have replied. An operational dashboard might show which warehouse updates are missing. An agent might continue after enough optional checks finish.
That can help only when the early output is useful and its omissions are acceptable. Counts alone do not tell you whether the missing warehouse holds the item you need, or whether a missing check is critical. Those jobs need a different definition of coverage.
Exact calculation, assumptions and saved evidence
When is a rule late?
If a rule requires q of n replies, it misses 200 ms when at least n − q + 1 replies take 800 ms. Write that slow-count threshold as k. With per-warehouse slow chance p, the expected slow count is np.
When the pattern is unknown, the sharp lower and upper risks are max(0, (np − k + 1) / q) and min(1, np / k). The upper end, rather than an average of named scenarios, decides admission. These are bounds on possible dependence laws with the specified marginals, not a confidence interval or uncertainty about an estimated p.
The independent law uses binomial slow counts. The shared law puts all replies in the fast or slow group together. The spread law mixes the two nearest integers to np, choosing the slow subset uniformly. An unknown pattern can have a higher risk than all three.
Every integer rule from 1 to n is checked. The minimum count is the coverage percentage times n, rounded upward. The greatest count satisfying that floor and the inclusive risk cap is returned. A faster rule with less required coverage is not the objective.
Exact risk fractions for every rule
| Required | Risk used | Coverage | Within cap |
|---|
Scope: simultaneous starts, identical per-call slow probabilities, two arrival times and interchangeable replies. No retries, timeouts, streaming, serial work, changing load or empirical service guarantees. Exact arithmetic is checked against an independent rational calculation. This is a decision exercise using established probability, not a new theorem.
The original exhibit was informed by Dean and Barroso’s The Tail at Scale. This extension uses an original two-time example. Return to the complete-report investigation.