AptifoldContact

Why a dashboard keeps you waiting

The Last One Back.

The dashboard requests stock counts from 20 warehouses at once. It can show the report once every reply arrives.

One stock report

Model time 0 ms
19 replies at 40 ms 1 reply at 800 ms
Final replyWarehouse 20

Due at 800 ms

Stock reportWaiting for replies

0 of 20 warehouses have replied

In this example, each warehouse replies in 40 or 800 milliseconds. The report is due by 200 ms. The timings are invented.

Set up the report

Chance the complete stock report is late

18.2%

About 18.2 in 100 complete reports would miss the 200 ms deadline.

One report needs a reply from every warehouse. Any reply arriving after 200 ms makes the whole report late.

Why this number?

Every reply has to be fast.

An all-fast report needs every independent call to finish quickly. One slow reply is enough to hold it up.

81.8%Every reply fast18.2%At least one slow

Probabilities in this model, not counts from a sampled workload.

One warehouse request1%
Complete report · 20 replies18.2%

Same scale: 0–100%. Change the warehouse count to compare.

How do the slow replies line up?

Chance the whole report is late

Each warehouse still has a 1% chance of a slow reply. Checking 20 warehouses gives a report 20 chances to encounter a slow reply; one is enough to hold up the whole report.

Each call: 1.0% slow40 ms when fast · 800 ms when slow

Change each call’s slow chance

These are authored probabilities for this model. No service is being measured.

One selected possible report

Will the stock report be ready?

0model ms

Select a warehouse to follow its reply. On the board, use Left/Right or Home/End.

Follow one warehouse replyThe report waits until all calls have returned.

Warehouse 20

Launch40 ms · fast200 ms · deadline800 ms

Playback runs at half speed. The clock shows model time.

All calls start together. Move the clock to see the gate work.

Same calls, different company

The per-call number doesn’t settle the report.

Knowing each call’s slow chance leaves a range of possible report risks. The way the replies coincide matters.

Same per-call chances allow1–20%of reports to miss the deadline

WHERE YOU CAN USE THIS

Measure the wait the person experiences.

The stock dashboard is one example of a job that needs several parallel results: a search page merging indexes, a dashboard reading several services, or an agent waiting for every required check. A quick individual call does not guarantee a quick complete result.

Two design questions follow. Does the person need every reply before they can continue? Do delays share a cause, such as an overloaded dependency? This bench changes the second question while keeping every reply required. Showing partial results, setting timeouts or changing that completion rule needs a different calculation.

What if a partial report is acceptable? Open Ready enough? to set a minimum reply coverage and a maximum lateness risk. Its separate brief finds the fullest qualifying rule and shows what an early snapshot omits. The settings on this complete-report bench stay here.

Behind the response room

The report waits for its slowest reply.

Each call has probability p of taking 800 ms. Otherwise it takes 40 ms. All begin together, and the report requires all n replies. Its completion time is therefore the largest call time. The 200 ms deadline separates the two outcomes.

Three ways to keep the same per-call chance

In all three models, each call has the same latency law and the same expected time. What changes is the joint law: which calls are slow in the same report. The expectation of the largest time can change even when the expectation of every individual time stays fixed.

Two calls, each slow half the time

This small exact example uses p=½. It is separate from the current controls. Each column adds to 100%, and each call is slow in exactly half its column.

Exact probabilities of the four ordered outcomes
Call 1 / call 2IndependentSharedSpread
Fast / fast25%50%0%
Fast / slow25%0%50%
Slow / fast25%0%50%
Slow / slow25%50%0%
Report past deadline75%50%100%

What the bounds say

A report that needs every reply must be slow whenever any one call is slow, so its risk is at least p. Adding the n slow-event probabilities gives the union bound: at most np, capped at one. Both ends can be attained by the declared shared and spread constructions.

This model does not identify a real system’s dependence or recommend synchronizing its failures. Changes to load, resources or background work can change the per-call probabilities too. Serial work, partial-result rules, timeouts, retries and hedging need different models. A percentile statement usually bounds a deadline exceedance probability; it does not supply the exact p used here.

Inspect this setup

Exact slow-count probabilities

Fractions are reduced integer ratios. Tiny positive probabilities remain possible even when a rounded percentage would look like zero.

Slow callsExact probabilityReport finishes

Source

Dean and Barroso’s The Tail at Scale (CACM,2013) explains how component outliers affect large fan-out services. The author-hosted paper, pp.76–78 also discusses concentrating disruptions and partial-result completion. This exhibit uses its own two-time example and original graphics; it reproduces no measured service table.

Probabilities are calculated with exact integer fractions. Independent implementation and small joint-state checks qualify the arithmetic; they do not validate an empirical workload or establish a new theorem.

APTIFOLD / AGENTIC AI

What would you like to work through?

Aptifold develops its own agentic AI to investigate questions, test ideas and build working software. This experience is one example of that work.

Explore Aptifold Discuss a question

Company links open in a new tab, so you can keep this experiment open.