Find your next learning step
Some queues react to their own length. Staff are added, processors speed up, customers balk, or retries arrive later. Once the queue changes the rates that govern it, capacity is endogenous and averages alone stop telling the whole story.
Open the chapter contents · Queueing Theory root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Separate cause from response
Route 2 · Work the threshold model
Route 3 · Read feedback and switching
Route 4 · Test evidence and limits
Route 5 · Check and continue
CHAPTER 1 OF 12 · SEPARATE CAUSE FROM RESPONSE
1. A queue can change the system that serves it
In a basic queue, arrivals occur at a rate λ and service completes at a rate μ fixed outside the model. Many real systems respond instead. A visible backlog may call in another worker, activate a faster processor, change triage rules or persuade some customers not to join. The queue length is then both an outcome and an input.
That distinction matters because an observed association is not automatically a cause. Busy periods may coincide with high capacity because managers respond to congestion, not because capacity created the congestion. A defensible model writes the direction explicitly: the current state n selects λ(n), μ(n), or a separate controller state. Only then can we ask whether the feedback stabilises the system, merely delays overload, or creates costly switching.
| Model | What is fixed externally? | What responds to the queue? |
|---|---|---|
| Exogenous M/M/1 | λ and μ | Nothing |
| State-dependent service | λ | μ(n) |
| Balking or admission control | μ | λ(n) |
| Controlled environment | transition rules and available actions | staffing or speed state |
CHAPTER 2 OF 12 · SEPARATE CAUSE FROM RESPONSE
2. Endogenous capacity needs a larger state
Suppose the server runs slowly below a threshold and quickly above it. If speed changes instantly with n, the queue length alone can be a Markov state. If activation takes time, the state must include both the queue and the current capacity mode, such as (n, slow) or (n, fast). Two systems with the same visible queue can then have different near-term departure rates.
The same principle applies when arrivals respond. A long line may reduce new entries now, yet create retries later. Counting only the present line misses the orbit of deferred jobs. Endogenous does not mean mysterious; it means the rates must be conditional on a state partly created by earlier events.
This also separates control policy from physical capacity. A machine capable of nine jobs per hour may be scheduled at five until a threshold is crossed. The model should record the applied rate, the switching rule and any delay or cost, not replace them with one blended number.
CHAPTER 3 OF 12 · WORK THE THRESHOLD MODEL
3. Work a threshold-service birth–death queue
Consider Poisson arrivals at λ = 6 jobs per hour. One server completes at μ(n) = 5 jobs per hour when 1 ≤ n ≤ 3 and μ(n) = 9 when n ≥ 4. The low mode alone is overloaded, but the high mode has spare capacity. For a birth–death chain,
πn = π0 × ∏ from k = 1 to n of λ(k−1)/μ(k).
The first three ratios are 6/5 = 1.2. From state 4 onward the ratio is 6/9 = 2/3. Using π0 as one unnormalised unit gives weights 1, 1.2, 1.44 and 1.728 for states 0 to 3. The entire tail has weight 1.728 × (2/3)/(1−2/3) = 3.456. Total weight is 8.824, so π0 ≈ 0.1133 and the probability of operating at n ≥ 4 is 3.456/8.824 ≈ 0.392.
| Quantity | Result | Interpretation |
|---|---|---|
| Tail ratio | 2/3 | High-mode drift is downward |
| P(n ≥ 4) | 0.392 | Fast capacity is active about 39% of steady-state time |
| Mean number E[N] | 30/8.824 ≈ 3.40 | The queue spends material time near and above the trigger |
| Long-run throughput | 6 per hour | In a stable lossless queue, completions balance admitted arrivals |
CHAPTER 4 OF 12 · WORK THE THRESHOLD MODEL
4. Why one average service rate can mislead
It is tempting to average 5 and 9 and declare a service rate of 7. That ignores how often each rate is used, and those usage fractions are created by the queue itself. In the worked model, the time-average applied capacity is not a simple 50–50 mixture. It equals 5 while the system is in states 1 to 3, 9 in states 4 and above, and zero completed work when empty.
Even the average applied capacity is not a complete waiting-time model. The order of slow and fast periods matters when switching is delayed. Costs also depend on what is being averaged: scheduled capacity, potential completions and actual completions are different quantities. Unused fast capacity after the backlog clears does not create negative waiting.
The safe sequence is to model the state-dependent rates, solve or simulate the joint state, and only then calculate the operational average needed for the decision.
CHAPTER 5 OF 12 · WORK THE THRESHOLD MODEL
5. Changed context: weaken the fast mode
Keep λ = 6 and the same threshold, but reduce the fast service rate from 9 to 7. The tail ratio becomes 6/7 ≈ 0.857. A stationary distribution still exists, yet the tail now decays much more slowly, so long queues become substantially more common. If the fast rate falls to 6, the tail ratio is one and the stationary weights no longer sum to a finite total. Matching arrivals only after congestion is not enough for positive recurrence in this infinite-buffer model.
This is the transfer check: do not ask only whether the controller turns capacity on. Ask whether the activated regime creates strict negative drift far into the tail. For a more general state-dependent birth–death queue, inspect the products of λ(n)/μ(n+1), not a single headline utilisation.
The conclusion has a limit. A finite buffer always has a finite state space, but it may achieve that mathematical stability by rejecting work. Blocking probability and admitted throughput then become essential evidence.
CHAPTER 6 OF 12 · READ FEEDBACK AND SWITCHING
6. Feedback can reduce demand now and create demand later
Queue-dependent arrivals may fall when customers see a long line. That is negative feedback if the lost demand disappears. If customers retry, refresh a page or rejoin another channel, part of the demand moves through time instead. A short visible queue can coexist with a large hidden orbit of deferred attempts.
Positive feedback is also possible. Slow responses can trigger status checks, duplicate tickets or escalations, increasing effective arrival pressure. The original request and the later attempts are not independent jobs if they represent the same need. Measure unique work, attempts and admitted work separately.
The model therefore needs a behavioural boundary. State whether λ(n) counts first attempts, all attempts, or accepted entries. Without that definition, a fitted fall or rise in arrivals can be interpreted in opposite ways.
CHAPTER 7 OF 12 · READ FEEDBACK AND SWITCHING
7. Switching delays and hysteresis change the path
Instant threshold switching assumes that capacity moves from 5 to 9 as soon as n reaches 4 and returns immediately when n falls. Real staffing, autoscaling and equipment warm-up take time. During the delay, the queue can continue to grow. A controller state is then required, and the pair (queue length, capacity mode) determines the next rates.
Hysteresis uses two thresholds: switch up at n = 6, for example, and switch down at n = 2. This can prevent rapid on–off chatter near one boundary. It also means the same queue length can be served differently depending on its recent path. At n = 4, a system descending from a peak may remain fast while a system rising from quiet may remain slow.
That memory is operationally useful but invalidates a one-dimensional state description. Record activation and release rules, delays, minimum run times and capacity failures before treating queue length as sufficient.
CHAPTER 8 OF 12 · READ FEEDBACK AND SWITCHING
8. Control objectives are not identical to short waits
A policy that always selects maximum speed may reduce waiting but consume more energy, staff attention or wear. A policy that protects cost may tolerate a longer queue. Dynamic service-rate control makes this trade-off explicit through holding costs, service costs and available actions.
The optimal action can often rise with congestion, but that is a model result under stated assumptions, not a universal law. Markov-modulated demand, abandonment, setup costs and information available to the controller can alter the threshold or make a simple heuristic poor in some regimes.
| Decision question | Evidence to retain | Common shortcut to avoid |
|---|---|---|
| When did capacity change? | controller log and effective timestamp | infer it only from departures |
| Why did demand change? | unique jobs, attempts, balks and retries | treat every attempt as new demand |
| Did the policy stabilise the tail? | long-run drift and tail occupancy | compare only daily averages |
| Was switching worth it? | wait, loss, cost and switch count | optimise one metric silently |
CHAPTER 9 OF 12 · TEST EVIDENCE AND LIMITS
9. What the model does not establish automatically
State dependence does not prove deliberate feedback. Both the queue and capacity may respond to a third variable such as time of day, weather or job complexity. Departures are also censored by the available backlog: a quiet completion count cannot reveal potential service speed when the server is starved.
Stationarity is another limit. A steady-state calculation assumes that the rules and environment persist long enough for long-run proportions to be meaningful. A one-hour peak with a ten-minute activation delay is a transient problem even if an infinite-horizon controlled chain is stable.
Finally, a threshold model compresses individual jobs. If service requirements vary with queue length because easy jobs are prioritised, the apparent rate change mixes capacity with case selection. Retain job class and service requirement evidence before giving the controller causal credit.
CHAPTER 10 OF 12 · TEST EVIDENCE AND LIMITS
10. A practical feedback-queue checklist
Start with event-level arrivals, service starts, completions, abandonments, rejections and retry identifiers. Add capacity-mode changes with requested and effective times. Confirm whether an arrival observes the pre-arrival or post-arrival queue and whether the threshold uses jobs waiting or jobs in system.
Estimate conditional rates only where exposure is adequate. Ten minutes spent at a rare queue length cannot support a precise hourly rate. Plot run lengths and transitions around each threshold, then compare an instantaneous model with a delayed or hysteretic alternative. Test the far-tail drift under plausible parameter uncertainty.
For a practical evidence check, reconstruct the balance equation over each interval: ending work = starting work + admitted arrivals − completions − abandonments. Unexplained residuals usually reveal timestamp, identifier or boundary problems before they reveal new queueing theory.
CHAPTER 11 OF 12 · CHECK AND CONTINUE
11. Questions readers should ask
Can a queue be stable if its low service rate is below arrivals?
Yes. The worked infinite-buffer example is stable because sufficiently large queues activate a high regime with μ greater than λ, making the far-tail drift negative. Stability does not imply short waits or low control cost.
Is the average of the slow and fast rates the effective service rate?
Not without the endogenous state fractions, empty time and switching path. The queue determines how often each mode is active.
Does a threshold require a Markov model?
No. Thresholds can be studied through simulation, renewal models or data-driven control. A Markov chain is useful only when the chosen state makes the next-step law adequately memoryless.
Why keep a separate controller state?
Because delay, hysteresis or minimum run time lets identical queue lengths have different current capacity and different next-event probabilities.
CHAPTER 12 OF 12 · CHECK AND CONTINUE
12. Continue through the verified queueing ecosystem
Begin with the Queueing Theory root, the Gold Standard of Education and the How X Works hub. Compare exogenous environmental modulation in Markov-Modulated Service, Joint Background States and Arrival–Service Dependence with delayed operational response in Time-Varying Arrivals and Capacity That Arrives Too Late.
For demand feedback, continue to Retries, Attempt Rates and Unique Jobs and Finite Buffers, Blocking and Admitted Throughput. For measurement limits, use Service-Time Second Moments and Mean Wait. Authoritative technical reading: Dynamic Service Rate Control for a Single Server Queue with Markov Modulated Arrivals develops controlled service rates under modulated demand, while Markovian Queues with Poisson Control shows why queue length and controller-selected speed may need to be modelled jointly.
Top · Contents · Previous · How X Works