The mathematics is the same. The constraints are not. Six industries, six versions of the same problem — and what each one teaches us.
Scheduling is everywhere — but each industry has its own language for it, its own constraints that make it hard, and its own catastrophic failure mode when things go wrong. Click any industry to dive in.
Select an industry above — or click any tab below — to explore the constraints, scale, and failure modes that define scheduling in each domain.
Manufacturing scheduling is where the field was born — and it remains one of the hardest instances in practice. The classical Job Shop Problem (JSP) asks: given n jobs, each requiring a sequence of operations across m machines, find a schedule that minimises total completion time (makespan). Real factories layer on top: sequence-dependent setup times (cleaning a machine between products), maintenance windows, shift patterns, material dependencies, and multi-plant routing.
Bottleneck machines sit idle waiting for upstream operations while urgent jobs queue behind lower-priority work. A single late delivery cascades into penalty costs, customer complaints, and stock outs downstream. In automotive plants, a 1-hour unplanned stoppage can cost £50,000 or more.
Each row is a machine; each coloured block is one operation from a job. The solver must find an ordering that avoids machine conflicts while minimising the end of the last block.
Constraint Programming (CP-SAT) or MILP for medium instances. Metaheuristics (tabu search, genetic algorithms) for large plants where exact solvers time out. Rolling-horizon re-scheduling when new urgent jobs arrive.
Healthcare scheduling is uniquely human — the constraints aren't just mathematical, they're ethical. Nurse rostering must balance legal requirements (the Working Time Directive caps duty at 48 hrs/week with minimum 11-hour rests), clinical safety (minimum nurse-to-patient ratios, right skill mix), and fairness (equitable weekend load, respecting preferences). Operating room (OR) scheduling must sequence surgeries across limited theatres while accounting for surgeon availability, equipment sterilisation, and unpredictable overruns.
Understaffed night shifts create patient safety risks. Over-reliance on agency staff costs 3–4× base salary. Unfair rosters drive burnout and resignations — a particular crisis in post-pandemic healthcare systems where every nurse retained matters.
The solver must ensure every shift has enough cover, no nurse works consecutive night+day shifts, and weekends are distributed fairly across the 4-13 week horizon.
ILP or CP-SAT with a soft/hard constraint hierarchy. Hard constraints encode legal requirements; soft constraints encode preferences and fairness objectives. Commercial solutions include Optima Nursing, Allocate, and custom CP-SAT models.
Aviation scheduling is among the most complex operational planning problems in existence — and one of the most consequential. It happens in three linked stages: fleet assignment (which aircraft type flies each route), crew pairing (building multi-day trip sequences that start and end at a crew base, respecting FAA/EASA duty-time rules), and crew rostering (assigning pairings to individual crew members for the monthly period). A single scheduling error can propagate through an entire network — grounding aircraft, stranding crew, and triggering millions in recovery costs.
A single delayed flight cascades: crew go "out of position," aircraft sit at the wrong airport, and the disruption ripples through the network for hours. Southwest Airlines' December 2022 meltdown — partly attributed to scheduling system failures — cost over $800M and stranded 2 million passengers.
Set-partitioning formulation solved with column generation + branch-and-price. MILP cores are standard. Commercial platforms: Sabre AirCentre, Jeppesen, Carmen. Real-time disruption recovery uses heuristics — exact methods are too slow for the 20-minute decision window.
Logistics scheduling is fundamentally the Vehicle Routing Problem (VRP) — assign a set of deliveries to a fleet of vehicles and find the optimal route for each vehicle. In practice, the "base" VRP is almost never enough. Most real problems are VRPTW (with time windows), VRPC (with vehicle capacity), or both. Add multiple depots, mixed fleet types, driver shift hours, real-time traffic, and same-day delivery windows — and you have the problem that drives billions in annual software investment. Last-mile delivery alone accounts for 53% of total shipping costs.
Drivers return over capacity or run out of time mid-route. Deliveries miss time windows, triggering failed-delivery fees and redelivery costs. Poor routing inflates fuel spend and driver overtime. At scale, even a 1% routing improvement translates to millions in annual savings.
Three vehicles, each with a capacity limit and shift-end deadline, collectively cover 9 stops. The solver must partition stops across vehicles and sequence each route to minimise total distance while respecting time windows.
OR-Tools' routing library (Google), NVIDIA cuOpt for GPU acceleration at scale, or commercial platforms like Routific and OptimoRoute. Large instances use LNS (Large Neighbourhood Search) or guided local search — exact methods don't scale past ~30 stops.
Workforce scheduling in retail and service industries is demand-driven: the number of staff needed at any hour depends on forecasted footfall, call volume, or throughput — and that forecast changes weekly. The problem is to build a set of shifts, assign employees to them, and ensure coverage meets demand at every time interval, while honouring labour law (maximum weekly hours, minimum rest, break entitlements), union agreements, and individual preferences. Unlike manufacturing, the "product" is service — and the cost of failure is invisible until a customer gives up and leaves.
Over-scheduling during quiet periods burns payroll unnecessarily. Under-staffing during peaks creates queues, abandoned carts, and poor customer experience scores. Unfair schedules drive turnover — and in low-margin retail, replacing a trained employee costs 50–200% of their annual salary.
ILP for medium instances; decomposition (first build shifts, then assign) for large ones. Commercial platforms dominate: Kronos (UKG), Deputy, Rotageek, When I Work. Many use ML to forecast demand, then OR to solve the scheduling problem.
Project scheduling formalises as the Resource-Constrained Project Scheduling Problem (RCPSP): given a set of activities with precedence relations and resource requirements, find a start time for each activity that minimises the project makespan (or meets a deadline), subject to resource availability. Real projects add: activity durations that are uncertain, multi-mode activities (you can allocate more resources to speed up), time-varying resource availability, multi-project shared resource pools, and the ever-present scope change.
Critical path activities slip, compressing buffers downstream until deadlines are missed. Resource conflicts stall multiple workstreams simultaneously. The CHAOS Report consistently shows ~70% of software projects delivered late or over budget — with poor scheduling cited as a leading cause.
Critical Path Method (CPM) and PERT for deterministic baselines. CP-SAT or MILP for resource-constrained variants. Monte Carlo simulation for stochastic duration modelling. Tools: Microsoft Project, Primavera P6, or custom Python + OR-Tools models.
Two fully documented, award-winning deployments — UPS in last-mile logistics and Netherlands Railways in crew scheduling — that show what happens when scheduling theory meets operational scale, institutional resistance, and the messy constraints of real infrastructure.
UPS's On-Road Integrated Optimization and Navigation system is arguably the largest-scale VRP deployment in history. It re-routes 55,000 US drivers every morning using a blend of network optimisation, machine learning, and constraint satisfaction — in time for the first shift.
UPS operates ~55,000 drivers making 16 million deliveries per day across the US. Before ORION, routes were planned by experienced dispatchers using local knowledge and fixed templates. The question wasn't whether routes could be improved — it was whether you could prove it, systematically, and deploy it at scale across every UPS terminal simultaneously.
Domain knowledge still wins. UPS's engineers spent years encoding knowledge that experienced dispatchers carry implicitly — right-turn biases, local access restrictions, building-specific delivery norms. The solver provided optimality guarantees; the domain experts provided the constraints that made those guarantees meaningful. Ten years on, ORION continues to evolve, integrating real-time data and dynamically re-optimising routes mid-day.
In the early 2000s, NS faced a crisis: a new timetable had created an unworkable crew scheduling problem that planners could no longer solve by hand. What followed became one of the most celebrated OR deployments in history — a complete re-architecture of how Dutch railway operations are planned, from timetable to crew.
NS operates roughly 400,000 train kilometres per day across the Dutch network, staffed by around 2,800 train drivers and 3,000 conductors. When NS introduced a new demand-driven timetable in 2007, the resulting crew scheduling problem broke every existing manual and semi-automated planning process. The timetable created complex transfer patterns — crew had to be repositioned across the network constantly — and the scale made the combinatorial explosion intractable with prior methods.
Crucially, NS treated timetabling, rolling stock scheduling, and crew scheduling not as three independent problems solved sequentially, but as a single integrated planning challenge. Each stage's output fed the next, and the whole pipeline was redesigned around OR from the ground up.
The biggest insight from NS wasn't a solver breakthrough — it was an architectural one. Previous railway planning systems solved timetabling, rolling stock, and crew scheduling separately, in sequence, with each handoff introducing infeasibilities the next stage had to absorb. NS's team, working with researchers at CWI and Erasmus University Rotterdam, proved that integrating all three stages — even imperfectly — produced solutions that were fundamentally better than any sequential approach could reach. The timetable that emerged was one no human planner had imagined, because it was shaped from the start by what was actually feasible for crews and rolling stock to execute.
The published paper — Kroon et al. (2009), Interfaces — remains one of the clearest accounts of what it takes to move OR from theory into live national infrastructure.
The same five dimensions across all six industries. The patterns that emerge reveal why some domains converged on CP-SAT, others on metaheuristics, and a few on decomposition methods that mix both.
| Industry | Problem Family | Optimisation Goal | Primary Solver Approach | Typical Horizon | Failure Cost |
|---|---|---|---|---|---|
| 🏭 Manufacturing | Job Shop / Flow Shop | Makespan, tardiness | CP-SAT MILP Tabu search | Shift / day | £50k+/hr stoppage |
| 🏥 Healthcare | Nurse rostering / OR sched. | Fairness, cost, coverage | CP-SAT ILP | 4–13 weeks | Patient safety, burnout |
| ✈️ Aviation | Crew pairing / rostering | Cost, regulatory compliance | Column gen. B&P Heuristics (recovery) | Monthly cycle | $M per disruption |
| 🚚 Logistics | VRP / VRPTW | Distance, time, fuel | LNS Guided LS OR-Tools / cuOpt | Daily | Late fees, redelivery cost |
| 🛒 Retail / Workforce | Shift planning / rostering | Coverage, cost, fairness | ILP Decomposition Commercial WFM | 4–8 weeks | Service failures, turnover |
| 📐 Project Mgmt. | RCPSP / CPM | Makespan, budget | CP-SAT MILP Monte Carlo | Weeks–years | Deadline penalties, overruns |
Every domain above is a variation of the same core problem — allocate limited resources to tasks across time. What makes each hard is the constraint structure: aviation has strict regulatory cliffs (violate rest rules and the whole schedule is invalid), logistics has combinatorial route explosion, healthcare has soft fairness constraints that are genuinely hard to formalise. The solver choice follows from the constraint structure: CP-SAT excels where constraints are rich and interlocking; metaheuristics dominate where scale defeats exact methods; decomposition wins where the problem has natural sub-problems that can be solved independently and then composed.
Organised by domain — industry reports, Edelman Award deployments, vendor documentation, and regulator publications. Tags indicate the type of source.