By the end of this chapter you can

• Say what an advanced planning and scheduling (APS) module does and where it sits between planning and the floor.
• Describe a resource with its calendar, capabilities, rates and efficiency factor.
• Compute the run minutes of an operation and tell finite from infinite capacity scheduling.

What the module does

Production scheduling turns demand and released orders into a plan that says which operation runs on which machine, line, crew or tool, and when. Vendors call software that does this advanced planning and scheduling (APS) ^[1][2]. This module replaces the APS tools and scheduling boards a business pays for today, from dedicated products such as PlanetTogether, Siemens Opcenter APS (Preactor), Asprova and DELMIA Ortems to the Smartsheet, monday.com or Excel boards many planners fall back on, by building the same core on your own platform.

The standard that frames the job is ISA-95 Part 3 (IEC 62264-3), which describes the detailed production scheduling activity and its information flows between planning above it and execution on the floor below ^[3]. In this module's boundary, material requirements calculation belongs to the module that plans materials (M12), and recording what actually happened belongs to the execution module (M04). Scheduling sits between them: orders in, a feasible plan out, and the floor's actual times coming back.

Resources, calendars and rates

Everything the scheduler places work on is a resource. The module library uses six types, and a resource usually points at a piece of equipment already in the spine.

Each resource has a calendar of shifts, and calendar exceptions that change it for a stretch of time: unavailable, planned maintenance and holiday take time away; overtime adds it. A resource also lists the capabilities it has (cut, mix, fill) with a rate per hour and a preference that says which resource to try first when several can do the same job. The efficiency factor scales the rate to what the resource really achieves: 0.8 means it delivers four fifths of its rated output. One resource can be marked a bottleneck, the one that limits what the whole area can make.

From quantity to minutes

The run time of an operation is the quantity divided by the effective rate, then turned into minutes. The seeded case used through this module has one machine with a rate of 60 an hour and an efficiency factor of 0.8, and four jobs of 96 units each.

Finite and infinite capacity

An infinite-capacity schedule places every order where its due date says and never checks whether the resource has room. Drawn as a Gantt chart (a bar chart with a row for each resource and time running left to right) it looks like a plan, but it can show two jobs on one machine at the same moment, or a day longer than the shift. A finite-capacity schedule never loads a resource beyond what its calendar holds ^[2][4]. When a job cannot fit, it says so and gives the reason, instead of drawing it anyway.

Scheduling can run two ways. Forward scheduling places each job as early as it can start, which answers "when can it finish?". Backward scheduling places each job as late as it can while still meeting the due date, which answers "when must it start?". When backward scheduling needs a start before the resource opens, the dates cannot all be met, and the job is marked infeasible.

Exercise · Run the infinite-capacity test on your own schedule20 minutes

You need: One week of your current schedule as the planners print or export it, and the shift times of one machine

Use a real week, and the machine the planner calls the busiest. Leave customer names out of anything you save.

1. 1. Pick one machine and one day from the week's schedule.
2. 2. Add up the minutes of work the schedule places on that machine that day.
3. 3. Work out the minutes the machine is open that day: its shift, less breaks and planned maintenance.
4. 4. Look for two jobs drawn at the same time on that machine, and for any job that runs past the end of the shift.
5. 5. Write down the load as work minutes divided by open minutes, and whether the tool warned about anything.
6. 6. Repeat for two more machines and write one line for each: finite, or infinite capacity drawn as a Gantt chart.

Outcome: A short table of three machines with work minutes, open minutes, any overlap, and a verdict on whether your current schedule is finite.

References

1. PlanetTogether: APS frequently asked questions. https://www.planettogether.com/faq
2. SYMESTIC: what is an APS system. https://www.symestic.com/en-us/what-is/aps-system
3. ISA: ISA-95 series of standards (enterprise-control system integration). https://www.isa.org/standards-and-publications/isa-standards/isa-95-standard
4. Fabrico: what is advanced planning and scheduling (APS). https://www.fabrico.io/blog/what-is-advanced-planning-and-scheduling-aps/

By the end of this chapter you can

• Apply earliest due date and shortest processing time to a small case and compare the lateness.
• Use a setup matrix to choose the order of jobs that minimises changeover time.
• Explain why locked operations must survive a re-run, and what the scheduler does when a lock cannot be honoured.

Dispatch rules

A dispatch rule is a sort order: when several jobs wait for one resource, which goes first? The production-planning body of knowledge the specification points to (the ASCM/APICS CPIM material) covers dispatching rules alongside lot sizing and capacity planning ^[5], and APS tools build their sequencing on rules like these ^[1]. The module implements four as sort keys.

No rule wins every measure. Take one resource and three jobs (the seeded case has exactly one capable resource, so the scheduler has no other machine to choose and the results below hold), all released at 08:00: P takes 3 hours and is due at 11:00, Q takes 1 hour and is due at 12:00, R takes 2 hours and is due at 13:00.

Run backward on the same case and the plan needs P to start at 07:00, an hour before the shift opens. The scheduler must mark P infeasible and say why. It is better to tell the planner the dates cannot all be met than to draw the chart anyway.

Setups make the order matter

On many resources the time to change over depends on what ran before. A paint line moving from light to dark loses little; dark to light needs a long clean. The module stores this in a setup matrix: for a resource and an attribute of the product (the seeded case uses a family, A or B), the minutes to change from one value to another, and optionally a cost. The first job on a resource needs no setup. The setup minutes are written on the operation when it is placed, so changing the matrix later never rewrites a published schedule.

Seed 1 has four jobs of 96 units on one machine (again the only capable resource), 120 minutes each, in a shift from 08:00 to 17:00. J1 and J3 are family A; J2 and J4 are family B. Run in the order they arrive, the machine changes over three times.

The sequencer tries each job as the first, builds the rest by always taking the job with the smallest setup from the one just placed (ties go to the earlier due date, then the lower job number), and keeps the order with the smallest total setup. If two starting jobs give the same total setup, it keeps the one with the lower job number: starting from J1 or J3 both give 30 minutes in Seed 1, so J1 wins. Then it records the result as scenario KPIs: total setup hours (0.5), makespan (8.5 hours, from the start of the first job to the end of the last), on-time share and utilisation.

Locked operations

Planners override the machine: a customer is waiting, a tool is only free at 10:00, a supervisor has promised a crew. If the next run of the scheduler moves those operations, the planner's work is lost and trust goes with it. An operation marked locked keeps its resource and times in every re-run. Everything else is planned around it. A lock that falls in a closed period, or overlaps another lock, is refused with the reason.

Two further properties matter. The scheduler must be deterministic: the same input gives the same plan, so that a difference between two runs always has a cause. And when the cases get hard, a constraint-programming solver (a program that searches for the best plan among all the plans that satisfy your rules) can be added as an option; OR-Tools CP-SAT, named in the specification, is a free, open-source one. Rules stay the default because they are fast and easy to explain. If the bottleneck is known, the Theory of Constraints option (drum-buffer-rope) schedules around it first and lets the rest follow.

Exercise · Predict the seeded cases before any code25 minutes

You need: Paper or a spreadsheet; this chapter's seeded cases

Do this before your agent writes a scheduler. Your predictions become the tests it must pass.

1. 1. Compute the run minutes of one Seed 1 job: 96 ÷ (60 × 0.8) × 60.
2. 2. Place the jobs in the order J1, J2, J3, J4 from 08:00 and add the setups from the matrix. Write the end time and total setup.
3. 3. Find the order of the four jobs with the smallest setup by trying each job as the first. Write its end time.
4. 4. Change A to B to 90 on paper, find the new best order, and write the setup and end time.
5. 5. Place the Seed 2 jobs P, Q and R in EDD order and then SPT order. Write each job's end time and its lateness.
6. 6. Lock Q at 10:00 to 11:00, place R and P around it by EDD, and write the new end times.

Outcome: A one-page prediction sheet with the run minutes, both Seed 1 orders with their setup and end times, the A-to-B-90 result, and Seed 2 under EDD, SPT and with Q locked.

References

1. Fabrico: what is advanced planning and scheduling (APS). https://www.fabrico.io/blog/what-is-advanced-planning-and-scheduling-aps/
2. PlanetTogether: APS frequently asked questions. https://www.planettogether.com/faq
3. SYMESTIC: what is an APS system. https://www.symestic.com/en-us/what-is/aps-system
4. ISA: ISA-95 series of standards (enterprise-control system integration). https://www.isa.org/standards-and-publications/isa-standards/isa-95-standard
5. ASCM: APICS CPIM, Certified in Planning and Inventory Management. https://www.ascm.org/learning-development/certifications-credentials/cpim/

By the end of this chapter you can

• Copy a scenario, change an input and compare the six scenario KPIs side by side.
• Say what publishing does in one transaction and why a published scenario cannot change.
• Name the module's tables, their groups and the spine tables they point at.

Scenarios

A scenario is a version of the schedule with its own horizon, jobs, constraints, objective weights and solver. Its status is draft, published or archived. APS tools commonly let planners compare what-if scenarios ^[1][2]. The planner copies a scenario to try a change (a different rule, an extra shift, a rush order), re-runs it, and compares. The copy records where it came from in created_from_id, and the original is not touched.

Compare shows six values per scenario, stored as scenario KPIs: on-time percentage, total lateness in hours, total setup in hours, utilisation percentage, makespan in hours and work in progress count. The scenario's objective weights say how much each of lateness, setup and utilisation counts when the sequencer chooses between plans.

Publishing

Publishing makes one scenario the plan the floor works to. It happens in a single transaction so that nothing is half done: the route first refuses a scenario with any resource loaded beyond capacity, then sets the status to published with the time and the person, archives the previously published scenario, writes the status history, and leaves two kinds of event in the outbox (the spine table where a module leaves events for other modules to pick up): aps.scenario.published, and one aps.job.late_predicted for each late job.

A published scenario is read-only. The database refuses any change to its jobs, operations, constraints or KPIs, and allows only one published scenario at a time. To change the plan, copy it, change the copy and publish that: the old plan stays on record. The operator's dispatch list is then simply the published scenario's operations for their area in start order, shown as one card each with the item, quantity, resource, planned start and end and any setup.

The data model

The module adds eleven tables with the prefix aps_, each with the standard columns every spine table has (id, tenant_id, created_at, created_by, updated_at, updated_by, row_version, archived_at, ext). It points at the spine for equipment, calendars, items, people and reason codes, and never keeps its own copy.

The database does the checking it can. Allowed values (a resource type, a job status, a constraint type) are check constraints; codes are unique per business; an exception cannot end before it starts. PostgreSQL's constraints and row-level security do this work for you.

Exercise · Check your tables against the specification20 minutes

You need: A SQL console on your branch database, the eleven aps_ tables, and the specification's column list

Use a local or branch database, never the live one. Your agent can write the statements; you read what they return.

1. 1. List the columns of each aps_ table with their types and compare them to the specification, noting any missing or extra column.
2. 2. List the check constraints on aps_resource, aps_job and aps_calendar_exception and compare the allowed values to the specification.
3. 3. Try to insert an aps_resource with a resource type that is not in the list, and confirm the database refuses it.
4. 4. Try to insert an aps_calendar_exception that ends before it starts, and a second aps_resource with an existing code. Confirm both are refused.
5. 5. Confirm every time column is a timestamp with time zone, and every aps_ table has tenant_id with row-level security switched on.
6. 6. Write one line for each refusal: the table, the rule and the message the database gave.

Outcome: A short list of schema differences (ideally none) and four refusals you saw the database make.

References

1. PlanetTogether: APS frequently asked questions. https://www.planettogether.com/faq
2. FitGap: PlanetTogether overview. https://us.fitgap.com/products/011934/planettogether

By the end of this chapter you can

• Explain why efficiency factors must come from measured OEE, not nameplate.
• Define schedule adherence, planned on-time delivery, changeover hours and bottleneck utilisation.
• Describe the adherence loop and name the three pitfalls the module is built to avoid.

Garbage capacity data, broken promises

Scheduling has no regulation of its own. The risk is credibility: if the capacity data is wrong, the schedule promises dates that cannot be kept, and the floor and the customers learn to ignore it. The main cause is scheduling on the rate printed on the machine's plate instead of the rate it achieves.

Overall equipment effectiveness (OEE) is availability (run time divided by planned production time) times performance (actual output rate divided by ideal rate) times quality (good units divided by all units), one of the KPIs ISO 22400-2 defines ^[1]. In this module the efficiency_factor on a resource is fed from measured OEE over the same planned time the calendar models. Planned maintenance is already an exception on the calendar, so it must not be counted twice. The factor records where it came from (the source, the period and the date) in the resource's ext field. If it has not been measured, say so and treat the resource as a known risk; do not guess.

The four schedule KPIs

ISO 22400-2 defines key performance indicators for manufacturing operations management, including schedule-related ones such as allocation and utilisation ^[1]. The module reports four every week and compares them between the old plan and the new.

Decide the tolerance in minutes that counts a start as on time before you collect any numbers, and write it down. Otherwise it will be chosen afterwards to make the result look right.

The adherence loop

When the schedule is published, each operator sees the operations for their area on a phone. Start and Finish are two large buttons; each tap records the time on the phone and is sent with a row id made on the phone. Start inserts the adherence row with that id and Finish updates the same row, so a repeated Start adds nothing and a repeated Finish returns success without a second write (the tap is idempotent: sending it twice has the same effect as once). The server, not the phone, computes how many minutes the start deviated from the plan. A supervisor gives a reason code to each start that missed the tolerance. If an execution module is installed, its completed-run events can feed the same table.

Where a resource consistently starts or finishes later than planned, compare its measured output with the factor you set, update the factor with a new source and period, and run again. A factor never changes without a recorded measurement.

Three pitfalls, and bringing data across

• Infinite capacity disguised as a Gantt chart: test it by finding an overload; a property-based test (hundreds of random cases checked against one rule) should find none.
• Nameplate instead of measured rates: check that every bottleneck resource has a measured source recorded.
• No locked operations: planner overrides vanish at the next run, so test that locked operations keep their place.

Incumbent tools rarely hold history worth migrating beyond adherence data. Bring over resources, routings and open orders from your ERP or MRP export, keep the old export as a record, and compare against it for a full week before you switch. Check the old tool's cancellation terms from your own account page, and cancel only after the cutover works and the export is safe.

Exercise · Replace one nameplate rate with a measured one25 minutes

You need: Planned time, run time, total count and good count for one bottleneck resource over a week, from your counters or a hand tally

Pick the resource that limits your area. If you have an execution system, take the figures from it; if not, a tally kept by the operator for one week is enough.

1. 1. Write down the resource's planned production time for the week, excluding planned maintenance.
2. 2. Write its run time, and compute availability as run time ÷ planned production time.
3. 3. Compute performance as the actual output rate ÷ the ideal rate, using total count ÷ run time.
4. 4. Compute quality as good count ÷ total count.
5. 5. Multiply the three to get OEE, and write it as the efficiency factor with its source, period and date.
6. 6. Recompute the run minutes of a typical order with the old and the new factor, and write the difference.

Outcome: One efficiency factor from measured data, its source recorded, and the minutes of difference it makes to a typical order.

References

1. ISO 22400-2:2014, Key performance indicators for manufacturing operations management: definitions and descriptions. https://www.iso.org/standard/54497.html
2. SYMESTIC: what is an APS system. https://www.symestic.com/en-us/what-is/aps-system
3. ISA: ISA-95 series of standards (enterprise-control system integration). https://www.isa.org/standards-and-publications/isa-standards/isa-95-standard

Choose one answer for each question, then submit. You will see the right answer and why for every question.

1. What does a finite-capacity scheduler never do?

ALoad a resource beyond what its calendar holdsBUse a dispatch ruleCSchedule backward from a due dateDRecord the plan in a database

2. A machine has a rate of 60 units an hour and an efficiency factor of 0.8. How long does an order of 96 units take?

A96 minutesB80 minutesC120 minutesD144 minutes

3. Which source should the efficiency factor of a bottleneck resource come from?

AThe machine's nameplateBThe vendor's brochureCA round number agreed in a meetingDMeasured OEE over the period, with the source recorded

4. Backward scheduling answers which question?

AWhen can it finish?BWhen must it start to meet the due date?CWhich machine is cheapest?DWho is on shift?

5. Three jobs wait for one machine: P takes 3 hours (due 11:00), Q takes 1 hour (due 12:00), R takes 2 hours (due 13:00), all released at 08:00. Under EDD, when does R finish and how late is it?

A12:00, on timeB13:00, on timeC14:00, 1 hour lateD14:00, 3 hours late

6. On the same three jobs, why does SPT give a worse total lateness than EDD?

AIt runs P last, and P finishes at 14:00 against its 11:00 due time, 3 hours lateBIt adds setup timeCIt ignores the release dateDIt leaves the machine idle

7. Why does a setup matrix store minutes for a pair of values rather than one number per product?

ATo save spaceBBecause the changeover time depends on what ran before, so A to B can differ from B to ACBecause operators prefer tablesDBecause setup is the same for every pair

8. Seed 1 has four 120-minute jobs (A, B, A, B), A to B 30 minutes and B to A 60, in a shift from 08:00 to 17:00. What does running them in the order A, B, A, B do?

AFinishes at 16:00BNeeds no setupCCosts 30 minutes of setupDNeeds 120 minutes of setup and ends at 18:00, after the shift closes

9. Why must a locked operation keep its place when the scheduler is re-run?

ALocks make the run fasterBThe database requires itCOperators cannot read a moving scheduleDOtherwise the planner's overrides are lost on every re-run

10. What is the rule for a published scenario?

AAnyone with access can edit itBIts jobs, operations, constraints and KPIs cannot change; copy it, change the copy and publish thatCIt is deleted when the next draft is madeDIt can be changed until midnight

11. Operators press Start and Finish on a phone with no signal, and the app sends the same tap twice. What is the result?

ATwo adherence rowsBThe server rejects bothCOne adherence row, because the row id made on the phone makes the repeat harmlessDThe phone's time is replaced by the server's

12. 80 operations were scheduled and 68 started within the tolerance. Which KPI is 85 percent?

ASchedule adherenceBPlanned on-time deliveryCBottleneck utilisationDChangeover hours

Submit answers