Distributed order management and the split shipment nobody priced
Nov 6, 2026, 10 min read

Line-by-line node selection is what manufactures split shipments, because each line is assigned to its own nearest node with no term for the parcel the order has already opened. Parcel pricing is mostly a fixed base per shipment, so a second parcel on a three pound order can add eighty per cent to the outbound cost while the sourcing rule records a saving. The correction is an opportunity cost per unit per node, estimated from the expected future shipping cost of depleting that node, which is the dual value of the fulfilment linear programme. Xu, Allgor and Graves showed in 2009 that holding the decision for a short window and re-evaluating it cuts splits without moving the promise.
An order comes in with two lines. A pair of boots and a pair of socks. The boots ship from the east coast DC, the socks ship from a store in Ohio that happened to be four miles closer to the customer, and the customer receives two parcels on two days for a total order value of ninety. Nobody in the business made that decision. A rule engine made it twice, once per line, and each of the two decisions was locally correct against the objective it was given.
The distributed order management layer is usually bought to make sourcing decisions and configured to make lookups. The lookups are fast, auditable and easy to explain to an operations team, and they systematically produce two outcomes that cost more than the sourcing rules save: parcels that should have been one parcel, and inventory stranded at nodes that then have to serve nothing.
Where the second parcel comes from
Almost every rules-based sourcing configuration works line by line. Take the line, filter the nodes that have available stock, sort by a priority list or by distance, take the top one. Repeat for the next line. The configuration has no term for the fact that a parcel is already being opened at one of those nodes.
Jasin and Sinha, in Operations Research in 2015, formalised exactly this failure for multi-item e-commerce orders and showed that assigning items independently is what drives excess splits, then gave an LP-based correlated rounding scheme that assigns the items of one order jointly. The mechanism matters more than the algorithm: the moment you decide the items of an order together, the second item's cost depends on where the first one went, and greedy per-line assignment cannot represent that.
The size of the penalty comes from how parcel carriers price. A rate card is mostly a base charge for the first unit of weight in a zone, plus a small increment per additional unit. Say the base is 7.90 and each additional pound adds 0.55. A single three pound parcel costs 7.90 plus two increments, so 9.00. The same goods as a one pound parcel and a two pound parcel cost 7.90 plus 8.45, so 16.35. The split adds 7.35 on a shipment that should have cost 9.00, which is over eighty per cent, and it does that in exchange for a zone saving on the second line that was worth well under a pound.
Every one of those splits is recorded by the sourcing engine as a successful nearest-node assignment. The cost lands in a freight account that nobody reconciles back to routing decisions, which is why split rate is often the single most useful number a fulfilment team can start reporting and often the one they do not have.
The cost function the rules engine is actually minimising
Ask what a sourcing configuration optimises and the answer is usually distance, or a hand-maintained node priority list, or a mixture where distance breaks ties inside priority bands. Neither of those quantities is a cost.
Distance is a poor proxy for parcel cost because zones are step functions. Two nodes forty miles apart can sit in the same zone, in which case the distance preference between them is worth nothing at all, or they can straddle a boundary where twenty miles is worth a pound and a half. A distance-sorted list treats those two cases identically.
Priority lists are worse in a different way. They encode a snapshot of the network at the moment somebody wrote them, they get edited under pressure during peak, and they persist for years. The practical test is to ask when your node priority list was last changed and what evidence changed it. If the answer is a capacity incident in a previous peak, the list is now optimising against a network that has since gained two nodes.
The cost function worth minimising has four components: outbound parcel cost for the assignment as a whole, the handling cost at each node, the expected cost of the splits the assignment creates, and the future cost of consuming the units it consumes. The first three are arithmetic. The fourth is the one that changes behaviour.
Putting a price on the unit you just consumed
Acimovic and Graves, in Manufacturing and Service Operations Management in 2015, set out the version of this that has held up. Rather than minimising today's shipping cost, minimise today's cost plus an estimate of the future cost your choice creates, where that estimate comes from the dual values of a transportation linear programme solved over the remaining horizon. Each unit at each node gets an opportunity cost, and the sourcing decision compares today's parcel saving against it.
The intuition is easier than the formulation. A node holding forty units against expected demand of a hundred and twenty over the horizon is scarce, and any unit you take from it is a unit that a nearer future order will have to source from further away. A node holding nine hundred units against the same demand is not scarce, and taking a unit from it costs nothing beyond today's parcel.
Put numbers on it. Node A is close to the customer, holds 40 units, and is the cheapest node for about 120 expected orders over the next three weeks. Node B is a zone further out, holds 900. Sourcing from A saves 1.20 on today's parcel. But A will run out roughly two thirds of the way through the horizon, and every order it cannot serve after that goes to B or further, at an average penalty of 3.40. If the marginal unit at A has around a 0.6 chance of being the one that tips a future order into that penalty, its opportunity cost is about 2.04, which is more than today's saving. The correct decision is to pay 1.20 more now.
Acimovic and Graves returned to the same effect in 2017 under the name spillover: orders that have to be served from a distant node because a near node is empty, with the cost showing up weeks after the sourcing decisions that caused it. Spillover is the reason a network can post excellent per-order routing metrics in week one of a peak and terrible outbound cost in week four.
Andrews, Farias, Khojandi and Yan published an implementation of primal-dual fulfilment optimisation at Urban Outfitters in INFORMS Journal on Applied Analytics in 2019, reporting a mid-single-digit percentage reduction in outbound shipping cost against the retailer's previous policy. Single digits on outbound freight is a large number in retail terms, and it came from changing the objective rather than from adding nodes.
Holding the decision instead of making it at click
Sourcing at the moment of click is a design choice that most systems make by default, and it is the most expensive moment to decide. At click you know one order. Ten minutes later you know several, and their overlaps are visible.
Xu, Allgor and Graves, in Manufacturing and Service Operations Management in 2009, examined re-evaluating fulfilment decisions rather than fixing them at order capture, and found that periodic reassignment before release reduces the number of shipments materially without changing what the customer was promised. The mechanism is that inventory positions and the order pool both move during the window, so a decision that was forced at 14:02 becomes unnecessary at 14:20 when a receipt lands or a second order for the same node arrives.
The implementation question is how long you can hold. The answer is bounded by the earliest carrier cutoff among the candidate nodes minus the pick and pack time at that node, and for most networks that leaves a window of tens of minutes on same-day releases and hours on next-day. That window is free. The reason it usually goes unused is that the order management system was configured to release immediately so that the warehouse queue never looks empty.
Two constraints make holding harder than it sounds. Customer-facing promise displays that quote a specific node or a specific date at checkout have to be satisfiable under every assignment the batch might produce, which is a handoff to the available-to-promise logic covered elsewhere (I7). And any hold has to fail open, so that a batch process which does not run releases everything on the previous rule rather than stalling.
When splitting is the right answer
Consolidation is not free and the objective should not be a split rate of zero. Splitting is correct in several ordinary situations, and a rule that forbids it will do damage of its own.
When no single node has the full order. Forcing consolidation then means either a stock transfer between nodes, which costs more than the second parcel almost every time, or making the customer wait for a receipt.
When the items are physically incompatible. A forty kilogram item and a small accessory ship on different services, and pretending they are one shipment produces a manual exception in a warehouse rather than a saving.
When the consolidation node has run out of picking capacity. Forcing every line of an order onto one node during peak moves work to a node that cannot absorb it, and the promise breaks in a way that costs more than freight.
When the second parcel avoids a cancellation. A split that ships eighty per cent of the order today and the remainder in three days is usually better for the customer than holding the whole order, and much better than cancelling a line.
The useful control is a threshold rather than a ban. Compute the incremental cost of the split against the incremental cost of the consolidated alternative including any delay, and allow the split when it wins. That comparison needs both numbers to exist, which is the part most configurations skip.
What to measure once it is running
Split rate per order is the headline, and it needs a denominator of multi-line orders only, because single-line orders can never split and their share moves with assortment. Alongside it, track outbound cost per order against a same-mix baseline, and units stranded, meaning units at nodes that ended the season without being sold or transferred.
Cancellation rate belongs in the same view because the failure mode of an aggressive consolidation policy is that it promises from a node that turns out to be short. If your split rate falls and your cancellation rate rises, the sourcing engine has moved cost rather than removed it.
Where this stops
The opportunity cost approach needs a demand forecast at node level over the horizon, and node-level forecasts are the weakest ones you own. Total demand for an item is forecastable; the share of it that will arrive at postcodes nearest to node A over the next three weeks is much noisier, and the dual values inherit that noise. The method still beats distance sorting because it only has to get the scarcity ranking of nodes roughly right, but a network with volatile geographic demand will see the advantage shrink.
Small networks may not need any of this. With two distribution centres and no store fulfilment, the assignment space is tiny, splits are rare, and a sensible rule set captures most of the available value. The methods start paying when the node count reaches somewhere around ten and the majority of orders have more than one feasible assignment.
Carrier contract structure can also flatten the argument. If you buy freight on a flat per-parcel rate with no zone differentiation, the distance term disappears from the objective and only the split and scarcity terms remain. That is worth checking before building anything, because it changes which of the four cost components is doing the work.
The last honest limit is that a re-optimising sourcing engine is harder to explain than a priority list. When a store manager asks why an order went to a node three zones away, the answer is a dual value, and that answer has to be renderable as a sentence in an audit record or the operation will lose confidence in it.
Take one week of multi-line orders, recompute what each would have cost as a single parcel from the cheapest node that could have filled the whole order, and compare that with what you actually paid.
More on this in omnichannel on the DemandLabs platform.




