A warehouse can only pick what it was told to pick, in the quantity the system believes it holds. Almost every fulfilment failure a seller experiences at an overseas warehouse is not a labour problem or a racking problem — it is a data problem that started hours earlier, somewhere between a storefront and a warehouse management system.
Sellers tend to think of integration as a single pipe: the order goes out, the tracking comes back. In practice a working setup carries three distinct flows, and each fails in its own way. The inbound flow tells the warehouse what is arriving, on which reference, in which cartons, so receiving can be checked against something. The order flow carries the picking instruction and the service level attached to it. The inventory flow keeps both sides agreeing on how many units exist and how many of those are already promised to someone else.
Underneath all three sits a fourth, quieter flow: events. A warehouse that only accepts orders and returns tracking is a black box with a delay. A warehouse that reports receipt discrepancies, short picks, damaged units, cancelled lines and returns as they happen gives you a supply chain you can actually manage. When you evaluate a 3PL, ask what they send back, not only what they accept.
There is no single correct transport. The right answer depends on who you are, who your trading partners are, and how much engineering you want to own after go-live. The mistake is choosing the fashionable option rather than the one that fits the volume and the counterparties.
| Transport | Best suited to | Strength | Weakness |
|---|---|---|---|
| REST / JSON API | Own website, marketplace connectors, in-house OMS | Real-time answers, per-order status, easy to test | You own retries, pagination, rate limits and reconciliation |
| Webhooks / callbacks | Order acknowledgements, shipment and stock events | Removes polling; event arrives with the change | Silent if your endpoint is down; needs replay and idempotency |
| EDI (X12 / EDIFACT) | Retail, wholesale and large 3PL networks | Standard document semantics: 940, 945, 947, 846, 856 | Setup lead time, mapping cost, less forgiving of bad data |
| Flat file over SFTP | Low volume, batch schedules, legacy systems | Simple, auditable, cheap to run | Stale by design; a bad batch can be loaded twice |
A practical pattern is a hybrid: API for orders and inventory where speed matters, EDI where a retailer or a large warehouse network expects standard documents, and a nightly file as the reconciliation of record. Whatever the mix, insist on one thing — a way to ask the system, at any moment, what it currently believes about a specific order reference. Without that, every dispute becomes archaeology.
The single most consequential number in a warehouse integration is the order cut-off, and it is usually the least precisely defined. "Same-day dispatch" means nothing until you know the time zone, whether weekends count, and whether the clock starts when the order is placed or when the warehouse accepts it. Two channels feeding one warehouse can have different cut-offs, which is fine — as long as the service level travels with the order rather than being inferred from arrival time.
Idempotency matters more than speed. Every realistic integration retries, and every retry risks a duplicate shipment. The fix is not vigilance; it is a required unique key on every order message plus a warehouse that treats a repeated key as an update rather than a new instruction. Ask directly whether a duplicate submission creates a second pick. If the answer is a discussion rather than a yes or no, treat that as a no.
Then there is the queue. Orders that fail validation — unknown SKU, missing address field, a value the warehouse cannot map — must land somewhere visible, with the raw payload retained. An integration that rejects silently into a log nobody reads will look perfect in testing and lose a weekend of orders at peak.
Most stock disagreements are semantic, not physical. A warehouse holds three quantities that matter: on hand, allocated to open orders, and available for new promises. Sellers who sync only on hand will oversell, because the units sitting in a pick wave are already spoken for. Sellers who sync only available will panic-buy when a promotion temporarily consumes the buffer.
Three settings decide whether the numbers behave. Reservation logic determines when stock is committed — at order acceptance, at wave release, or at pick. Safety buffer is the quantity you withhold from sale so a short pick does not become a cancelled order. And the refresh cadence decides how stale your storefront can be: real-time pushes are accurate but noisy, while scheduled snapshots are calm but blind to a spike.
Write down which quantity maps to which field on both sides, and reconcile weekly against a physical count rather than against the other system. Two systems agreeing is not evidence of accuracy; it is evidence of consistent assumption.
Outbound tracking is the minimum, not the standard. The events that change decisions are the less glamorous ones: the receipt confirmation that closes a purchase order, the discrepancy report that flags cartons short or damaged, the short-pick notice that lets you substitute or refund before the customer complains, the pick-and-pack confirmation that updates stock mid-wave, and the returns receipt that puts a unit back into sellable, quarantine or scrap with a reason code.
Each event should carry the reference the other side can match on, a timestamp in a stated time zone, and enough detail to act without a phone call. If a returns event says only that a unit came back, you cannot decide whether to relist it, and the unit will age in the warehouse while someone chases an email.
Labels are where integration touches money. A warehouse that generates labels through carrier APIs can rate-shop across services, apply the correct service for the destination zone, and produce compliant documentation without human typing. A warehouse that receives pre-printed labels from you inherits every address and zone error you made, and a wrong zone is not refunded after the parcel moves.
Decide who owns the label, who owns the tracking number, and who owns the claim when a parcel is lost. If the warehouse books the carrier, the claim sits with them and your recourse is contractual. If you book the carrier, you keep control of rates but own the exception handling. Both models work; mixing them per shipment usually does not.
Sign-off should be based on behaviour observed, not on a specification read. Run the following before peak season rather than during it.
| Test | Pass condition |
|---|---|
| Duplicate order submission | One pick task, one shipment, second message logged as a duplicate |
| Order with one invalid line | Valid lines proceed, rejected line reported with a reason code you can parse |
| Unknown SKU | Order parked in an exception queue, raw payload retrievable, alert raised |
| Endpoint down for an hour | Events replay in order after recovery, none lost, none doubled |
| Inbound with a short carton | Discrepancy filed against the ASN reference the same day |
| Stock reconciliation | On hand, allocated and available reconcile to a physical count within an agreed tolerance |
| Returns receipt | Unit dispositioned with a reason code and a sellable or quarantine flag |
| Peak simulation | Agreed multiple of normal daily volume processed without manual intervention |
| Cut-off boundary | Orders either side of the cut-off dispatch on the correct day |
The questions that separate a mature integration from a connector are the same everywhere. Which fields are mandatory, and what happens to a message that omits one? Is there a sandbox, and does it behave like production? What is the documented retry and replay policy? How are changes versioned, and how much notice do you get before a schema change? Who is paged when an interface fails at the weekend? What is the contractual remedy if the interface is unavailable for a day?
Then ask for the boring artefact that reveals everything: a sample of every message type they send, with real field values. A 3PL that can produce those quickly has done this before. One that needs to check usually has not.
This is operational guidance, not legal or tax advice, and marketplace and carrier specifications change without much notice, so confirm current requirements with the platform or the warehouse before go-live. 明杰供应链 treats integration design as part of the shipping brief rather than an afterthought, because a lane is only as reliable as the data that travels with the cargo.
Send us the channels you sell on, the systems you already run, and the volume you expect at peak. We will come back with the flows we would build, the fields we need from you, the events you would receive, and a testing plan you can run before your next promotion.
Discuss Your Integration →