Executive summary
- What: From GraphQL Admin API version 2026-10, an exclusive greater-than bound on a timestamp filter excludes the entire second rather than only its first microsecond.
- Why: Queries that feed a returned timestamp back in as an incremental cursor can now miss records created later in that same second.
- Who: Teams running custom apps, ERP or OMS connectors, and middleware jobs that poll orders, products, or inventory on a schedule.
What changed
Shopify returns timestamps such as createdAt at second precision, but several resources store them internally with fractional seconds. Before 2026-10, a filter of created_at:>'2026-09-15T12:00:02Z' was compared against 12:00:02.000000, so a record created at 12:00:02.500 matched it. That same record also matched created_at:<='2026-09-15T12:00:02Z', which meant the two ranges overlapped.
From 2026-10, a greater-than value carrying no fractional component starts the window after the end of that second, at 12:00:02.999999. A record matches only when the timestamp the API returns for it is genuinely later than the bound, so the greater-than and less-than-or-equal ranges stop overlapping. Nothing changes, on any version, for the other comparators, for date-only values, or for values that already include fractional seconds.
Why it matters
The old overlap was quietly absorbing a design flaw in a lot of integrations. A connector that stores the latest createdAt it has seen and passes it back as a greater-than bound was relying on second-level truncation to re-fetch the boundary record, then discarding it as a duplicate. Under the new behaviour that same pattern does the opposite: anything created later in the boundary second is skipped, permanently, because the cursor has already moved past it.
The upside is that windows are now deterministic. Adjacent reporting ranges no longer double count the same order, which makes reconciliation against finance systems cleaner. The cost is that incremental sync now needs explicit deduplication. This is the same class of issue we surface in most ERP integration reviews: the sync logic is correct in testing and loses records only under volume.
Role-specific impact
- Marketers: Reports built on timestamp-bounded windows will stop double counting orders that sit on a window boundary. Expect small, one-time shifts in daily totals after an upgrade, and do not read them as a traffic change.
- Developers: Audit every query argument sent to an Admin API connection and flag each greater-than bound without fractional seconds. Either move to the inclusive comparator and deduplicate by id, or carry full microsecond precision in the cursor.
- Store admins: If an ERP or OMS connector appears to drop a handful of orders shortly after an app update, ask the vendor which API version and which filter the sync uses before you start reconciling by hand.
Use-case example
Real-world scenario
A Plus merchant polls orders every five minutes and passes the createdAt of the last order it processed as a greater-than bound. During a promotion running near 90 orders per minute, more than one order frequently lands inside the same second. With 288 polling windows a day, if even a third of boundaries fall on a second holding a second order, roughly 95 orders a day never reach the ERP. Those surface as missing fulfillments and inventory drift. Switching the bound to the inclusive comparator and deduplicating by id removed the gap without changing the polling interval.
Implementation checklist
- Inventory every integration that sends a query argument to an Admin API connection, including custom apps, scheduled middleware jobs, and internal reporting scripts.
- Search those queries for timestamp comparators and flag each greater-than bound that has no fractional seconds.
- For each flagged query, decide whether a missed record is tolerable. For order, payment, and inventory sync it is not.
- Replace the exclusive bound with the inclusive one and deduplicate results by id in your own code before writing downstream.
- Persist the set of processed ids alongside the last timestamp so that a rerun stays idempotent.
- In staging, move the integration from its current version to 2026-10 and replay a high-volume hour, comparing record counts against the old version.
- Schedule the production version bump outside peak trading and watch downstream record counts for a full day afterwards.
FAQ
Q: Do we need to change anything if our apps stay on 2026-07 or earlier?
A: No. The behaviour applies only from 2026-10 onward, so this belongs in your version upgrade test plan rather than on your incident list.
Q: Is the inclusive comparator plus deduplication the only safe pattern?
A: It is the simplest and our default recommendation. If your cursor can store and send full fractional precision, passing a microsecond-accurate value to the exclusive comparator also works, provided every component in the chain preserves that precision.
Resources
Shopify API search syntax reference
Need guidance? Talk to Makro.