Write a SQL query that returns the next page of orders using `CreatedAtUtc` and `OrderId` as the cursor.

Table: Orders(OrderId, CustomerId, CreatedAtUtc, Status). Return @PageSize non-deleted orders ordered newest first. For the first page both cursor parameters are null. Later pages receive the last row's CreatedAtUtc and OrderId.

If the previous page ends at (2026-07-18 10:00, OrderId 42), the next page includes rows with an earlier timestamp, plus rows at 10:00 whose id is lower than 42.

• Sort by CreatedAtUtc DESC, OrderId DESC.
• Cursor comparison must exactly match that ordering.
• Return at most @PageSize rows.
• Validate a sensible positive page-size limit in the API.

For descending order, rows after the cursor have an earlier timestamp, or the same timestamp and a smaller id. I keep both cursor values together because the timestamp alone is not unique. The matching composite index lets the database continue from the cursor instead of reading and discarding a large offset.

SELECT TOP (@PageSize)
    o.OrderId,
    o.CustomerId,
    o.CreatedAtUtc,
    o.Status
FROM dbo.Orders AS o
WHERE o.Status <> 'Deleted'
  AND
  (
      @AfterCreatedAtUtc IS NULL
      OR o.CreatedAtUtc < @AfterCreatedAtUtc
      OR
      (
          o.CreatedAtUtc = @AfterCreatedAtUtc
          AND o.OrderId < @AfterOrderId
      )
  )
ORDER BY o.CreatedAtUtc DESC, o.OrderId DESC;

An index on (CreatedAtUtc DESC, OrderId DESC) including the projected columns supports the ordered range. A filtered index may help if deleted rows are common and the database can match the filter. The null first-page branch may lead to plan variation, so production workloads should inspect actual plans.

• Many orders share the same timestamp.
• The first page has null cursor values.
• The final page contains fewer rows than requested.
• Rows are inserted between page requests.

• Create more same-timestamp rows than fit on one page.
• Walk every page and assert no id repeats.
• Insert a newer row between requests and verify forward paging stays stable.
• Compare the concatenated pages with the same ordered snapshot.

• Using only the timestamp as the cursor.
• Reversing one comparison while keeping descending sort.
• Using OFFSET and calling it keyset pagination.
• Accepting an unbounded page size.