Cursor pagination

Cursor pagination asks for the next page by naming where the last one stopped, rather than by counting how many rows to skip. The request carries a pointer, usually the sort key of the final row (after=2026-04-11T09:14Z, or an opaque token encoding the same thing), and the answer is the rows immediately after it. Offset paging instead says “skip 30, give me 10”, which is a position in a list that is assumed to be holding still. On a feed, an inbox, or an audit log, it is not holding still, and every row inserted above the reader’s position pushes one row down across the boundary. That row is served twice and the row it displaced is served never, which is the failure the demonstration above shows: page two opens with a post the reader has already read.

The visible consequence is the reason this belongs in an interface vocabulary rather than a database one. A cursor cannot jump to page seven, because nobody knows what the seventh page starts with until the sixth has been read. So a cursor-paged list offers Next and Previous and nothing else: no numbered pages, no last page, and usually no total count either, since counting the whole set is the expensive query the design was avoiding. If a product wants numbered pages, someone has to accept offsets and their drift, or precompute page boundaries and accept that they go stale. This is a trade, not an oversight, and it is worth naming out loud in a design review before somebody adds a page number field to the mock.

Where it sits among neighbours is mostly a question of who does the paging. Pagination is the numbered control and the mental model of a fixed set. Load more and infinite scroll are the two places cursors are already the norm without anybody saying the word: both only ever go forward, both are appending to what is on screen, and both would be visibly broken by offset drift, since a duplicate row appears directly under the original. If your feed occasionally shows a repeated item, this is very often why.

Design the empty and the edge honestly. Disable Previous on the first page rather than hiding it, so the control does not move. Say “no more results” instead of showing an empty page. Keep the cursor in the URL so a page can be shared and reloaded, and treat an expired or invalid cursor as a soft failure that resets to the top with an explanation, never as a blank screen. And if the list is short and static (a settings table, a small catalogue), offsets are fine: this pattern earns its constraints on data that changes while it is being read.