Ordering, Pagination & Projections

Graphs do not store an inherent “table order.” When you MATCH movies and people, the engine is free to return rows in any order that is correct for the pattern. User interfaces, however, need ranked lists: newest films first, highest review averages, alphabetical actor directories. ORDER BY sorts the current result rows by one or more expressions before later clauses such as SKIP, LIMIT, or another WITH.

Sort keys may be variables already in scope, property accesses, or computed expressions. Each key can be ASC (default) or DESC. Null ordering is version- and configuration-sensitive in edge cases; if nulls appear in production sorts, test the behavior on your Neo4j version and consider coalescing to a sentinel when business rules demand it.

After aggregation (Lesson 06), you typically order by the aggregate aliases—movie counts, average ratings—because those values only exist after the grouping boundary. You can also order inside a WITH pipeline so a later collect sees a deterministic sequence (for example, collect titles newest-first).

MATCH (m:Movie)
RETURN m.title AS title, m.released AS released
ORDER BY released DESC, title ASC
SKIP 0
LIMIT 10

LIMIT n keeps at most n rows. SKIP k discards the first k rows of the ordered stream. Together they implement classic offset pagination: page 1 is SKIP 0 LIMIT 10, page 2 is SKIP 10 LIMIT 10, and so on. Clause order after RETURN is ORDER BY, then SKIP, then LIMIT.

Stability requires a total order. If two movies share the same released year and you order only by that year, page boundaries can shuffle between requests. Always add a unique tie-breaker such as title when unique in your domain, or elementId(m) when you need a guaranteed unique secondary key.

Deep offsets are expensive: the engine still produces and discards the skipped rows. For large catalogs, prefer keyset (seek) pagination: remember the last sort key from the previous page and filter WHERE released < $lastReleased OR (released = $lastReleased AND title > $lastTitle) with a small LIMIT. Offset pagination remains fine for admin UIs and small result sets.

Returning entire nodes ships every property—including ones clients never display—and can bloat payloads as models grow. Map projection syntax m { .title, .released } builds a map with only the selected properties. You can add aliases, literals, and computed entries: m { .title, year: m.released, source: 'movie-graph' }.

Sub-projections nest related data. A common pattern matches a movie and its actors, then returns m { .title, .released, actors: collect(p { .name, .born }) }. The result is a JSON-like document ideal for REST or GraphQL resolvers without over-fetching.

UNWIND is the inverse of collect: it turns a list into rows. Use it for parameter lists of ids, for post-processing collected values, or for generating ranges when combined with list functions. After UNWIND, you can MATCH from each element, filter, and re-aggregate.