What does Hibernate's @BatchSize annotation — or the hibernate.default_batch_fetch_size setting — change about how lazy associations are loaded?

## The problem it addresses Lazy loading defers a query until the association is touched. When code walks a list of parents and touches a lazy association on each, the deferral fires once per parent: one query for the parents, then one per row. Batch fetching does not remove the extra queries; it **groups** them. ## The mechanism When Hibernate loads entities lazily it does not immediately go to the database for the one you touched. It first scans the persistence context for other objects of the same batchable kind that are still uninitialised — other proxies of the same entity class, or other uninitialised collections of the same role (that is, the same mapped attribute of the same entity). It then issues one query for up to `batchSize` of them: ```sql select ... from order_line where order_id in (?, ?, ?, ?, ?) ``` All of those associations are initialised in one round trip; the one you touched is returned to your code and the others are simply ready when you reach them. With 100 parents and `size = 25`, you get four queries instead of 100. Crucially the association remains **lazy**. Nothing is loaded unless something touches it, so a request path that never reads the collection pays nothing. ## Where you configure it **Per collection** — annotate the collection field. This batches initialisation of that specific collection role: ```java @OneToMany(mappedBy = "order") @BatchSize(size = 25) private Set<OrderLine> lines; ``` **Per entity class** — annotate the entity. This batches initialisation of *proxies* of that entity, which is what you want for the many-to-one direction, where each parent row holds a proxy to the same target type: ```java @Entity @BatchSize(size = 25) class Customer { ... } ``` A `@BatchSize` on the `@ManyToOne` field itself is not how the to-one case is configured; the annotation belongs on the target entity class (or you rely on the global setting). **Globally** — `hibernate.default_batch_fetch_size` applies a default to every batchable association. It is off by default (no batching). Setting it to a sane value is one of the highest-value single properties in a Hibernate application, because it silently improves every accidental N+1 rather than only the ones someone remembered to annotate. A per-association `@BatchSize` overrides the global value. ## Which side of the query it affects Batch fetching works on the **loading of associations**, not on the root query. It does not affect how many rows your JPQL returns, does not duplicate roots, and does not interfere with `setMaxResults`. That last property is what makes it the natural companion to paginated queries, where a fetch join forces Hibernate to paginate in memory. ## Practical notes and edge cases **The IN list and parameter churn.** A batch of 25 does not always find 25 pending items, so Hibernate must issue queries with varying numbers of parameters. Different parameter counts mean different SQL strings, which means more entries in the statement cache and more parse work on the database. Hibernate mitigates this by padding IN-clause parameter counts to powers of two, controlled by `hibernate.query.in_clause_parameter_padding` (on by default in modern Spring-managed configurations; verify on plain Hibernate). Padding repeats the last id to fill, so results are unaffected. **It only helps when several proxies are pending.** Loading one parent with `find` and touching its collection cannot batch anything — there is one collection. Batching pays off precisely in the list-processing case. **Order of loading is not guaranteed to be your iteration order.** Hibernate picks pending items from the persistence context; the batch may include entities you have not reached yet, which is the point, but it means you cannot reason about "the first N". **It does not deduplicate work across sessions.** Batching is a persistence-context-scoped optimisation. Across requests, caching is the tool, not batching. **Choosing a size.** Values between 10 and 50 cover most cases. Too small and the round-trip count stays high; too large and you risk long IN lists, plan variance, and loading much more than the request needs. Databases also have parameter limits, so very large batch sizes are not free. ## Interview framing The crisp version: "associations stay lazy; batch fetching groups the deferred loads with an IN predicate, configured per collection, per entity class, or globally, and it is the strategy that survives pagination". Adding the padding detail and the "only helps with many pending proxies" caveat shows operational familiarity.