Java and backend interview questions

Boot keeps the Hibernate session open until the view or Jackson finishes rendering, so lazy loads in the serializer work.

It shares entity data across persistence contexts, which helps read-mostly reference data.

Optimistic uses a version column and fails on conflict, which is great for rare collisions.

Find requests that borrow a connection and never return it: missing try-with-resources, a long HTTP call inside @Transactional, or Open Session in View.

A composite index that matches equality then range then order, such as (category_id, price, created_at), can serve the query without a sort if the planner agrees.

A phantom is a new row that matches your previous range query when you re-query in the same transaction.

Partitioning is still one database, split for maintenance and pruning large tables, like orders by month.

Replicas take SELECT load off the primary.

Each nested field resolver hits the database per parent, which is N+1 by architecture.

Pure stateless JWT cannot be revoked except by waiting.

A gateway JWT proves the user to the first hop

Use environment variables injected by the orchestrator, or a vault such as HashiCorp Vault or cloud secret manager, mounted at runtime.

Blue-green stands up a full new fleet and flips the load balancer, which is fast rollback but doubles capacity.

Liveness should restart a deadlocked JVM

allkeys-lru evicts any key under memory pressure, so even locks and sessions can vanish unless they are tiny and you size Redis for them.

One key taking a huge share of QPS saturates a single Redis thread and CPU.

When members join or miss heartbeats, Kafka revokes partitions and the group stops processing during the rebalance.

OFFSET 1e6 still walks that many index entries.

Persist a unique reservation id per attempt and decrement with UPDATE ...

Structured logs with a correlation id, RED metrics on each endpoint, and traces that follow the request into SQL and Kafka.