Java and backend interview questions

On a network partition you cannot have both linearizable consistency and availability.

L4 (TCP) balances connections without reading HTTP, which is simple and fast.

Stickiness pins a user to one instance so in-memory session state works, which defeats even load and complicates deploys.

Do not start a global XA transaction across services.

Choreography: OrderPlaced triggers Inventory to reserve

You write the business row and an outbox row in the same database transaction, then a publisher streams outbox events to Kafka.

The place-order API commits the order as CREATED and returns quickly

Each client key has a bucket of tokens that refill at a constant rate

The merchant sends a unique merchant transaction id

SET key token NX EX seconds so only one caller wins and the lock expires if the process dies.

Idempotent producers prevent duplicate writes of the same produce batch to a partition after retries, using sequence numbers.

Check processing time versus poll interval, stop-the-world GC, downstream DB pool exhaustion, and whether one hot partition has a skewed key.

G1 is the common default: good throughput with pause goals in tens of milliseconds at moderate heaps.

A heap leak shows old-gen occupancy rising across GC cycles and retained objects in a heap dump, often caches without bounds.

Thread dump if requests hang, CPU is high on a few threads, or you suspect deadlock: you see stack traces and lock owners.

Many threads blocking on the same monitor or striped lock serialize work and burn CPU on context switches.

parallel streams steal from the common ForkJoinPool and can starve the whole JVM under load.

Always lock accounts in a total order, for example by ascending account id, so two transfers cannot wait on each other.

They are proxy-based AOP.

A method that catches nothing and throws IOException can still commit partial JDBC work.