PrepZone Logo
PrepZone

Interview Reference Sheet

Dense reference card — architecture, semantics, tuning knobs, and trade-offs for senior backend interviews.

Why this matters

  • Single-page review before VaultCommerce loop or external interviews.
  • Covers what interviewers probe at staff level: EOS, KRaft, partitioning, ops.
  • Cross-links mental models from all 51 prior articles.
  • Not a substitute for depth — a structured recall aid.

Cloud building blocks toolbox

NETWORKRoute 53DNS + GeoDNS
NETWORKCloudFrontCDN edge
NETWORKAWS ALBL7 balance
NETWORKAPI Gatewayauth + throttle
COMPUTEAmazon EKScontainer fleet
DATABASEElastiCacheRedis cache
DATABASERDS / Aurorarelational
DATABASEDynamoDBkey-value
INTEGRATIONMSK / SQSasync events
STORAGEAmazon S3object store
ANALYTICSOpenSearchfull-text
OPSCloudWatchmetrics · logs
SECURITYAWS WAFedge security
SECURITYCognitoidentity
COMPUTELambdaserverless
INTEGRATIONSNS / SESnotify
Every StreamHub product design combines subsets of these AWS components.

Architecture quick reference

Broker hosts partition leaders. Topic → partitions → ordered offsets. Consumer groups divide partitions. KRaft controllers manage metadata. RF=3, min.insync.replicas=2, acks=all for durability. VaultCommerce baseline cluster: 6 brokers, 12 partitions on order topics.

Key points

  • KRaft — metadata quorum; no ZooKeeper
  • ISR + acks=all — durable produce path
  • Partition key — ordering scope
  • Consumer lag — primary consumer health metric
  • Transactional outbox — DB-Kafka consistency pattern

Semantics and patterns

At-least-once default; idempotent producer for log dedup; transactions for read-process-write; outbox for DB+Kafka atomicity; saga via choreography; DLT for poison messages. Ordering per partition key. Schema Registry BACKWARD compatibility.

Tuning and ops knobs

Producer: linger.ms, compression.type=lz4, enable.idempotence. Consumer: max.poll.interval, CooperativeStickyAssignor, manual commit. Ops: lag alerts, URP alerts, rolling restarts, MM2 for DR. Spring: KafkaTemplate, @KafkaListener, DefaultErrorHandler.

Java
// VaultCommerce interview skeleton — producer + consumer durability baseline
// Producer: acks=all, idempotence, Avro
props.put(ProducerConfig.ACKS_CONFIG, "all");
props.put(ProducerConfig.ENABLE_IDEMPOTENCE_CONFIG, true);

// Consumer: manual commit, cooperative rebalance, idempotent handler
props.put(ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG, false);
props.put(ConsumerConfig.PARTITION_ASSIGNMENT_STRATEGY_CONFIG,
    CooperativeStickyAssignor.class.getName());
// + processed_events dedup table or natural SQL idempotency

Quick recall

Everything you need if you only revisit this box.

  1. Log not queue; offsets not deletion.
  2. Durability = acks=all + min ISR + RF.
  3. End-to-end EOS needs outbox or idempotent consumers.

Test yourself

Answer these before moving on — recall is what makes it stick.