Streaming Data Language Exercises

1 / 5

A Kafka architect explains: "The topic has 12 partitions with a replication factor of 3."

What does a replication factor of 3 mean?

Replication factor = number of broker copies of each partition:

Role	Function
Leader	Handles all reads and writes for the partition
Followers	Replicate leader data; promote on leader failure

Minimum recommended: RF=3 for production. RF=3 tolerates up to 2 broker failures. Key vocabulary: leader partition, follower replication, ISR (In-Sync Replicas), unclean leader election, acks=all for durability.

2 / 5

A team has two consumer instances in the same consumer group reading from a 6-partition topic.

How are partitions assigned?

Consumer group = cooperative partition consumption — each partition is consumed by exactly one consumer in the group:

Scenario	Assignment
6 partitions, 2 consumers	3 partitions each
6 partitions, 6 consumers	1 partition each (max parallelism)
6 partitions, 8 consumers	2 consumers sit idle

Key vocabulary: consumer group rebalance, partition assignment, lag monitoring, committed offset, at-least-once vs. exactly-once.

3 / 5

A monitoring alert fires: "The consumer group is at 45,000 messages lag on partition 3."

What does consumer lag indicate?

Consumer lag = latest_offset − committed_offset

Lag situation	Likely cause
Sudden large lag spike	Consumer crash, temporary partition reassignment
Slowly growing lag	Consumer throughput below producer rate
Stable lag	Consumer keeping pace but behind (acceptable in some systems)

Key vocabulary: consumer lag, offset commit, lag monitoring (Kafka UI, Burrow), backpressure, lag recovery rate.

4 / 5

An engineer explains: "The streaming pipeline uses event-time windowing rather than processing-time windowing."

What is the key difference?

Event-time vs. processing-time windowing:

Time domain	Based on	Challenge
Event-time	Timestamp embedded in the event payload	Late/out-of-order events require watermarks
Processing-time	Wall-clock time when event arrives at processor	Inaccurate for events delayed in transit

Key vocabulary: watermark, late data handling, allowed lateness, window trigger, tumbling vs. sliding vs. session window.

5 / 5

A system design document states: "The pipeline guarantees exactly-once delivery semantics."

What does exactly-once semantics mean and why is it hard to achieve?

Delivery guarantees — the fundamental trade-off in distributed messaging:

Guarantee	Trade-off	Implementation
At-most-once	May lose messages, no duplicates	Fire and forget
At-least-once	No loss, may duplicate	Retry on failure
Exactly-once	No loss, no duplicates (hardest)	Idempotent producer + transactional API

Key vocabulary: idempotent producer, transactional API, 2PC in streaming, semantic guarantee trade-offs, Flink checkpointing.