引導式測驗 · 硬 等級

Kafka 和 Flink 的串流功能測驗

了解 Kafka 的角色、事件時間和浮水印、後製資料處理、Flink 檢查點以及一次性保證的限制。

問題 1 的 8

Apache Kafka 在串流媒體功能管道中扮演什麼角色?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. Apache Kafka 在串流媒體功能管道中扮演什麼角色?
    • It acts as a durable event log, with record ordering guaranteed within each partition
    • It trains the machine learning model directly
    • It stores the final trained feature store snapshot
    • It replaces the need for any stream-processing engine
  2. 翻滾窗和滑動窗有何不同?
    • Sliding windows only work with Kafka, tumbling windows only work with Flink
    • Tumbling windows divide time into fixed, non-overlapping intervals; sliding windows compute a continuously updating rolling aggregation
    • Tumbling windows never expire; sliding windows expire every second
    • There is no meaningful difference between them
  3. 為什麼詢問過去 10 分鐘內發生了多少筆交易的詐騙模型通常會使用滑動視窗而不是滾動視窗?
    • Sliding windows are always faster to compute
    • Tumbling windows cannot count transactions
    • Because it needs a continuously updating rolling count rather than a fixed periodic bucket
    • Sliding windows use less memory in every case
  4. Flink 的水印解決了什麼問題?
    • Compressing event data to save storage
    • Encrypting events in transit between Kafka and Flink
    • Choosing which Kafka topic to read from
    • Estimating how far behind in event time the stream has progressed, to handle late-arriving events
  5. 在配置的允許延遲期之後到達的事件時間視窗記錄會發生什麼情況?
    • It is dropped from that window or routed to a configured late-data side output
    • Memory usage increases without bound
    • Kafka stops accepting new events
    • Windows never finalize their results
  6. Which resource tradeoff can result from a long allowed-lateness period for an event-time window?
    • Late events are dropped immediately
    • Window state is retained longer so late records can revise results, using more state
    • Kafka partitions are automatically rebalanced
    • The feature store stops updating entirely
  7. According to the guide, what is required for end-to-end exactly-once stream output?
    • Only a Flink checkpoint, regardless of the sink
    • A high watermark setting by itself
    • A checkpointed/replayable source and a sink that participates in compatible exactly-once or idempotent handling
    • Disabling event timestamps
  8. What does the guide identify as a common misconception about streaming features?
    • That streaming features are always less accurate than batch features
    • That Kafka cannot be used for feature engineering
    • That Flink requires a relational database to function
    • That they are simply batch, but faster, when streaming introduces distinct correctness concerns like out-of-order handling