The log — events, state, retention and replay

A partition is an append-only log. Offsets always increase and never change. Reading does not remove records. Compaction (keep the latest record per key) is an optional policy. It is not the base model.

Source: https://kafka.apache.org/documentation/#intro_concepts_and_terms

offsets_for_times maps a timestamp to the first offset at or after it (for each partition). Then you seek to that offset and read forward. Reading from the beginning and filtering also works. But it rereads retained history that you do not need.

Source: https://kafka.apache.org/documentation/#intro_consumers

Kafka guarantees order within a partition. The default partitioner sends a given key to one partition, so per-key order is preserved. There is no cross-partition order, so there is no global order.

Source: https://kafka.apache.org/documentation/#intro_concepts_and_terms

Time and size retention delete old segments. Compaction keeps the latest value per key. You can replay only back to the earliest retained offset. Reading never deletes records, so D is false.

Source: https://kafka.apache.org/documentation/#compaction

State is a projection of events. You fold the log to compute it. A compacted topic stores exactly the latest record per key. Changelogs and __consumer_offsets work this way. acks is a durability setting and is unrelated.

Source: https://developer.confluent.io/courses/apache-kafka/topics/

Kafka decouples producers and consumers in time. Synchronous request/response is the opposite pattern. It belongs to a service call or RPC, not to a log.

Source: https://developer.confluent.io/courses/apache-kafka/