Kafka Streams GlobalKTable sychronization to applications

With working normal k-streams, kafka stores the offsets of each application on its internal offset topics. On a application restart, applications reprocess the topics depending on auto.offset.reset policy. This is indeed explained here.

I am using kafka-stream's GlobalKTable to replicate the data over applications. However I'm a bit confused about applications' restarts since it's not populated on applications whose id (StreamsConfig.APPLICATION_ID_CONFIG) does not change after restart (due to a deployment or a crash). Whenever I start a new instance of streams application with new id, GlobalKTable is populated.

A GlobalKTable is nothing different but a topic with log-compaction feature enabled. javadoc of creating a StreamsBuilder#globalTable states:

streamsBuilder.globalTable("some-topic", Materialized.as("kglobaltable-store"))

Note that GlobalKTable always applies "auto.offset.reset" strategy "earliest" regardless of the specified value in StreamsConfig.

Hence I expect, regardless of the application id, my streams applications read the kglobaltable-store topic from the start and populate store locally like this github issue. It seems the topic the javadoc refers is some-topic instead of kglobaltable-store.

Is this the intended behaviour for GlobalKTable? And additionally is there a retention policy on topics which are backing GlobalKTables?

This behaviour also results in stale data on kglobaltable-store topic when we have a retention policy on some-topic. An example would be as follows:

At time t0, let;

some-topic: (1, a) -> (2, b) -> (1, c)

kglobaltable-store: [(1, c), (2, b)]

After some time (2, b) is subject to retention, I start my streams application (with a new id) and my GlobalKTable only stores the record (1, c) If this is the case.

EDIT: I am using InMemoryKeyValueStore.