On the JVM, Array.equals and Array.hashCode do not incorporate array contents; they inherit Object.equals/hashCode. This implies that Collections that rely upon equals/hashCode (eg, HashMap/HashSet and variants) treat two arrays with equal contents as distinct elements.
Many of the Kafka Streams internal classes currently use generic HashMaps and Sets to manage caches and invalidation status. For example, RocksDBStore.cacheDirtyKeys is a HashSet<K>. Then, in RocksDBWindowStore, the Elements are constructed as RocksDBStore<byte, byte>.
Similarly, the MemoryLRUCache<K, RocksDBCacheEntry> internally holds a LinkedHashMap<K,V> map, and a HashSet<K> keys, and these end up holding byte keys. Finally, user-code may attempt to use any of these provided types with byte, with undesirable results.
Keys that are byte-arrays should be wrapped in a type that incorporates the content in their computation of equals/hashCode. java.nio.ByteBuffer is one such type that could be used, but a purpose-built immutable class would likely be a better solution.