Major Description
This is the same bug as KAFKA-7912 which was then re-introduced by KAFKA-8802.
Any iterator returned from InMemoryKeyValueStore may end up throwing a ConcurrentModificationException because the backing map is not concurrent safe. I expect that this only happens when the store is retrieved from KafkaStreams.store() from outside of the topology since any usage of the store from inside of the topology should be naturally single-threaded.
To start off, a reminder that this behaviour explicitly violates the interface contract for ReadOnlyKeyValueStore which states
The returned iterator must be safe from java.util.ConcurrentModificationExceptions
It is often complicated to make code to demonstrate concurrency bugs, but thankfully it is trivial to reason through the source code in InMemoryKeyValueStore.java to show why this happens:
- All of the InMemoryKeyValueStore methods that return iterators do so by passing a keySet based on the backing TreeMap to the InMemoryKeyValueIterator constructor.
- These keySets are all VIEWS of the backing map, not copies.
- The InMemoryKeyValueIterator then makes a private copy of the keySet by passing the original keySet into the constructor for TreeSet. This copying was implemented in
KAFKA-8802, incorrectly intending it to fix the concurrency problem. - TreeSet then iterates over the keySet to make a copy. If the original backing TreeMap in InMemoryKeyValueStore is changed while this copy is being created it will fail-fast a ConcurrentModificationException.
This bug should be able to be trivially fixed by replacing the backing TreeMap with a ConcurrentSkipListMap but here's the rub:
This bug has already been found in KAFKA-7912 and the TreeMap was replaced with a ConcurrentSkipListMap. It was then reverted back to a TreeMap in KAFKA-8802 because of the performance regression. I can see from one of the PRs that it was believed the concurrency problem with the TreeMap implementation was fixed by copying the keyset when the iterator is created but the problem remains, plus the fix creates an extra copy of the iterated portion of the set in memory.
For what it's worth, the performance difference between TreeMap and ConcurrentSkipListMap do not extend into complexity. TreeMap enjoys a similar ~2x speed through all operations with any size of data, but at the cost of what turned out to be an easy-to-encounter bug.
This is all unfortunate since the only time the state stores ever get accessed concurrently is through the `KafkaStreams.store()` mechanism, but I would imagine that "correct and slightly slower) is better than "incorrect and faster".
Too bad BoilerBay's AirConcurrentMap is closed-source and patented.
Attachments
Issue Links
- links to