Minor Description
Purgatory is the data structure in Kafka broker that manages delayed operations. There is a ConcurrentHashMap (Kafka Pool) maps each operation key to the operations (in a ConcurrentLinkedQueue) that are interested in the key.
When an operation is done or expired, it's removed from the list (ConcurrentLinkedQueue). When the list is empty, it's removed from the ConcurrentHashMap. The 2nd operation has to be protected by a lock, to avoid adding new operations into a list that is being removed. This is currently done by a globally shared ReentrantReadWriteLock. All the read operations on purgatory have to acquire the read permission of this lock. The list removing operations needs the write permission of this lock.
Our profiling result shows that Kafka broker is spending a nontrivial amount of time on this read write lock.
The problem is exacerbated when there are a large amount of short operations. For example, when we are doing sync produce operations (acks=all), a DelayedProduce operation is added and then removed for each message. If the QPS of the topic is not high, it's very likely that, when the operation is done and removed, the list of that key (topic partitions) also becomes empty, and has to be removed when holding the write lock. This operation blocks all the read / write operations on entire purgatory for awhile. As there are tens of IO threads accessing purgatory concurrently, this shared lock can easily become a bottleneck.
Actually, we only want to avoid concurrent read / write on the same key. The operations on different keys do not conflict with each other.
I suggest to shard purgatory into smaller partitions, and lock each individual partition independently.
Assuming there are 10 io threads actively accessing purgatory, sharding purgatory into 512 partitions will make the probability for 2 or more threads accessing the same partition at the same time to be about 2%. We can also use ReentrantLock instead of ReentrantReadWriteLock. When the read operations are not much more than write operations, ReentrantLock has lower overhead than ReentrantReadWriteLock.