search

python-zk / kazoo

Kazoo is a high-level Python library that makes it easier to use Apache Zookeeper.
https://kazoo.readthedocs.io
Apache License 2.0
1.3k stars 386 forks source link

TreeCache recipe creating heavy load on ZK while reconnecting. #664

Open Buffer0x7cd opened 2 years ago

Buffer0x7cd commented 2 years ago

The TreeCache recipe seems to generate a large amount of traffic while reconnecting to the ZK node ( This usually happens when there is a leader election in the cluster, which forces every client to get disconnected from the cluster and then go into a reconnection loop, which lasts until the leader election finishes. )

https://github.com/python-zk/kazoo/blob/9bb849941deb1d62ca3ca4d74882ed41cf569513/kazoo/recipe/cache.py#L210

Here we can see that the TreeCache tries to reload the entire Tree after getting reconnected to the ZK node. To test this, we ran a 3 node ZK cluster with 5k znodes and around 200 TreeCache clients (The clients use TreeCache to maintain the view of ZK data in memory ). During the leader election process, we noticed a 40 to 50x increase in read traffic (read traffic going from 600rps to 30k rps). This ends up causing a thundering herd problem for the cluster.

After some discussion, we came up with a few ideas that can probably help us to avoid the situation

  1. Instead of reloading the entire Tree after a reconnect event, Only reload the znodes that have been updated between client disconnect and reconnect event. Zk guarantees that watches for the znodes will be triggered once clients get reconnected. we can use this to do a selective update of the znodes.
  2. Introduce a client-side rate limiting which can help us to smooth the traffic burst after a reconnection. Clients should be able to use this to avoid overwhelming the cluster.

Let me know what you think about the issue.