Hello community,
I came across this blog post:
https://banzaicloud.com/blog/kafka-on-etcd/
And I thought it would be a good idea to discuss the criticism as a community.
Let me copy the points here and add some notes:
• Unlike Kafka it does not have a vibrant and huge community (merge
those PR’s please, anyone?)
I have personally met and worked with a lot of great people in this community
over the years, and as such, I probably have a pretty biased view. But, it is a
common concern that we are not fast enough at responding. We also don't have
conferences and large meetups compared to other communities. Are those really
necessary, though? What can we do to be a better community?
• It uses a protocol which is hard to understand and it’s hard to
maintain a large Zookeeper cluster
I can't really speak for the hard to understand part, and I don't understand
what "maintain a large ZooKeeper cluster" is referring to. How large is it and
why do we need it to be large? We have features like observers that enable
large clusters, but whether it solves the problem depends on what they are
after.
• It’s a bit outdated, compared say with Raft
When we wrote about Zab years back, we had as a goal to explain the protocol in
a way that could be reproduced. We had other goals too, like explaining how we
had been successful in implementing a system like ZooKeeper with that protocol,
the properties it guaranteed and so on. Raft focused on the simplicity of
understanding, which makes a lot of sense given that there was interest in
reproducing it. Given its focus, and clearly the quality of the people behind
it, Raft has been more successful in popularizing the implementation of
replicated state machines. At a protocol level, however, I don't think there is
anything that makes Zab outdated with respect to Raft.
• It’s written in Java (yes, it’s opinionated but this is a problem for
us as ZK is an infrastructure component)
This is arguable, there are pros and cons both ways.
• We run everything in Kubernetes and k8s by default has an in-built
Raft implementation, etcd
I can totally understand this point. No one wants to have to operate two
systems doing similar things. To consolidate operations, it clearly makes sense
to pick one. Ironically, this post talks about plugability, but Kubernetes does
not really give the option of using zk rather than etcd if that's what I want
to use.
• Linearizability (if there is a word like this) - check this
comparison chart
We do provide linearizable reads with sync(), although I understand that it is
arguable whether that is truly linearizable. There has been a long running
discussion about whether we should make sync() truly linearizable by making it
a first-class txn. Back in the day, we haven't done it because we wanted reads
to be fast, so we implemented it in a way that it didn't have to go through the
whole pipeline of request processors, but it still reaches out to the leader.
See the issue for more detail:
https://issues.apache.org/jira/browse/ZOOKEEPER-2136
• Performance and inherent scalability issues
I don't know if those experiments were done using a dedicated device to the txn
log, which is a well-known fact about zk's performance. Incremental
snapshotting is clearly a good way to reduce the amount of disk load for
snapshots, but I wonder whether that's really a primary concern given that
servers these days often have multiple devices.
I don't understand that max CPU utilization for zk
(https://coreos.com/blog/performance-of-etcd.html). Perhaps this is something
to be investigated.
• Client side complexity and thick clients
Due to the set of features we wanted to offer, we have indeed chosen this path.
• Lack of service discovery
I don't have a good sense of how many users are actually bothered by this. I
have heard complaints over time about service discovery with ZooKeeper, but I'm
not sure there was any conclusion about whether service discovery is a good use
case for such coordination systems, including etcd for that matter.
Any feedback?
Thanks,
-Flavio
