zookeeper、doozerd、etcd都有着相似的架构，这三者的服务节点都需要一个仲裁节点来操作，它们是强一致的，并提供各种操作原语。应用程序可以通过客户端lib库来构建分布式的系统。在一个单datacenter中，consul的server节点工作在一种简单的方式下，consul server需要一个仲裁操作，并提供强一致性。consul原生的支持多datacenter，就像多gossip系统链接sever节点和clients一样。

         如果任何这些系统都用于K/V值存储，它们都提供了相同的语义，读取是强一致性的，在面对网络分区的时候，牺牲一致性确保可用性，然而，当系统使用高级特性的时候这些差异更加明显。这些系统提供的语义对构建服务发现有很大作用。zookeeper只提供原始的K/V值存储，并要求开发人员自己构建自己的系统来提供服务发现功能。consul提供了一个坚固的服务发现框架，这样就提升了开发的工作效率。客户端简单的注册服务，然后使用DNS或者HTTP接口来发现服务。其他其他则需要你自己定制自己的解决方案。一个令人信服的服务发现框架必须包含健康检测和考虑失败的可能性。原生的系统使用心跳检测、周期性的更新和TTL来确保发现服务异常。这个系统需要知道工作节点的数量和固定数量服务器上的需求。此外，故障检测窗口需要TTL机制。zookeeper提供了短暂节点的K/V条目，当客户端断开链接则删除该条目。这是比心跳检测更复杂的系统，但是也有增加客户端难的问题。所有客户端必须维护到zookeeper服务的连接活跃，并发送活跃消息，而且，这种客户端比较厚重，很难编写，易出BUG。consul使用一个完全不同的体系进行健康检查。不只是在server节点，consul client运行在集群中的每一个节点上，这些clients是gossip pool的一部分，提供包括分布式健康监测的功能。gossip协议提供了一个高效的故障检测机制，可以扩展到任何集群规模，而没有任何工作集中在某台服务器上客户端也支持在本地进行更丰富的健康监测。而zookeeper的短暂节点是一个非常原始的活跃度检查。客户端可以检查web服务器的状态返回码，内存利用率、磁盘使用情况等等。consul clients暴露出了一个HTTP接口，避免像zookeeper一样暴露给客户端一些复杂的系统。consul提供一流的服务发现、健康检查、K/V存储、多数据中心服务。支持任何简单的K/V存储，所有这些其他系统都需要额外的工具和lib库。通过client节点，consul提供了一个简单的API接口。

What do Etcd, Consul, and Zookeeper do?
  - Service Registration:
    - Host, port number, and sometimes authentication credentials, protocols, versions
      numbers, and/or environment details.
  - Service Discovery:
    - Ability for client application to query the central registry to learn of service location.
  - Consistent and durable general-purpose K/V store across distributed system.
    - Some solutions support this better than others.
    - Based on Paxos or some derivative (i.e. Raft) algorithm to quickly converge to a consistent state.
    - Centralized locking can be based on this K/V store.
  - Leader Election:
    - Not to be confused with leader election within the quorum of Etcd/Consul nodes. This is an
      implementation detail that is transparent to the user. What we are talking about here is leader
      election among the services that are registered against Etcd/Consul.
    - Etcd tabled their leader election module until the API stabilizes.
  - Other non-standard use cases:
    - Distributed locking
    - Atomic broadcast
    - Sequence numbers
    - Pointers to data in eventually consistent stores.
 
- How do they behave in a distributed system?
  - All of the solutions under consideration are primarily CP systems in the CAP context.
    That is, they favor consistency over availability. This means that all nodes have a
    consistent view of written data but at the expense of availability in the event that
    a network partitions occurs (i.e. loss of node).
    - Some of these solutions will support "stale reads" in the event of node loss.
  - Each solution can work with only one node. It is generally advised that we have one etcd/
    consul per VM/physical host. We do not want to have an etcd/consul per container!
 
- Immediate problems that we are trying to solve:
  - Get and set dynamic configuration across a distributed system (e.g. things in moc.config.json):
    - This is perhaps the most pressing problem that we need to solve.
    - An SCM tool like Puppet/Anisble are great for managing static configurations but
      they are too heavy for dynamic changes.
  - Service registration:
    - We need to be able to spin up a track and have services make themselves visible
      via DNS.
    - This would be useful primarily outside of production where we would want to regularly
      spin up and destroy tracks.
    - That said, we don't have a highly-distributed and elastic architecture so we could get
      by without this for a while.
  - Service discovery:
    - Services must be able to determine which host to talk to for a particular service.
    - This may not be as important for production if we have a loadbalancer. In fact, a
      loadbalancer would be more transparent to our existing apps as they work at the IP level.
    - That said, we don't have a highly-distributed and elastic architecture so we could get
      by without this for a while.
 
- Features that we don't need for now:
  - Leader election. Many of our apps are currently not designed to scale horizontally.
    However, it should be noted that Consul has the ability to select a leader based on
    health checks.
 
- Problems that these tools are not designed to solve:
  - Load-balancing.
 
- Things that I've explored:
  - Etcd:
    - Basic info:
      - Service registration relies on using a key TTL along with heartbeating from the service
        to ensure the key remains available. If a services fails to update the key’s TTL, Etcd
        will expire it. If a service becomes unavailable, clients will need to handle the
        connection failure and try another service instance.
      - There would be a compelling reason to favor Etcd if we ever planned to use CoreOS
        but I don't see this happening anytime soon.
    - Pros:
      - Service discovery involves listing the keys under a directory and then waiting for
        changes on the directory. Since the API is HTTP based, the client application keeps a 
        long-polling connection open with the Etcd cluster.
      - Has been around for longer than Consul. 150% more github watches/stars.
      - 3 times as many contributors (i.e. more eyes) and forks on github.
    - Cons:
      - There are claims that the Raft implementation used by Etcd (go-raft) is not quite right (unverified).
      - Immature, but by the time its use is under consideration in production, it should
        have reached 1.0.
      - Serving DNS records from Etcd may require a separate service/process (verify):
        - http://probablyfine.co.uk/2014/03/02/serving-dns-records-from-etcd/
        - SkyDNS is essentially DNS on top of Etcd
 
  - Consul:
    - Pros:
      - Has more high-level features like service monitoring.
      - There is another project out of Hashicorp that will read/set environment variable
        for processes from Consul.
        - https://github.com/hashicorp/envconsul
      - Better documentation.
        - I had an easier time installing and configuring this over Etcd, not that Etcd was
          particularly hard. Docs make all the difference.
        - Stuff like this makes me want to shed a tear. I commend the KIDS at Hashicorp.
          - http://www.consul.io/docs/internals/index.html
      - You can make DSN queries directly against Consul agent! Nice! No need for SkyDNS or Helix
      - We can add arbitrary checks! Nice, if we are into that sort of thing.
      - Understands the notion of a datacenter. Each cluster is confined to datacenter but the
        cluster is able to communicate with other datacenters/clusters.
        - At Skybox, we might use this feature to separate docker tracks, even if they live on same host.
      - It has a rudimentary web UI:
        - http://demo.consul.io/ui/
    - Cons:
      - There are claims that Consul's implementation of Raft is better (unverified).
      - Immature. Even younger than Etcd (though there are no reason to believe that there are problems with it).
 
- Etcd and Consul similarities:
  - HTTP+JSON based API. Curl-able.
  - Docker containers can talk directly with Etcd/Consul over the docker0 interface (i.e. default gateway).
  - Atomic look-before-you-set:
    - Etcd: Compare-and-set by both value and version index.
    - Consul: Check-and-set by sequence number (ModifyIndex)
  - DNS TTLs can be set to something VERY low.
    - Etcd: supports TTL (time-to-live) on both keys and directories, which will be honoured:
      if a value has existed beyond its TTL
    - Consul: By default, serves all DNS results with a 0 TTL value
  - Has been tested with Jepsen (tool to simulate network partitions in distributed databases).
    - Results were not 100% for either but still generally promising.
    - https://news.ycombinator.com/item?id=7884640
  - Both work with Confd by Kelsey Hightower.
    - A tool that watches Etcd/Consul and modifies config files on disk.
    - https://github.com/kelseyhightower/confd
  - Long polling for changes:
    - Etcd: Easily listen for changes to a prefix via HTTP long-polling.
    - Consul: A blocking query against some endpoints will wait for a change to potentially
      take place using long polling.
 
- Things that I have not explored:
  - SkyDNS: Anyone have good input on this one?
  - Zookeeper: It seems mature but it would take a lot more work to make it work for us.
    - We would be have to configure and use it without high-level features.
    - Provides only a primitive K/V store.
    - Requires that application developers build their own system to provide service discovery.
    - Java dependency (and Dan Streit hates Java)
    - All clients must maintain active connections to the ZooKeeper servers, and perform keep-alives.
    - Zookeeper not recommended for virtual environments? Why? I just read this somewhere.
  - Corosync/Pacemaker (not sure if this is a viable solution, actually)
  - Redis is not viable! It is an in-memory K/V that does not persist data. Nope.
  - Smartstack + Synapse + Nerve from AirBnB (not viable as it only does TCP through HAproxy).
    - Ruby dependencies and many moving parts.
 
- References:
  http://www.hashicorp.com/blog/twelve-factor-consul.html   (heroku's excellent 12-factor thing).
  http://12factor.net/
  http://www.consul.io/intro/vs/index.html
  http://www.consul.io/docs/internals/index.html
  https://news.ycombinator.com/item?id=7604787
  https://news.ycombinator.com/item?id=7623317
  https://news.ycombinator.com/item?id=7884640
  http://www.activestate.com/blog/2014/03/brandon-philips-explains-etcd
  http://jpmens.net/2013/10/24/a-key-value-store-for-shared-configuration-etcd-confd/
  http://igor.moomers.org/smartstack-vs-consul/
  http://jasonwilder.com/blog/2014/02/04/service-discovery-in-the-cloud/
  http://nerds.airbnb.com/smartstack-service-discovery-cloud/