Feb 4, 2014 · 14 minute read · Comments
Service discovery is a key component of most distributed systems and service oriented architectures.
The problem seems simple at first: How do clients determine the IP and port for a service that
exist on multiple hosts? Usually, you start off with some static configuration which gets you pretty far.
Things get more complicated as you start deploying more services. With a live
system, service locations can change quite frequently due to auto or manual scaling,
new deployments of services, as well as hosts failing or being replaced. Dynamic service registration and discovery becomes much more important in these scenarios in
order to avoid service interruption. This problem has been addressed in many different ways and is continuing to evolve. We’re going to look at some open-source or openly-discussed solutions to this problem to understand how they work. Specifically,
we’ll look at how each solution uses strong or weakly consistent storage, runtime dependencies, client
integration options and what the tradeoffs of those features might be. We’ll start with some strongly consistent projects such as Zookeeper,
Doozer and Etcd which are typically
used as coordination services but are also used for service registries as well. We’ll then look at some interesting solutions specifically designed for service registration and discovery.
We’ll examine Airbnb’s SmartStack,
Netflix’s Eureka, Bitly’s NSQ,
Serf,
Spotify and DNS and finally SkyDNS. There are two sides to the problem of locating services. Service Registration and Service Discovery. Any service registration and discovery solution also has other development and operational aspects to consider: These first three registries use strongly consistent protocols and are actually general purpose, consistent
datastores. Although we’re looking at them as service registries, they are typically used for coordination
services to aid in leader election or centralized locking with a distributed set of clients. Zookeeper is a centralized service for maintaining configuration
information, naming, providing distributed synchronization, and providing group services. It’s written in Java, is strongly consistent (CP) and uses the Zab protocol to coordinate changes across the ensemble (cluster). Zookeeper is typically run with three, five or seven members in the ensemble. Clients use language
specific bindings in order to access the ensemble.
Access is typically embedded into the client applications and services. Service registration is implemented with ephemeral nodes
under a namespace. Ephemeral nodes only exist while the client is connected so typically a backend service registers itself, after startup, with its
location information. If it fails or disconnects, the node disappears from the tree. Service discovery is implemented by listing and watching the namespace for the service. Clients
receive all the currently registered services as well as notifications when a service becomes unavailable
or new ones register. Clients also need to handle any load balancing or failover themselves. The Zookeeper API can be difficult to use properly and language bindings might have subtle differences
that could cause problems.
If you’re using a JVM based language, the Curator Service Discovery Extension might be of some use. Since Zookeeper is a CP system, when a partition occurs,
some of your system will not be able to register or find existing registrations even if they could function properly during the
partition. Specifically, on any non-quorum side, reads and writes will return an error. Doozer is a consistent, distributed data store. It’s written in Go,
is strongly consistent and uses Paxos
to maintain consensus. The project has been around for a number of years but has stagnated for
a while and now has close to 160 forks. Unfortunately, this makes it
difficult to know what the actual state of the project is and whether is is suitable for production
use. Doozer is typically run with three, five or seven nodes in the cluster. Clients use language
specific bindings to access the cluster and, similar to Zookeeper, integration is embedded
into the client and services. Service registration is not as straightforward as with Zookeeper because Doozer does not have any
concept of ephemeral nodes. A service can register itself under a path but if the service becomes
unavailable, it won’t be removed automatically. There are a number of ways to address this issue. One option might be to add a timestamp and
heartbeating mechanism to the registration process and
handle expired entries during the discovery process or with another cleanup processes. Service discovery is similar to Zookeeper in that you can list all the entries under a path and
then wait for any changes to that path. If you use a timestamp and heartbeat during registration, you
would ignore or delete any expired entries during discovery. Like Zookeeper, Doozer is also a CP system and has the same consequences when a partition occurs. Etcd is a highly-available, key-value store for shared configuration and service discovery. Etcd
was inspired by Zookeeper and Doozer. It’s written in Go, uses Raft
for consensus and has a HTTP+JSON based API. Etcd, similar to Doozer and Zookeeper, is usually run with three, five or seven nodes in the cluster.
Clients use a language specific binding or implement one using an HTTP client. 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. 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. Since Etcd uses Raft, it
should be a strongly-consistent system. Raft requires a leader to be elected and all client requests are handled by
the leader. However, Etcd also seems to support reads from non-leaders using this undocumented consistent parameter which would improve
availabilty in the read case. Writes would still need to be handled by the leader during a partition and could
fail. These next few registration services and approaches are specifically tailored to service registration
and discovery. Most have come about from actual production use cases while others are interesting
and different approaches to the problem. Whereas Zookeeper, Doozer and Etcd could also be used for
distributed coordination, these solutions don’t have that capability. Airbnb’s SmartStack is a combination
of two custom tools, Nerve and Synapse
that leverage haproxy and Zookeeper to handle
service registration and discovery. Both Nerve and Synapse are written in Ruby. Nerve is a sidekick style process that runs as a separate process alongside the application service.
Nerve is reponsible for registering services in Zookeeper. Applications expose a The sidekick model eliminates the need for a service to interact with Zookeeper. It simply needs
a monitoring endpoint in order to be registered. This makes it much easier to support services in different
languages where robust Zookeeper binding might not exist. This also provides many of benefits of the
Hollywood principle. Synapse is also a sidekick style process that runs as a separate process alongside the service.
Synapse is responsible for service discovery. It does this by querying Zookeeper for currently
registered services and reconfigures a locally running haproxy instance. Any clients on the host
that need to access another service always accesses the local haproxy instance which will route the
request to an available service. Synapse’s design simplifies service implementations in that they do not need to implement any client
side load balancing or failover and they do not need to depend on Zookeepr or its language bindings. Since SmartStack relies on Zookeeper, some registrations and discovery may fail during a partition.
They point out that Zookeepr is their “Achilles heel” in this setup.
Provided a service has been able to discover the other services, at least once, before a partition, it should still
have a snapshot of the services after the partition and may be able to continue operating during the
partition. This aspect improves the availability and reliability of the overall system. Update: If you’re intested in a SmartStack style solution for docker containers, check out docker service discovery. Eureka is Netflix’s middle-tier, load balancing and discovery
service. There is a server component as well as a smart-client that is used within application
services. The server and client are written in Java which means the ideal use case would be for
the services to also be imlemented in Java or another JVM compatible language. The Eureka server is the registry for services. They recommend running one Eureka server in each
availability zone in AWS to form a cluster. The servers replicate their state to each
other through an asynchronous model which means each instance may have a slightly, different picture
of all the services at any given time. Service registration is handled by the client component. Services embed the client in their application
code. At runtime, the client registers the service and periodically sends heartbeats to renew its leases. Service discovery is handled by the smart-client as well. It retrieves the current registrations from the
server and caches them locally. The client periodically refreshes its state and also handles load
balancing and failovers. Eureka was designed to be very resilient during failures. It favors availabilty over
strong consistency and can operate under a number of different failure modes. If there is a partition within the cluster,
Eureka transitions to a self-preservation state. It will allow services to be discovered and registered
during a partition and when it heals, the members will merge their state again. NSQ is a realtime, distributed messaging platform. It’s written in Go
and provides an HTTP based API. While it’s not a general purpose service registration and discovery tool,
they have implemented a novel model of service discovery in their
nsqlookupd agent
in order for clients to find nsqd instances at
runtime. In an NSQ deployment, the nsqd instances are essentially the service. These are the message stores.
nsqlookupd is the service registry. Clients connect directly to nsqd instances but since these may
change at runtime, clients can discover the available instances by querying nsqlookupd instances. For service registration, each nsqd instance periodically sends a heartbeat of its state to each nsqlookupd
instance. Their state includes their address and any queues or topics they have. For discovery, clients query each nsqlookupd instance and merge the results. What is interesting about this model is that the nsqlookupd instances do not know about each other.
It’s the responsibility of the clients to merge the state returned from each stand-alone nsqlookupd instance to
determine the overal state. Because each nsqd instance heartbeats its state, each nsqlookupd eventually
has the same information provided each nsqd instance can contact all available nsqlookupd instances. All the previously discussed registry components all form a cluster and use strong or weakly consistent
consensus protocols to maintain their state. The NSQ design is inherently weakly consistent but
very tolerant to partitions. Serf is a decentralized solution for service discovery and orchestration. It is also
written in Go and is unique in that uses a gossip based protocol, SWIM for membership, failure detection and custom event propogation. SWIM was designed to address the unscalability of traditional heart-beating protocols. Serf consists of a single binary that is installed on all hosts. It can be run as an agent, where it
joins or creates a cluster, or as a client where it can discover the members in the cluster. For service registration, a serf agent is run that joins an existing cluster. The agent is started
with custom tags that can identify the hosts role, env, ip, ports, etc. Once joined to the cluster, other
members will be able to see this host and it’s metadata. For discovery, serf is run with the Serf is a relatively new project and is evolving quickly. It is the only project in this post that
does not have a central registry architectural style which makes it unique. Since it uses a asynchronous, gossip based protocol, it is inherently weakly-consistent but more fault tolerant and available. Spotify described their use of DNS for service discovery in their post
In praise of “boring” technology. Instead of using a newer, less
mature technology they opted to build on top of DNS. Spotify views DNS as a
“distributed, replicated database tailored for read-heavy loads.” Spotify uses the relatively unknown SRV record which is intended
for service discovery. SRV records can be thought of as a more generalized MX record. They allow you
to define a service name, protocol, TTL, priority, weight, port and target host. Basically, everything
a client would need to find all available services and load balance against them if necessary. Service registration is complicated and fairly static in their setup since they manage all zone files
under source control. Discovery uses a number of different DNS client librarires and custom tools. They
also run DNS caches on their services to minimize load on the root DNS server. They mention at the end of their post that this model has worked well for them but they are starting
to outgrow it and are investigating Zookeeper to support both static and dynamic registration. SkyDNS is a relatively new project that is written in Go,
uses RAFT for consensus and also provides a client API over HTTP and DNS. It has some
similarities to Etcd and Spotify’s DNS model and actually uses the same RAFT implementation as Etcd,
go-raft. SkyDNS servers are clustered together, and using the RAFT protocol, elect a leader. The SkyDNS servers
expose different endpoints for registration and discovery. For service registration, services use an HTTP based API to create an entry with a TTL. Services must
heartbeat their state periodically. SkyDNS also uses SRV records but extends them to also support
service version, environment, and region. For discovery, clients use DNS and retrieve the SRV records for the services they need to contact.
Clients need to implement any load balancing or failover and will likely cache and refresh service
location data periodically. Unlike Spotify’s use of DNS, SkyDNS does support dynamic service registration and is able to do this
without depending on another external service such as Zookeeper or Etcd. If you are using docker, skydock might be worth checking out to integrate your containers with SkyDNS automatically. Overall, this is an interesting mix of old (DNS) and new (Go, RAFT) technology and will be interesting
to see how the project evolves. We’ve looked at a number of general purpose, strongly consistent registries (Zookeeper, Doozer, Etcd)
as well as many custom built, eventually consistent ones (SmartStack, Eureka, NSQ, Serf, Spotify’s DNS, SkyDNS). Many use embedded client libraries (Eureka, NSQ, etc..) and some use separate sidekick processes
(SmartStack, Serf). Interestingly, of the dedicated solutions, all of them have adopted a design that
prefers availability over consistency. (1) If using the (2) If using a caching DNS client in front of SkyDNS, reads could be inconsistent
service discovery doozer etcd zookeeper architecture smartstack nsq skydns
The Problem
General Purpose Registries
Zookeeper
Doozer
Etcd
Single Purpose Registries
Airbnb’s SmartStack
/health endpoint,
for HTTP services, that Nerve continuously monitors. Provided the service is available, it will be
registered in Zookeper.Netflix’s Eureka
Bitly’s NSQ lookupd
Serf
members command which returns the current members of the
cluster. Using
the members output, you can discover all the hosts for a service based on the tags their agent is
running.Spotify and DNS
SkyDNS
Summary
Name
Type
AP or CP
Language
Dependencies
Integration
Zookeeper
General
CP
Java
JVM
Client Binding
Doozer
General
CP
Go
Client Binding
Etcd
General
Mixed (1)
Go
Client Binding/HTTP
SmartStack
Dedicated
AP
Ruby
haproxy/Zookeeper
Sidekick (nerve/synapse)
Eureka
Dedicated
AP
Java
JVM
Java Client
NSQ (lookupd)
Dedicated
AP
Go
Client Binding
Serf
Dedicated
AP
Go
Local CLI
Spotify (DNS)
Dedicated
AP
N/A
Bind
DNS Library
SkyDNS
Dedicated
Mixed (2)
Go
HTTP/DNS Library
consistent parameter, inconsistent reads are possible