● This works from (almost) any host in Yelp
● This works from Python, Java, command line etc.
● If a service supports HTTP or TCP then it can be made discoverable.
○ This includes third-party services such as MySQL and scribe
● It’s dynamic: for a given service, if new instances are added then they
will automatically become available.
Very Important Things to Note
● SmartStack (nerve and synapse) were written by Airbnb
● We’ve added some features
● The work here has been carried out by many people across Yelp
Nerve registers service instance in ZooKeeper:
│ └── server_1_0000013614
│ └── server_1_0000000959
│ ├── server_1_0000002468
│ └── server_2_0000002467
The data in a znode is all that is required to connect to the corresponding
We’ll shortly see how this is used for discovery.
Normally hacheck just acts as a transparent proxy for our healthchecks:
$ curl -s yocalhost:6666/http/service_1/1234/status | jq .
We can also use it to fail healthchecks before we shut down a service.
This allows us to gracefully shutdown a service.
(Also provides a 1s cache to limit healthcheck rate.)
$ hadown service_1
$ curl -v yocalhost:6666/http/service_1/1234/status
Service service_1 in down state since 1443217910: billings
How do we know what services to advertise? Every service host
periodically runs a script to regenerate the nerve configuration, reading
from the following sources:
● mesos slave API
● By default bind to 0.0.0.0
● Bind only to yocalhost on public servers.
● HAProxy gives us a lot of goodies for all clients:
○ Redispatch on connection failures
○ Zero-downtime restarts (once you know how :)
○ Easy to insert connection logging
● Each host also exposes an HAProxy status page for easy introspection
Every client host periodically runs a script to regenerate the synapse
configuration, reading service definitions from yelpsoa-configs.
For each service reads a smartstack.yaml file.
Restarts synapse if configuration has changed.
Some client libraries like to do their own load balancing e.g. cassandra,
memcached. Use synapse to dump the registration information to disk:
$ cat /var/run/synapse/services/devops.demo.json | jq .
docker container 1
docker container 2
● We’d like to run only one nerve / synapse / haproxy per host
● What address should we bind haproxy to?
● 127.0.0.1 won’t work from within a container
● Instead we pick a link-local address 169.254.255.254 (yocalhost)
● This also works on servers without docker
We run services in both our own datacenters as well as AWS.
We logically group these environments according to latency.
Service authors get to decide how ‘widely’ their service instances are
Everything is controlled via smartstack.yaml files.
or AZs in AWS
Regions within 5ms
round-trip e.g. ‘pacific
advertise / discover
Synapse should look in the
habitat directory in its local
Nerve should register this
service in the habitat directory
of its local ZooKeeper
“Wouldn’t it be useful if we could make a service running in datacenter A
available in an (arbitrary) datacenter B?”
● Makes it easier to bring up a new datacenter
● Makes it easier to add more capacity to a datacenter in an emergency
● Makes it easier to keep a datacenter going in an emergency if a service
● Lots of little components, each doing doing one thing well
● Very simple interface for clients and services
○ If it speaks TCP or HTTP we can register it
● Easy to independently replace components
○ HAProxy -> NGINX?
● Easy to observe behavior of components
It’s OK if ZooKeeper fails
● Nerve and Synapse keep retrying
● HAProxy keeps running but with no updates
● HAProxy performs its own healthchecks against service instances
○ If a service instance becomes unavailable then it will stop receiving
traffic after a short period
● The website stays up :)
Does it blend scale?
● Used to have scaling issues with internal load balancers, this is not a
problem with SmartStack :)
● Hit some scaling issues at 10s of thousands of ZooKeeper connections
○ Addressed this by using just a single ZooKeeper connection from
each nerve and synapse
● Used to have lots of HAProxy healthchecks hitting services
○ hacheck insulates services from this
○ We limit HAProxy restart rate
What about etcd / consul / …?
● We try to use boring components :)
● We’re already using Zookeeper for Kafka and ElasticSearch so it’s
natural to use it for our service discovery system too.
● etcd would probably also work, and is supported by SmartStack
● Conceptually similar to consul / consul-template
What about DNS?
● What TTL are you going to use?
● Are you clients even going to honor the TTL?
● Does the DNS resolution happen inline with requests?
● We’ve used SmartStack to create a robust service discovery system
● It’s UNIXy: lots of separate components, each doing one thing well
● It’s flexible: locality-aware discovery
● It’s reliable: new devs at Yelp view discovery as a solved problem
● It’s useful: SmartStack is the glue that holds our SOA together