Skip to content

On-zone DNS registration agent (see also: binder)

License

Notifications You must be signed in to change notification settings

TritonDataCenter/registrar

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

98 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Registrar

This repository is part of the Joyent Manta and Triton projects. For contribution guidelines, issues, and general documentation, visit the main Triton and Manta project pages.

Table of Contents:

Service discovery in Triton and Manta

Triton and Manta components generally discover their dependencies through DNS. There are three main components that make up this service discovery system:

  • a ZooKeeper cluster, which keeps track of the list of instances of each different type of component
  • Registrar (this component), a small agent that runs alongside most Triton and Manta components to register that component's presence in ZooKeeper
  • Binder, a server that answers DNS queries using the ZooKeeper state as the backing store

Let's examine what happens when an operator deploys a new instance of the Manta service called "authcache". We'll assume the deployment uses DNS suffix "emy-10.joyent.us":

  1. The operator provisions a new instance of "authcache". This creates a new SmartOS zone (container). The new container gets a uuid. In this example, the uuid is "a2674d3b-a9c4-46bc-a835-b6ce21d522c2".
  2. When the new zone boots up, various operating system services start up, including Registrar (this component).
  3. Registrar reads its configuration file, which specifies that it should register itself at "authcache.emy-10.joyent.us".
  4. Registrar connects to the ZooKeeper cluster and inserts a ZooKeeper ephemeral node called "/us/joyent/emy-10/authcache/a2674d3b-a9c4-46bc-a835-b6ce21d522c2". The contents of this node are a JSON payload that includes the IP address of this zone, as well as the ports it's listening on. The ZooKeeper protocol requires that Registrar periodically heartbeat to the cluster in order to maintain its session and keep this ephemeral node alive.
  5. Some time later, a client of authcache does a DNS query for "authcache.emy-10.joyent.us" using its configured nameservers, which are running instances of Binder. Assuming Binder doesn't have the answer to this query cached, it fetches the state out of ZooKeeper under "/us/joyent/emy-10/authcache". From this, it finds the IP addresses and ports of the authcache instances, including those of our newly-provisioned zone "a2674d3b-a9c4-46bc-a835-b6ce21d522c2". Binder caches this information for subsequent queries.
  6. Binder translates this information into the appropriate DNS answers (usually "A" records or "SRV" records, depending on what the DNS client asked for).

If the zone is destroyed, or the server on which it's running reboots or powers off or becomes partitioned, Registrar will become disconnected from ZooKeeper and its session will expire. This causes its ephemeral node to disappear from ZooKeeper. Once the Binder caches expire and they re-fetch the state from ZooKeeper, they will no longer find information about the zone that's gone, so they will stop including that zone in DNS answers. Clients will shortly stop using the zone.

In this way:

  • clients of a service (like "authcache") discover instances using DNS
  • instances are added to DNS automatically when they start up
  • instances are removed from DNS automatically for many (but not all) failures

Note that even in the best of cases, there is a non-trivial delay between when an instance fails and when clients see that and stop using it. For the instance to fall out of DNS, Registrar's ZooKeeper session timeout must expire (usually 30-40 seconds), Binder's cache must expire (currently another 60 seconds), and the DNS records fetched by clients must also expire (usually another 30-60 seconds, but this is configurable per-service, and it may actually take much longer than that). Clients must still deal with failures of instances that are present in DNS. DNS is the way that clients discover what instances they might be able to use, but it should not be assumed that all those instances are currently operating.

Health checking: Registrar runs separately from the program that actually provides service (the actual authentication cache process, in this example). So it's also possible for the real service to crash or go offline while registrar is still online. Registrar supports basic health checking by executing a command periodically and unregistering an instance if the command fails too many times in too short a period. However, as of this writing, this mechanism is extremely buggy. See HEAD-2282 and HEAD-2283.

SRV-based discovery: Many services (like Moray) use multiple processes in order to make use of more than one CPU. The recommended way to implement this is to configure registrar to publish specific port information. This allows Binder to answer queries with SRV records, which allow clients to discover not just individual zones, but the specific ports available in those zones.

Operating Registrar

Configuration

The configuration file is almost always immutable and based on a template that's checked into the repository of a component that uses registrar. Details are described under "Developing with Registrar" below.

Removing instances from service

There are many reasons why it's useful to remove an instance from DNS so that clients stop using it:

  • in development, this is useful to direct traffic at particular instances, or to test client behavior when instances come and go
  • in production, this is useful to isolate malfunctioning instances. You can remove these instances from service without actually halting them so that you can continue debugging them.

The usual way to do this is to disable the registrar SMF service in the zone you want to remove from DNS:

svcadm disable -st registrar

You can bring this zone back into DNS using:

svcadm enable -s registrar

Developing with Registrar

Incorporating Registrar into a component

As mentioned above, Registrar is deployed in nearly all Triton and Manta component zones. Incorporating Registrar into a new component usually involves only a few steps:

  • the component's repository should include a template configuration file, usually in sapi_manifests/registrar/template and an associated config-agent snippet in the same directory
  • the Mountain Gorilla (build system) configuration snippet for this component should depend on the registrar tarball

In the most common case, the Registrar configuration file is itself generated by config-agent using the template that's checked into the repository. In some cases (notably Moray), there's an additional step either at build time or zone setup time where the template itself is templatized by the list of TCP ports that should be exposed.

When new instances (zones) of the component start up, config-agent writes the registrar configuration file, populating variables such as the DNS domain suffix ("emy-10.joyent.us" in the example above) from the SAPI configuration for the current deployment. After that, Registrar starts up, reads the configuration file, and runs through the process described at the top of this README.

Configuration reference

The configuration file is specified using the "-f" argument to "main.js". The file is a JSON object with top-level properties:

Property Type Description
"adminIp" optional string IPv4 address to include in DNS records
"zookeeper" object describes how to connect to the ZooKeeper cluster
"registration" object describes which DNS names should be created for this component

adminIp: The address for all DNS answers related to this component is whatever is provided by the "adminIp" configuration property. This is the IP address used by clients that use DNS to discover this component. For Triton components, this should usually be an addresss on the "admin" network (hence the name). For Manta components, this should usually be an address on the "manta" network. If "adminIp" is not specified in the configuration, then Registrar picks an up, non-loopback IP address on the system and uses that, but this is not recommended.

zookeeper: Service discovery records are maintained in a ZooKeeper cluster. The "zookeeper" top-level property describes how to reach that cluster. This should be a configuration block appropriate for node-zkplus. See that project for details, but there's an example below that includes "timeout" and "servers" properties.

registration: The "registration" object describes the service discovery records that will be inserted into ZooKeeper. These control the DNS names that are available for this component. Broadly, there are two types of service discovery records:

  • Host records essentially allow Binder to answer DNS "A" and "SRV" queries with the IP address (and possibly port numbers) for a single instance. More precisely, host records are individual nodes in the ZooKeeper namespace that provide address and port information for a single zone. These are ephemeral, which means they disappear when Registrar's ZooKeeper session expires. That's by design so that if a zone disappears, it stops showing up in DNS.
  • Service records allow Binder to answer DNS "A" and "SRV" queries for a single logical service that's provided by any number of interchangeable instances. The list of instances available are represented by host records that are child nodes of the service record (within the ZooKeeper namespace).

Using host and service records

Let's look at an example. In Manta, instances of the "authcache" service publish host records under "$zonename.authcache.$suffix". As a result, if you have a Manta deployment whose DNS suffix is "emy-10.joyent.us", you can find the IP address for the "authcache" zone that's called "a2674d3b-a9c4-46bc-a835-b6ce21d522c2" by looking up "a2674d3b-a9c4-46bc-a835-b6ce21d522c2.authcache.emy-10.joyent.us":

$ dig +nocmd +nocomments +noquestion +nostats a2674d3b-a9c4-46bc-a835-b6ce21d522c2.authcache.emy-10.joyent.us
a2674d3b-a9c4-46bc-a835-b6ce21d522c2.authcache.emy-10.joyent.us. 30 IN A 172.27.10.62

We've just looked up the host record for a particular authcache instance. The authcache service also writes a service record for the higher-level DNS name "authcache.emy-10.joyent.us". This lets clients of the authcache service use the DNS name "authcache.emy-10.joyent.us" to find all available authcache instances:

$ dig +nocmd +nocomments +noquestion +nostats authcache.emy-10.joyent.us
authcache.emy-10.joyent.us. 30  IN      A       172.27.10.67
authcache.emy-10.joyent.us. 30  IN      A       172.27.10.62

Summary: A service can provide host records (when there's only one IP address for a given DNS name) or service records (when there may be multiple interchangeable instances).

Configuring the registration

The registration block of the configuration file determines which records are created. This block contains properties:

Property Type Description
"domain" string DNS name under which records will be created for this instance
"aliases" optional array of string array of fully-qualified DNS names to create as additional host records for this instance
"type" string the specific subtype of record to use for the host records created for this instance
"service" optional object if present, a service record will be created with properties described by this object (see below)
"ttl" optional number if present, this may be used for the TTL of the host record. See "Using TTLs" below.

With this information, Registrar creates the following records:

  • A host record is always created at $(hostname).$domain. The use of $(hostname) here refers to the system's hostname (see hostname(1)) and $domain refers to the configuration property above.
  • If the aliases array is present, then additional host records are created for each string in the array. These should be fully qualified — they should generally end with the value of domain.
  • If service is present, a service record is created at the DNS name $domain itself. The service object is described below.

Registration of host records

All records — host records and service records — internally have a specific type. The "type" property above controls the types used for the host records that Registrar creates. (Service records always have type "service", and any type other than "service" indicates a host record.) The specific "type" determines exactly how Binder uses them. The following types are supported:

Type Can be queried directly? Can be used for Service?
"db_host" yes no
"host" yes no
"load_balancer" yes yes
"moray_host" yes yes
"ops_host" no yes
"redis_host" yes yes
"rr_host" no yes

For types that cannot be queried directly ("ops_host" and "rr_host"), if you query the corresponding DNS name, Binder will behave as though they weren't there. This is not generally useful in new components.

For types that cannot be used as a service ("db_host" and "host"), if these records are found as child nodes of a "service" record, they will not be included in the DNS results for the service itself. This is not generally useful in new components.

In a simpler world, all host record types could be queried directly (meaning that when you look up a DNS name that maps to a host record of that type, Binder answers with the address information in that record), and they could also be used as backing hosts for a "service" record. For historical reasons, that's not true, and it's not easy to change because there are services (notably "webapi" and "loadbalancer") that share a DNS name today, but where only one of them is intended to be enumerable.

Summary: The most common case is that each instance of a component is interchangeable, and clients can talk to any one of them. In that case, you should use host records of type "load_balancer" and separately configure a "service" record. This will cause "domain" to be a DNS name that lists all of the active instances' IP addresses, and $zonename.$domain can be used to find the address of specific instances when that's needed (mostly for debugging). On the other hand, if you want to create standalone host records that aren't part of a logical service, use type "host" and do not create an associated service record. This is not common.

Example: Here's an example Registrar configuration:

{
    "registration": {
        "type": "load_balancer",
        "domain": "example.joyent.us",
        "aliases": [
            "host-1a.example.joyent.us",
            "host-1b.example.joyent.us"
        ]
    },
    "adminIp": "172.27.10.72",
    "zookeeper": {
        "timeout": 60000,
        "servers": [ { "host": "172.27.10.35", "port": 2181 },
                     { "host": "172.27.10.32", "port": 2181 },
                     { "host": "172.27.10.33", "port": 2181 } ]
    }
}

In these and subsequent examples, "172.27.10.72" is the IP address of the Registrar instance on the network on which we're providing the "example.joyent.us" service. The "zookeeper" block is specific to this deployment and points Registrar at the ZooKeeper cluster.

This example specifies that we're registering an instance of the service "example.joyent.us". This will create three host records: one for the Registrar instance's hostname (which is "b44c74d6" in this case) and one for each of the two aliases "host-1a.example.joyent.us" and "host-1b.example.joyent.us". We can look up any of these:

$ dig host-1a.example.joyent.us +short
172.27.10.72
$ dig host-1b.example.joyent.us +short
172.27.10.72
$ dig b44c74d6.example.joyent.us +short
172.27.10.72

This configuration did not specify "service", so there are no service-level records, so there's no way to list all of the hosts under "example.joyent.us". (As of this writing, if you try to lookup "example.joyent.us", Binder will crash because of MANTA-3058.)

Registration of service records

As mentioned above, service records are used for two purposes:

  • to indicate that a particular DNS name is served by any of several interchangeable instances
  • to provide service, protocol, and port information so that Binder can answer SRV queries

To have Registrar create a service record, specify a "service" property under the "registration" object. The "service" object must have a property "type" with value "service" and another "service" object with properties:

Property Type Meaning
"srvce" string service to use for SRV answers. (The name "srvce" is correct, not a typo.)
"proto" string protocol to use for SRV answers
"port" number port to use for SRV answers when a child host record does not contain its own array of ports.
"ttl" optional number TTL to use for SRV answers. See "Using TTLs" below.

Note that the presence of "service" causes Registrar to create a service record, and the various fields are required, so it's not possible to specify a service record without also providing the information required to answer SRV queries.

Let's augment the configuration above to specify a service record:

{
    "registration": {
        "type": "load_balancer",
        "domain": "example.joyent.us",
        "service": {
            "type": "service",
            "service": {
                "srvce": "_http",
                "proto": "_tcp",
                "port": 80
            }
        }
    },
    "adminIp": "172.27.10.72",
    "zookeeper": {
        "timeout": 60000,
        "servers": [ { "host": "172.27.10.35", "port": 2181 },
                     { "host": "172.27.10.32", "port": 2181 },
                     { "host": "172.27.10.33", "port": 2181 } ]
    }
}

(We also dropped the aliases from this example because those were just for demonstration.)

With the service configuration in place, we can still look up the IP address for a specific instance:

$ dig b44c74d6.example.joyent.us +short
172.27.10.72

but we can also list all instances:

$ dig example.joyent.us +short
172.27.10.72

If we start up another Registrar instance with a similar configuration with IP address 172.27.10.73, then we'd get both results:

$ dig example.joyent.us +short
172.27.10.72
172.27.10.73

Note that because the service record configuration includes port numbers, Binder can now answer SRV queries as well. To make these queries with dig(1), we specify -t SRV (to ask for SRV answers) and specify a DNS name starting with _http._tcp. (because it's conventional for SRV requests to prepend the service and protocol names to the DNS name that you'd otherwise be using for the service):

$ dig -t SRV +nocmd +nocomments +noquestion +nostats _http._tcp.example.joyent.us
_http._tcp.example.joyent.us. 60 IN     SRV     0 10 80 b44c74d6.example.joyent.us.
b44c74d6.example.joyent.us. 30  IN      A       172.27.10.72

SRV queries allow clients to discover each of several instances running inside a container. This is the preferred approach for new services because it eliminates the need for a local loadbalancer, which improves performance and availability and also simplifies debugging.

ZooKeeper data format

This section describes the structure of service discovery records that are stored in ZooKeeper. You don't need this unless you're debugging or developing Registrar or ZooKeeper.

The service discovery information in ZooKeeper is always written by Registrar and read by Binder. It's thus a contract between these components. However, it was historically not documented, and several pieces are redundant or confusing. Additionally, this information is not thoroughly validated in Binder.

Caveat: This information is provided for reference only. The existing implementation is not crisp enough, validated enough, or committed enough to use this information to write an alternate implementation and expect that it will interoperate with the existing one.

Before reading this section, be sure to read and understand the "Configuration reference" section. It covers the basic underlying concepts that are used in this section.

ZooKeeper paths

ZooKeeper provides a filesystem-like API: there's a hierarchical, slash-delimited namespace of objects. Data about DNS domains is stored into ZooKeeper in paths derived from the domains by reversing the components of the domain and replacing dots (".") with slashes ("/"). So the information for domain "authcache.emy-10.joyent.us" is contained under "/us/joyent/emy-10/authcache" in the ZooKeeper namespace.

For a service like authcache, the typical ZooKeeper node structure looks like this:

  • "/us/joyent/emy-10/authcache" contains the service record for "authcache".
  • Nodes underneath this path (like "/us/joyent/emy-10/authcache/a2674d3b-a9c4-46bc-a835-b6ce21d522c2") contain host records for individual instances of "authcache".

The ZooKeeper analog of directory nodes can themselves contain data, so the node at "/us/joyent/emy-10/authcache" acts as both an object and a directory.

Overview of service discovery records

All of the ZooKeeper nodes written by Registar contain JSON payloads. We call these service discovery records. Internally, every service discovery record includes:

  • a required "type", a string identifying the specific type of this record.
  • a required property with the same name as the type that provides type-specific details, described below. For example, if the type has value "service", then there will be a top-level property called "service" that contains more information. If the type is "moray_host", the top-level property with the rest of the details will be called "moray_host".

There are broadly two kinds of records: host records and service records. As described above, host records indicate that a DNS name maps to a particular host (usually a zone or container). Service records indicate that a DNS name is served by one or more other hosts that are specified by child nodes in the ZooKeeper tree. Binder will reply to DNS requests with information about all of the hosts that it finds underneath a "service" record.

If we query Binder for A records for authcache.emy-10.joyent.us, we expect to get a list of IP addresses for the various instances of authcache. We expect we can connect to any of these instances on some well-known port to use the authcache service. How does this work?

When we query Binder for authcache.emy-10.joyent.us, assuming the result is not cached, Binder fetches the ZooKeeper node at "/us/joyent/emy-10/authcache". There, it finds a service record (with "type" == "service"):

{
  "type": "service",
  "service": {
    "type": "service",
    "service": {
      "srvce": "_redis",
      "proto": "_tcp",
      "port": 6379,
      "ttl": 60
    },
    "ttl": 60
  }
}

Seeing a service record, Binder then lists the children of the ZooKeeper node "/us/joyent/emy-10/authcache" to find host records for individual instances of the authcache service. (Remember, ZooKeeper's namespace looks like a filesystem, but the nodes that you'd think of as directories can themselves also contain data. In this case, the data at "/us/joyent/emy-10/authcache" is the service record. The child nodes in that directory describe the specific instances.) In this example, that includes two instances:

  • a2674d3b-a9c4-46bc-a835-b6ce21d522c2
  • a4ae094d-da07-4911-94f9-c982dc88f3cc

Binder also fetches the contents of the child nodes. These records look like this:

{
  "type": "redis_host",
  "address": "172.27.10.62",
  "ttl": 30,
  "redis_host": {
    "address": "172.27.10.62",
    "ports": [ 6379 ]
  }
}

{
  "type": "redis_host",
  "address": "172.27.10.67",
  "ttl": 30,
  "redis_host": {
    "address": "172.27.10.67",
    "ports": [ 6379 ]
  }
}

The record includes the IP address and TTLs that will be included in DNS answers. In this case, there are two addresses for "authcache.emy-10.joyent.us":

$ dig +nocmd +nocomments +noquestion +nostats authcache.emy-10.joyent.us
authcache.emy-10.joyent.us. 30  IN      A       172.27.10.67
authcache.emy-10.joyent.us. 30  IN      A       172.27.10.62

In order to use these, clients need to know the TCP port that the authcache service uses.

Note that clients can also query for the host records directly:

$ dig +nocmd +nocomments +noquestion +nostats a2674d3b-a9c4-46bc-a835-b6ce21d522c2.authcache.emy-10.joyent.us
a2674d3b-a9c4-46bc-a835-b6ce21d522c2.authcache.emy-10.joyent.us. 30 IN A 172.27.10.62

In this case, Binder answers the query by fetching the ZooKeeper node "/us/joyent/emy-10/authcache/a2674d3b-a9c4-46bc-a835-b6ce21d522c2", finding the host record there, and producing an "A" record. The service record is not involved here.

Service records also include the information required for Binder to answer DNS SRV queries. These allow clients to find all available servers running inside a container, which allows for more effective load balancing and resiliency than using a separate load balancer. For a worked example, see the "Configuration reference".

Host record reference

Host records are usually ephemeral nodes in ZooKeeper, which means they are removed when the corresponding Registrar becomes disconnected from ZooKeeper.

Host records have the following top-level properties:

Property Type Meaning
"address" string Apparently unused. Possibly historical.
"type" string Subtype of host record. See below.
"ttl" optional integer See "About TTLs" below.
type object The property name always matches the value of "type". See below for details.

Host records are distinguished from service records by having any "type" other than "service". Supported values of "type" (and the semantics of each type) are the same as those supported in the Registrar configuration reference. The various types of host records largely function the same way: each of these records causes Binder to produce either one "A" record with the IP address of that instance or multiple "SRV" records with the IP address and ports of the various instances contained inside the zone. (There's also a vestigial type called "database" which was historically produced by Manatee, but this type of record is neither produced nor consumed any more.)

The inner object (that has the same name as the value of "type") has properties:

Property Type Meaning
"address" string IPv4 address to use for A and SRV query responses.
"ports" optional array of integers TCP ports to use for SRV records. Binder generates one SRV answer for each element of this array.
"ttl" optional integer See "About TTLs" below.

Here's a host record created from our example above:

{
  "type": "load_balancer",
  "address": "172.27.10.72",
  "load_balancer": {
    "address": "172.27.10.72",
    "ports": [ 80 ]
  }
}

When queried for "A" records, Binder reports one for 172.27.10.72. When queried for SRV records, assuming the parent node (in ZooKeeper) for this host record is a service record, then Binder uses the protocol and service mentioned in the "service" record to generate an SRV answer for each port contained in this record.

Service record reference

Service records are those with "type" equal to "service". These are persistent nodes in ZooKeeper.

Service records have the following top-level properties:

Property Type Meaning
"type" string Always has value "service". (Otherwise, this is a host record.)
"service" object Describes the service name, protocol name, and TTL used for answering SRV queries about this service. The format of this object exactly matches the registration.service object in the Registrar configuration.

Here's a service record created from our example above:

{
  "type": "service",
  "service": {
    "type": "service",
    "service": {
      "srvce": "_http",
      "proto": "_tcp",
      "port": 80,
      "ttl": 60
    }
  }
}

DNS SRV records also support weights, but these are not supported by Registrar or Binder.

About TTLs

The semantics around which TTLs are used in DNS responses are surprisingly complex. In all cases, Binder provides default TTL values when no value is specified by any of the mechanisms below.

When looking up the IP address of a specific instance (not a service), you generally query Binder for "A" records for a DNS name that corresponds to a host record. In this case, the TTL is selected from whichever of the following is specified, in this order:

  • a TTL on the inner object within the host record (e.g., hostRecord[hostRecord.type].ttl).
  • a TTL on the host record itself (e.g., hostRecord.ttl)

When looking up a service's DNS name, you query Binder for "A" or "SRV" records for a DNS name that corresponds to a service record. In this case, Binder produces both "SRV" answers (that describe the instances available, using the DNS name for each host and a port number) and "A" answers as additionals (containing the resolutions for the hostnames provided in the "SRV" answers). For example, when you ask for the SRV records for "_http._tcp.example.joyent.us", you get:

$ dig +nocmd +nostats -t SRV _http._tcp.example.joyent.us
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 6415
;; flags: qr rd ad; QUERY: 1, ANSWER: 1, AUTHORITY: 0, ADDITIONAL: 2
;; WARNING: recursion requested but not available

;; OPT PSEUDOSECTION:
; EDNS: version: 0, flags:; udp: 1470
;; QUESTION SECTION:
;_http._tcp.example.joyent.us.  IN      SRV

;; ANSWER SECTION:
_http._tcp.example.joyent.us. 60 IN     SRV     0 10 80 b44c74d6.example.joyent.us.

;; ADDITIONAL SECTION:
b44c74d6.example.joyent.us. 30  IN      A       172.27.10.72

Binder is telling us that there's an instance at "b44c74d6.example.joyent.us" port 80, and it's separately telling us that the address for "b44c74d6.example.joyent.us" is 172.27.10.72. These two results can have different TTLs. The TTL on the SRV records indicates how long the client should cache the list of instances. Changes to this list should be propagated quickly, so we typically use TTLs on the order of 30 to 60 seconds. However, the TTLs on the "A" resolutions can be much longer, because it's almost unheard of for the IP address to change for a specific Triton or Manta zone.

There are basically two possibilities here:

  • If the TTL on the SRV records is shorter than or equal to that of the additional A records, then the client will generally re-resolve the SRV name as frequently as its TTL indicates. Since Binder always provides additionals for the A records, the client never needs to re-resolve the A records. The client ends up re-resolving all records on an interval specified by the SRV TTL.
  • If the TTL on the SRV records is longer than that of the A records, then a client would have to re-resolve the A records more frequently than the SRV records. That would likely result in significantly more load on Binder than is desirable. There's also not much reason to do this, since the addresses for individual zones don't generally change.

As a result, there's not generally much reason to have the TTLs differ between SRV records and A records, at least without significant changes to the way Binder and clients usually work.

Given all that, when you make an "SRV" query for a DNS name corresponding to a service (not a host), the TTL for the SRV records is selected from whichever of the following is specified, in this order:

  • a TTL on the service record's service details (serviceRecord.service.service.ttl)
  • a TTL on the inner record on the service record (serviceRecord.service.ttl)
  • a TTL on the service record itself (serviceRecord.ttl)

For the same query, the TTL for the "A" records is selected from:

  • a TTL on the child host record's inner record (hostRecord[hostRecord.type].ttl)
  • a TTL on the child host record itself (hostRecord.ttl)

When you make an "A" query for the same DNS name, the TTL used is the minimum of the above two TTLs (since the response represents both the list of instances and the addresses of each instance).

Debugging Notes

When Binder isn't reporting the results that you expect, ask the following questions.

Do you expect that the DNS results recently changed?

DNS results may change any time a Registrar instance establishes a new ZooKeeper session to ZooKeeper or loses its ZooKeeper session. There are a few reasons why this can take some time to be reflected in Binder:

  • Binder caches all ZooKeeper queries for up to 60 seconds.
  • ZooKeeper sessions can take up to 60 seconds to expire (depending on the Registrar configuration).
  • DNS clients cache the answers to queries for up to the TTL on each answer. This is typically 30-60 seconds, but can be several minutes, depending on the component.

As a result, when Registrar first starts up, it can take up to a minute for the new registration to be reflected in any particular Binder instance. When Registrar shuts down (or the underlying server powers off, panics, or becomes partitioned), it can take at least two minutes for that to be reflected in Binder, and clients may not discover the change for an additional TTL seconds.

Since the Binder instances operate independently (and cache independently), you can get inconsistent results if you make the same query against different Binders before they've all learned about recent updates.

Does the structure of records in ZooKeeper reflect what you expect?

Use zkCli.sh (the ZooKeeper CLI) inside any "binder" or "nameservice" zone to answer this question. (Run it with no arguments and use the help command to get started.)

  • If you're querying a service (e.g., "example.joyent.us"), you should find a "service" record at the corresponding ZooKeeper node (e.g., get /us/joyent/example). You can ls /us/joyent/example that path to see the child nodes. Do you see the entries for hosts you expect to be there? You can "get" each of these. You should find host records for the addresses that appear in DNS.
  • If you're querying a host (e.g., "$hostname.example.joyent.us"), you should just find a host record at the corresponding ZooKeeper node.

Remember too that only certain types of host records can be used with service DNS names. If you're using a record of type "db_host" or "host", it won't show up when you query the service DNS name. See "Registration of host records" under the configuration reference above for details.

If you don't find the records you expect in ZooKeeper, your Registrar configuration may be incorrect, or Registrar may not be functioning. If you do, Binder may not be working correctly.

License

Registrar is licensed under the Mozilla Public License version 2.0.