Jul 17, 2022 / SERVICE DISCOVERY, DEVOPS, APACHE APISIX

Writing your own service discovery client for Apache APISIX

API Gateways in general, and Apache APISIX in particular, provide a single entry point into one’s information system. This architecture allows for managing load balancing and failover over similar nodes. For example, here’s how you can create a route balanced over two nodes in Apache APISIX:

curl http://localhost:9080/apisix/admin/routes/1 -H 'X-API-KEY: edd1c9f034335f136f87ad84b625c8f1' -X PUT -i -d '{
  "uri": "/*",
  "upstream": {
    "type": "roundrobin",
    "nodes": {
      "192.168.0.1:80": 1,           (1)
      "192.168.0.2:80": 1            (1)
    }
  }
}'

1	Every request has a 50/50 chance of being sent to either node

It worked for a long time, but in this day and age, nodes are probably not pets but cattle: they come, and they go. Hence, it’s essential to dynamically update the nodes' list when it happens.

In this post, I’d like to explain how to do it.

Existing service discovery registries

Please, don’t reinvent the wheel! Apache APISIX comes with a bunch of existing service discovery registries out-of-the-box.

Registry	Provider	Description	Integration
DNS			Link
Consul	HashiCorp	Consul uses service identities and traditional networking practices to help organizations securely connect applications running in any environment.	Link
nacos	Alibaba	An easy-to-use dynamic service discovery, configuration and service management platform for building cloud native applications.	Link
Eureka	Netflix	Eureka is a RESTful (Representational State Transfer) service that is primarily used in the AWS cloud for the purpose of discovery, load balancing and failover of middle-tier servers. It plays a critical role in Netflix mid-tier infra.	Link
Zookeeper	Apache	ZooKeeper is a centralized service for maintaining configuration information, naming, providing distributed synchronization, and providing group services. All of these kinds of services are used in some form or another by distributed applications. Each time they are implemented there is a lot of work that goes into fixing the bugs and race conditions that are inevitable. Because of the difficulty of implementing these kinds of services, applications initially usually skimp on them, which make them brittle in the presence of change and difficult to manage. Even when done correctly, different implementations of these services lead to management complexity when the applications are deployed.	Link
Kubernetes	CNCF	Kubernetes, also known as K8s, is an open-source system for automating deployment, scaling, and management of containerized applications.	Link

Registry

Provider

Description

Integration

DNS

Link

Consul

HashiCorp

Consul uses service identities and traditional networking practices to help organizations securely connect applications running in any environment.

Link

nacos

Alibaba

An easy-to-use dynamic service discovery, configuration and service management platform for building cloud native applications.

Link

Eureka

Netflix

Eureka is a RESTful (Representational State Transfer) service that is primarily used in the AWS cloud for the purpose of discovery, load balancing and failover of middle-tier servers. It plays a critical role in Netflix mid-tier infra.

Link

Zookeeper

Apache

ZooKeeper is a centralized service for maintaining configuration information, naming, providing distributed synchronization, and providing group services. All of these kinds of services are used in some form or another by distributed applications. Each time they are implemented there is a lot of work that goes into fixing the bugs and race conditions that are inevitable. Because of the difficulty of implementing these kinds of services, applications initially usually skimp on them, which make them brittle in the presence of change and difficult to manage. Even when done correctly, different implementations of these services lead to management complexity when the applications are deployed.

Link

Kubernetes

CNCF

Kubernetes, also known as K8s, is an open-source system for automating deployment, scaling, and management of containerized applications.

Link

Before you write your own, make sure your platform is not listed above.

Setting up the environment

To ease my life and make it easier to reproduce the steps, I chose to use Docker and Docker Compose. Here’s the sample that you’re welcome to reuse for development purposes:

version: "3"

services:
  apisix:
    image: apache/apisix:2.14.1-alpine                                  (1)
    volumes:
      - ./config/config.yaml:/usr/local/apisix/conf/config.yaml:ro         (2)
      - ./yaml:/usr/local/apisix/apisix/discovery/yaml:ro               (3)
      - ./sample:/var/apisix:ro                                         (3)
    ports:
      - "9080:9080"
      - "9090:9090"
    restart: always                                                     (4)
    depends_on:
      - etcd
  etcd:
    image: bitnami/etcd:3.5.2
    environment:
      ETCD_ENABLE_V2: "true"
      ALLOW_NONE_AUTHENTICATION: "yes"                                  (5)
      ETCD_ADVERTISE_CLIENT_URLS: "http://0.0.0.0:2397"
      ETCD_LISTEN_CLIENT_URLS: "http://0.0.0.0:2397"
    ports:
      - "2397:2397"                                                     (6)

1	Use the latest image at the time of this writing
2	Minimal configuration, see the complete file for more details
3	Bear with me; I’ll explain later
4	Apache APISIX starts faster than etcd. It will look for etcd, concludes it’s not available, and stop. We want to start it again afterward.
5	Don’t do this in production!
6	If running on Docker Desktop with Kubernetes enabled, a port conflict occurs with the default port. We need to change it.

The use-case

Let’s imagine a YAML file that references the available nodes. An ad hoc process listens to changes to the topology: it re-generates the file with the new nodes. Our client reads the file regularly and updates its internal nodes list.

Here’s the proposed structure:

nodes.yaml

nodes:
  "192.168.1.62:81": 1
#END

Developing the discovery service client

To create a discovery client, the following structure is required:

yaml                             (1)
  |_ schema.lua                  (2)
  |_ init.lua                    (3)

1	Give it a name; `yaml` is as good as any other
2	List configuration parameters - name, type, whether it’s required, etc.
3	For the code itself

For Apache APISIX to use the client, you need to set the yaml folder as a child folder of /usr/local/apisix/apisix/discovery; hence the mount in the Docker Compose file above.

The client needs to follow a specific structure.

init.lua

local _M = {}

-- Initialize the client
function _M.init_worker()
end

-- Get available nodes.
--
-- @param service_name Not used
-- @treturn table
-- @return Available nodes, e.g., { [1] = { ["port"] = 81, ["host"] = 127.0.0.1, ["weight"] = 1 }}
function _M.nodes(service_name)
end

-- Dump existing nodes.
--
-- @return Debugging information
function dump_data()
end

Let’s start with the easy part:

init.lua

local nodes

function _M.nodes(service_name)
  return nodes                         (1)
end

1	Return the `nodes` table. We fill the nodes in the `_M.init()` function

We want the client to read the YAML file regularly. For this, we can leverage the power of the Lua Nginx module. It’s part of OpenResty, which Apache APISIX is built upon. The module offers additional APIs, and two of them are particularly useful:

ngx.timer.at to call a function after an initial delay
ngx.timer.every to call a function repeatedly at regular intervals

init.lua

local ngx_timer_at    = ngx.timer.at
local ngx_timer_every = ngx.timer.every

function _M.init_worker()
    ngx_timer_at(0, read_file)                 (1)
    ngx_timer_every(20, read_file)             (2)
end

1	Call the `read_file` function immediately
2	Call the `read_file` function every 20 seconds

Now is time to write the read_file function.

local util = require("apisix.cli.util")                                   (1)
local yaml = require("tinyyaml")                                          (2)

local function read_file()
    local content, err = util.read_file("/var/apisix/nodes.yaml")         (3)
    if not content then
        return
    end
    local nodes_conf, err = yaml.parse(content)                           (4)
    if not nodes_conf then
        return
    end
    if not nodes then
        nodes = {}
    end
    for uri, weight in pairs(nodes_conf.nodes) do                         (5)
        local host_port = {}
        for str in string.gmatch(uri, "[^:]+") do
            table.insert(host_port, str)                                  (6)
        end
        local node = {
          host = host_port[1],
          port = tonumber(host_port[2]),
          weight = weight,
        }                                                                 (7)
        table.insert(nodes, node)                                         (8)
    end
end

1	Import the library to read file
2	Import the library to convert YAML content to Lua tables
3	Read the file
4	Parse its content
5	Iterate over the lines, which should be formatted as `"<ip>:<port>":<weight>`. I’m too lazy to handle all corner cases, be my guest.
6	Parse each key - the `"<ip>:<port>"` string
7	Create a Lua table for each node
8	Insert it into the `nodes` file local variable

Putting the code to the test

I used the default Apache web server available on my Mac to test the code.

I changed the port from 80 to 81 to avoid conflicts
I started it with sudo apachectl start
I noted the IP of my machine, which is available from Docker containers
I updated the configuration file:
```
nodes:
  "192.168.1.62:81": 1
#END
```
I started the Docker Compose containers - docker compose up

At this point, I used the admin API to create a route with the new YAML service discovery client:

curl http://127.0.0.1:9080/apisix/admin/routes/1 -H 'X-API-KEY: edd1c9f034335f136f87ad84b625c8f1' -X PUT -i -d '{
  "uri": "/",
  "upstream": {
    "service_name": "MY-YAML",              (1)
    "type": "roundrobin",
    "discovery_type": "yaml"                (2) (3)
  }
}'

1	Matches the `service_name` parameter in the `_M.nodes(service_name)` function. It potentially allows returning different nodes based on it. We didn’t use it here, so anything works.
2	The magic happens here. The label must match the name of the discovery folder in `/usr/local/apisix/apisix/discovery/`.
3	No nodes are set; the service discovery client will dynamically return them

Let’s test it:

curl localhost:9080

It returns the root page served by the Apache Server as expected:

<html><body><h1>It works!</h1></body></html>

Nitpicking

While the above code works as expected, we can improve it.

Logging

Relevant logging can help your future self solve nasty bugs in production.

local core = require("apisix.core")

local function read_file(premature)
    local content, err = util.read_file("/var/apisix/nodes.yaml")
    if not content then
        log.error("Unable to open YAML discovery configuration file: ", err)    (1)
        return
    end

1	Trace the error

Parameterization

So far, we didn’t use any parameters. The configuration file path and the fetch interval are hard-coded. We can do better by making them configurable.

return {
    type = "object",
    properties = {
        path = { type = "string", default = "/var/apisix/nodes.yaml" },    (1)
        fetch_interval = { type = "integer", minimum = 1, default = 30 },  (1)
    },
}

1	Parameters with their type and default value. None of them are mandatory.

On the code side, we can use them accordingly.

local core       = require("apisix.core")
local local_conf = require("apisix.core.config_local").local_conf()

function _M.init_worker()
    local fetch_interval = local_conf.discovery and
                           local_conf.discovery.yaml and
                           local_conf.discovery.yaml.fetch_interval
    ngx_timer_every(fetch_interval, read_file)

Premature

Finally, the ngx.timer.every API calls our function with a dedicated premature parameter:

Premature timer expiration happens when the Nginx worker process is trying to shut down, as in an Nginx configuration reload triggered by the HUP signal or in an Nginx server shutdown. When the Nginx worker is trying to shut down, one can no longer call ngx.timer.at to create new timers with nonzero delays and in that case ngx.timer.at will return a "conditional false" value and a string describing the error, that is, "process exiting".

— ngx.timer.at

Let’s be a good citizen-developer and handle the parameter accordingly:

local function read_file(premature)
    if premature then
        return
    end
end

Conclusion

Most modern infrastructures are dynamic - servers are cattle, not pets. In this case, it doesn’t make much sense to configure the nodes of an upstream statically.

For this reason, Apache APISIX provides service discovery clients. While it comes with a bundle out-of-the-box, it’s possible to write your own by following a couple of steps. In this post, I described these steps to implement a node registry based on a YAML file.

The complete source code for this post can be found on Github.

To go further:

Follow @nicolas_frankel

Writing your own service discovery client for Apache APISIX

Existing service discovery registries

Setting up the environment

The use-case

Developing the discovery service client

Putting the code to the test

Nitpicking

Logging

Parameterization

Premature

Conclusion

To go further:

Discussing Backend For Front-end

Secure your web apps with an API Gateway