apisix/docs/en/latest/plugin-develop.md

title
Plugin Develop

table of contents

• where to put your plugins
• check dependencies
• name and config
• schema and check
• choose phase to run
• implement the logic
• write test case
• register public API
• register control API

where to put your plugins

There are two ways to add new features based on APISIX.

1. modify the source of APISIX and redistribute it (not so recommended)
2. setup the extra_lua_path and extra_lua_cpath in conf/config.yaml to load your own code. Your own code will be loaded instead of the builtin one with the same name, so you can use this way to override the builtin behavior if needed.

For example, you can create a directory structure like this:

├── example
│   └── apisix
│       ├── plugins
│       │   └── 3rd-party.lua
│       └── stream
│           └── plugins
│               └── 3rd-party.lua

Then add this configuration into your conf/config.yaml:

apisix:
    ...
    extra_lua_path: "/path/to/example/?.lua"

Now using require "apisix.plugins.3rd-party" will load your plugin, just like require "apisix.plugins.jwt-auth" will load the jwt-auth plugin.

check dependencies

if you have dependencies on external libraries, check the dependent items. if your plugin needs to use shared memory, it needs to declare in apisix/cli/ngx_tpl.lua, for example :

    lua_shared_dict plugin-limit-req     10m;
    lua_shared_dict plugin-limit-count   10m;
    lua_shared_dict prometheus-metrics   10m;
    lua_shared_dict plugin-limit-conn    10m;
    lua_shared_dict upstream-healthcheck 10m;
    lua_shared_dict worker-events        10m;

    # for openid-connect plugin
    lua_shared_dict discovery             1m; # cache for discovery metadata documents
    lua_shared_dict jwks                  1m; # cache for JWKs
    lua_shared_dict introspection        10m; # cache for JWT verification results

The plugin itself provides the init method. It is convenient for plugins to perform some initialization after the plugin is loaded.

Note : if the dependency of some plugin needs to be initialized when Nginx start, you may need to add logic to the initialization method "http_init" in the file apisix/init.lua, and you may need to add some processing on generated part of Nginx configuration file in apisix/cli/ngx_tpl.lua file. But it is easy to have an impact on the overall situation according to the existing plugin mechanism, we do not recommend this unless you have a complete grasp of the code.

name and config

Determine the name and priority of the plugin, and add to conf/config-default.yaml. For example, for the example-plugin plugin, you need to specify the plugin name in the code (the name is the unique identifier of the plugin and cannot be duplicate), you can see the code in file "apisix/plugins/example-plugin.lua" :

local plugin_name = "example-plugin"

local _M = {
    version = 0.1,
    priority = 0,
    name = plugin_name,
    schema = schema,
    metadata_schema = metadata_schema,
}

Note : The priority of the new plugin cannot be same to any existing ones, you can use the /v1/schema method of control API to view the priority of all plugins. In addition, plugins with higher priority value will be executed first in a given phase (see the definition of phase in choose-phase-to-run). For example, the priority of example-plugin is 0 and the priority of ip-restriction is 3000. Therefore, the ip-restriction plugin will be executed first, then the example-plugin plugin.

in the "conf/config-default.yaml" configuration file, the enabled plugins (all specified by plugin name) are listed.

plugins:                          # plugin list
  - limit-req
  - limit-count
  - limit-conn
  - key-auth
  - prometheus
  - node-status
  - jwt-auth
  - zipkin
  - ip-restriction
  - grpc-transcode
  - serverless-pre-function
  - serverless-post-function
  - openid-connect
  - proxy-rewrite
  - redirect
  ...

Note : the order of the plugins is not related to the order of execution.

To enable your plugin, copy this plugin list into conf/config.yaml, and add your plugin name. For instance:

apisix:
  admin_key:
    - name: "admin"
      # yamllint disable rule:comments-indentation
      key: edd1c9f034335f136f87ad84b625c8f1 # using fixed API token has security risk, please update it when you deploy to production environment
      # yamllint enable rule:comments-indentation
      role: admin

plugins: # copied from config-default.yaml
  ...
  - your-plugin

If your plugin has a new code directory of its own, and you need to redistribute it with the APISIX source code, you will need to modify the Makefile to create directory, such as:

$(INSTALL) -d $(INST_LUADIR)/apisix/plugins/skywalking
$(INSTALL) apisix/plugins/skywalking/*.lua $(INST_LUADIR)/apisix/plugins/skywalking/

schema and check

Write Json Schema descriptions and check functions. Similarly, take the example-plugin plugin as an example to see its configuration data :

{
    "example-plugin" : {
        "i": 1,
        "s": "s",
        "t": [1]
    }
}

Let's look at its schema description :

local schema = {
    type = "object",
    properties = {
        i = {type = "number", minimum = 0},
        s = {type = "string"},
        t = {type = "array", minItems = 1},
        ip = {type = "string"},
        port = {type = "integer"},
    },
    required = {"i"},
}

The schema defines a non-negative number i, a string s, a non-empty array of t, and ip / port. Only i is required.

At the same time, we need to implement the check_schema(conf) method to complete the specification verification.

function _M.check_schema(conf, schema_type)
    return core.schema.check(schema, conf)
end

Note: the project has provided the public method "core.schema.check", which can be used directly to complete JSON verification.

In addition, if the plugin needs to use some metadata, we can define the plugin metadata_schema, and then we can dynamically manage these metadata through the admin api. Example:

local metadata_schema = {
    type = "object",
    properties = {
        ikey = {type = "number", minimum = 0},
        skey = {type = "string"},
    },
    required = {"ikey", "skey"},
    additionalProperties = false,
}

local plugin_name = "example-plugin"

local _M = {
    version = 0.1,
    priority = 0,        -- TODO: add a type field, may be a good idea
    name = plugin_name,
    schema = schema,
    metadata_schema = metadata_schema,
}

You might have noticed the key-auth plugin has type = 'auth' in its definition. When we set the type of plugin to auth, it means that this plugin is an authentication plugin.

An authentication plugin needs to choose a consumer after execution. For example, in key-auth plugin, it calls the consumer.attach_consumer to attach a consumer, which is chosen via the apikey header.

To interact with the consumer resource, this type of plugin needs to provide a consumer_schema to check the plugins configuration in the consumer.

Here is the consumer configuration for key-auth plugin:

{
    "username": "Joe",
    "plugins": {
        "key-auth": {
            "key": "Joe's key"
        }
    }
}

It will be used when you try to create a Consumer

To validate the configuration, the plugin uses a schema like this:

local consumer_schema = {
    type = "object",
    additionalProperties = false,
    properties = {
        key = {type = "string"},
    },
    required = {"key"},
}

Note the difference between key-auth's check_schema(conf) method to example-plugin's:

-- key-auth
function _M.check_schema(conf, schema_type)
    if schema_type == core.schema.TYPE_CONSUMER then
        return core.schema.check(consumer_schema, conf)
    else
        return core.schema.check(schema, conf)
    end
end

-- example-plugin
function _M.check_schema(conf, schema_type)
    return core.schema.check(schema, conf)
end

choose phase to run

Determine which phase to run, generally access or rewrite. If you don't know the Openresty life cycle, it's recommended to know it in advance. For example key-auth is an authentication plugin, thus the authentication should be completed before forwarding the request to any upstream service. Therefore, the plugin must be executed in the rewrite phases. In APISIX, only the authentication logic can be run in the rewrite phase. Other logic needs to run before proxy should be in access phase.

The following code snippet shows how to implement any logic relevant to the plugin in the OpenResty log phase.

function _M.log(conf)
-- Implement logic here
end

Note : we can't invoke ngx.exit or core.respond.exit in rewrite phase and access phase. if need to exit, just return the status and body, the plugin engine will make the exit happen with the returned status and body. example

implement the logic

Write the logic of the plugin in the corresponding phase. There are two parameters conf and ctx in the phase method, take the limit-conn plugin configuration as an example.

curl http://127.0.0.1:9080/apisix/admin/routes/1 -H 'X-API-KEY: edd1c9f034335f136f87ad84b625c8f1' -X PUT -d '
{
    "methods": ["GET"],
    "uri": "/index.html",
    "id": 1,
    "plugins": {
        "limit-conn": {
            "conn": 1,
            "burst": 0,
            "default_conn_delay": 0.1,
            "rejected_code": 503,
            "key": "remote_addr"
        }
    },
    "upstream": {
        "type": "roundrobin",
        "nodes": {
            "39.97.63.215:80": 1
        }
    }
}'

conf parameter

The conf parameter is the relevant configuration information of the plugin, you can use core.log.warn(core.json.encode(conf)) to output it to error.log for viewing, as shown below:

function _M.access(conf, ctx)
    core.log.warn(core.json.encode(conf))
    ......
end

conf:

{
    "rejected_code":503,
    "burst":0,
    "default_conn_delay":0.1,
    "conn":1,
    "key":"remote_addr"
}

ctx parameter

The ctx parameter caches data information related to the request. You can use core.log.warn(core.json.encode(ctx, true)) to output it to error.log for viewing, as shown below :

function _M.access(conf, ctx)
    core.log.warn(core.json.encode(ctx, true))
    ......
end

write test case

For functions, write and improve the test cases of various dimensions, do a comprehensive test for your plugin! The test cases of plugins are all in the "t/plugin" directory. You can go ahead to find out. APISIX uses test-nginx as the test framework. A test case (.t file) is usually divided into prologue and data parts by _data_. Here we will briefly introduce the data part, that is, the part of the real test case. For example, the key-auth plugin:

=== TEST 1: sanity
--- config
    location /t {
        content_by_lua_block {
            local plugin = require("apisix.plugins.key-auth")
            local ok, err = plugin.check_schema({key = 'test-key'}, core.schema.TYPE_CONSUMER)
            if not ok then
                ngx.say(err)
            end

            ngx.say("done")
        }
    }
--- request
GET /t
--- response_body
done
--- no_error_log
[error]

A test case consists of three parts :

• Program code : configuration content of Nginx location
• Input : http request information
• Output check : status, header, body, error log check

When we request /t, which config in the configuration file, the Nginx will call "content_by_lua_block" instruction to complete the Lua script, and finally return. The assertion of the use case is response_body return "done", "no_error_log" means to check the "error.log" of Nginx. There must be no ERROR level record. The log files for the unit test are located in the following folder: 't/servroot/logs'.

The above test case represents a simple scenario. Most scenarios will require multiple steps to validate. To do this, create multiple tests === TEST 1, === TEST 2, and so on. These tests will be executed sequentially, allowing you to break down scenarios into a sequence of atomic steps.

Additionally, there are some convenience testing endpoints which can be found here. For example, see proxy-rewrite. In test 42, the upstream uri is made to redirect /test?new_uri=hello to /hello (which always returns hello world). In test 43, the response body is confirmed to equal hello world, meaning the proxy-rewrite configuration added with test 42 worked correctly.

Refer the following document to setup the testing framework.

Attach the test-nginx execution process:

According to the path we configured in the makefile and some configuration items at the front of each .t file, the framework will assemble into a complete nginx.conf file. "t/servroot" is the working directory of Nginx and start the Nginx instance. according to the information provided by the test case, initiate the http request and check that the return items of HTTP include HTTP status, HTTP response header, HTTP response body and so on.

Register public API

A plugin can register API which exposes to the public. Take jwt-auth plugin as an example, this plugin registers GET /apisix/plugin/jwt/sign to allow client to sign its key:

local function gen_token()
    --...
end

function _M.api()
    return {
        {
            methods = {"GET"},
            uri = "/apisix/plugin/jwt/sign",
            handler = gen_token,
        }
    }
end

Note that the public API is exposed to the public. You may need to use interceptors to protect it.

Register control API

If you only want to expose the API to the localhost or intranet, you can expose it via Control API.

Take a look at example-plugin plugin:

local function hello()
    local args = ngx.req.get_uri_args()
    if args["json"] then
        return 200, {msg = "world"}
    else
        return 200, "world\n"
    end
end


function _M.control_api()
    return {
        {
            methods = {"GET"},
            uris = {"/v1/plugin/example-plugin/hello"},
            handler = hello,
        }
    }
end

If you don't change the default control API configuration, the plugin will be expose GET /v1/plugin/example-plugin/hello which can only be accessed via 127.0.0.1.