A Performant API Gateway based on NGINX and Openresty
Clone or download
ddragosd added REMOTE_CONFIG_SYNC_CMD env var to customize the sync cmd (#69)
* added REMOTE_CONFIG_SYNC_CMD env var to customize the sync cmd
* updated README
Latest commit 835f20b Jul 27, 2018



A performant API Gateway based on Openresty and NGINX.

Table of Contents


The current project is considered production ready.

Quick start


$ docker run --name="apigateway" \
            -p 80:80 \
            -e "LOG_LEVEL=info" \

With Apache Mesos

$ docker run --name="apigateway" \
            -p 80:80 \
            -e "MARATHON_HOST=http://<marathon_host>:<port>" \
            -e "LOG_LEVEL=info" \

This command starts an API Gateway that automatically discovers the services running in Marathon. The discovered services are exposed on individual VHosts as you can see in the config file.

Accessing a Marathon app locally

For example, if you have an application named hello-world you can access it on its VHost in 2 ways:

  1. Edit /etc/hosts and add <docker_host_ip> hello-world.api.localhost then browse to http://hello-world.api.localhost
  2. Sending the Host header in a curl command: curl -H "Host:hello-world.api.localhost" http://<docker_host_ip>

The discovery script is provided as an example for a quick-start and it can be replaced with your favourite discovery mechanism. The script updates a configuration file containing all the NGINX upstreams that are used in the config file.

Accessing a Marathon app remotely

All you need to do is to create a DNS entry like *.api.example.com or *.gw.example.com and have it resolve to the nodes running the Gateway.

Assuming there is an application deployed in Marathon called hello-world it can be accessed at the URL: http://hello-world.api.example.com The Gateway is automatically proxying the request to the hello-world application in Marathon.

If you call http://my-custom-app.api.example.com the Gateway will proxy to a Marathon app named my-custom-app. If the application is not deployed in Marathon a 5xx error is returned back to the client. This behaviour is ofcourse configurable.

Running with a different configuration folder

One way to simplify the configuration of the Gateway is to copy the api-gateway-config folder in S3, having all nodes syncing from that location. For the moment only AWS S3 is supported but the plan is to add support for other clouds too.

In order to work with S3 the Gateway needs to know the bucket. The command is similar to the one above with an extra environment variable named REMOTE_CONFIG

docker run --name="apigateway" \
            -p 80:80 \
            -e "MARATHON_HOST=http://<marathon_host>:<port>" \
            -e "REMOTE_CONFIG=s3://api-gateway-config-bucket" \
            -e "LOG_LEVEL=info" \

s3://api-gateway-config-bucket should contain something similar to what's in the api-gateway-config folder. For any customization it would be good to start from there. There are only 2 files that won't be synchronised:

  • /etc/api-gateway/conf.d/includes/resolvers.conf . Read bellow for more info
  • https://github.com/adobe-apiplatform/apigateway/blob/master/api-gateway-config/environment.conf.d/api-gateway-upstreams.http.conf which is automatically generated by the Marathon Discovery script.

By default the access to the S3 bucket is done using IAM roles, but AWS Credentials can also be specified through environment variables:

docker run --name="apigateway" \
            -p 80:80 \
            -e "MARATHON_HOST=http://<marathon_host>:<port>" \
            -e "REMOTE_CONFIG=s3://api-gateway-config-bucket" \
            -e "AWS_ACCESS_KEY_ID=--change-me--" \
            -e "AWS_SECRET_ACCESS_KEY=--change-me--" \
            -e "LOG_LEVEL=info" \

By default the remote config is checked for changes every 10s. This interval is configurable through the REMOTE_CONFIG_SYNC_INTERVAL env variable. There are a few things to take into consideration when changing this value:

  • How often the configs really change
  • Cost. Some tools may make 1 API request per file to compare it. I.e. in S3 72 config files checked every 10s costs $7.46 but when checked every 30s it's only $2.4, times number of GW nodes.
  • Average time for an API Request. When reloading the GW the existing NGINX processes handling active connections are kept in the background until the request completes. So reloading the Gateway too fast may have the side effect of keeping too many processes running at the same time. This may, or may not be a problem but it's good to be aware of it.

Customizing the sync command

The sync command used for downloading the configuration files can be controlled via REMOTE_CONFIG_SYNC_CMD as well. This ENV VAR overrides the REMOTE_CONFIG one.


While starting up this container automatically creates the /etc/api-gateway/conf.d/includes/resolvers.conf config file using /etc/resolv.conf as the source. To learn more about the resolver directive in NGINX see the docs.

Running the API Gateway outside of Marathon and Mesos

Besides the discovery part which is dependent on Marathon at the moment, the API Gateway can run on its own as well. The Marathon service discovery is activated with the -e "MARATHON_HOST=http://<marathon_host>:<port>/".

What's inside

Module Version Details
Openresty Installed in /usr/local/sbin/api-gateway
Openresty compiled --with-debug Installed in /usr/local/sbin/api-gateway-debug which enables debugging log
Test Nginx 0.24 Useful for executing integration tests from the container.
It's installed in /usr/local/test-nginx-0.24/.
It's also used during Docker build to execute make test on lua modules.
PCRE 8.37 Enables PCRE JIT support
ZeroMQ 4.0.5 ZeroMQ
CZMQ 2.2.0 CZMQ - High-level C Binding for ZeroMQ
Modules for API Management and Logging
Module Version Description
api-gateway-config-supervisor 1.0.3 Syncs config files from Amazon S3 reloading the gateway with the updates
api-gateway-cachemanager 1.0.1 Lua library for managing multiple cache stores
api-gateway-hmac 1.0.0 HMAC support for Lua with multiple algorithms, via OpenSSL and FFI
api-gateway-aws 1.7.1 AWS SDK for Nginx in Lua
api-gateway-request-validation 1.2.4 API Request Validation framework
api-gateway-async-logger 1.0.1 Performant async logger
api-gateway-zmq-logger 1.0.0 Lua logger for ZMQ with FFI and CZMQ
api-gateway-request-tracking 1.0.1 Usage and Tracking Handler for the API Gateway
Other Lua Modules
Module Version Description
lua-resty-http v0.07 Lua HTTP client cosocket driver for OpenResty / ngx_lua
lua-resty-iputils v0.2.0 Utility functions for working with IP addresses in Openresty


The following performance tests results have been obtained on a virtual machine with 8 CPU cores and 4GB Memory.

The API Gateway container has been started with 4 CPU cores and net=host :

docker run --cpuset-cpus=0-3 --net=host --name="apigateway" -e "LOG_LEVEL=notice" adobeapiplatform/apigateway:latest

WRK test has been started with 4 CPU cores and net=host:

docker run --cpuset-cpus=4-7 --net=host williamyeh/wrk:4.0.1 -t4 -c1000 -d30s http://<docker_host_ip>/health-check
Running 30s test @
  4 threads and 1000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    30.38ms   73.80ms   1.16s    90.90%
    Req/Sec    35.26k    11.98k   83.70k    68.72%
  4214013 requests in 30.06s, 1.28GB read

Requests/sec: 140165.09

Transfer/sec:     43.57MB

Developer guide

To build the docker image locally use:

 make docker

To SSH into the newly built image use ( note that this is not the running image):

 make docker-ssh

Running and Stopping the Docker image

 make docker-run

The main API Gateway process is exposed to port 80. To test that the Gateway works see its health-check:

 $ curl http://<docker_host_ip>/health-check
   API-Platform is running!

If you're up for a quick performance test, you can play with Apache Benchmark via Docker:

 docker run jordi/ab ab -k -n 200000 -c 500 http://<docker_host_ip>/health-check

To run docker mounting the local api-gateway-config directory into /etc/api-gateway/ issue:

$ make docker-debug

In debug mode the docker container starts a special api-gateway compiled --with-debug providing very detailed debugging information. When started with -e "LOG_LEVEL=info" the output is quite verbose. To learn more about this option visit NGINX docs.

When done stop the image:

make docker-stop

Enabling API KEY Management through Redis

This command starts two docker containers: redis and gateway

make docker-compose

SSH into the running image

make docker-attach

Running the container in Mesos with Marathon

Make an HTTP POST on http://<marathon-host>/v2/apps with the following payload. For optimal performance leave the network on HOST mode. To learn more about the network modes visit the Docker documentation.

  "id": "api-gateway",
  "container": {
    "type": "DOCKER",
    "docker": {
      "image": "adobeapiplatform/apigateway:latest",
      "forcePullImage": true,
      "network": "HOST"
  "cpus": 4,
  "mem": 4096.0,
  "env": {
    "MARATHON_HOST": "http://<marathon_host>:<marathon_port>"
  "constraints": [  [ "hostname","UNIQUE" ]  ],
  "ports": [ 80 ],
  "healthChecks": [
      "protocol": "HTTP",
      "portIndex": 0,
      "path": "/health-check",
      "gracePeriodSeconds": 3,
      "intervalSeconds": 10,
      "timeoutSeconds": 10
  "instances": 1

To run the Gateway only on specific nodes marked with slave_public you can add the property bellow to the main JSON object:

"acceptedResourceRoles": [ "slave_public" ]
Auto-discover and register Marathon tasks in the Gateway

To enable auto-discovery in a Mesos with Marathon framework define the following Environment Variables:


So the Docker command is now :

docker run --name="apigateway" \
            -p 8080:80 \
            -e "MARATHON_HOST=http://<marathon_host>:<port>/" \
            -e "LOG_LEVEL=info" \