mTLS Everywhere: How to Configure TLS for APISIX

TLS in a Few Words

TLS offers several capabilities:

Server authentication: the client is confident that the server it exchanges data with is the right one. It avoids sending data, which might be confidential, to the wrong actor
Optional client authentication: the other way around, the server only allows clients whose identity can be verified
Confidentiality: no third party can read the data exchanged between the client and the server
Integrity: no third party can tamper with the data

TLS works through certificates. A certificate is similar to an ID, proving the certificate's holder identity. Just like an ID, you need to trust who delivered it. Trust is established through a chain: if I trust Alice, who trusts Bob, who in turn trusts Charlie, who delivered the certificate, then I trust the latter. In this scenario, Alice is known as the root certificate authority.

TLS authentication is based on public key cryptography. Alice generates a public key/private key pair and publishes the public key. If one encrypts data with the public key, only the private key that generated the public key can decrypt it. The other usage is for one to encrypt data with the private key and everybody with the public key to decrypt it, thus proving their identity.

Finally, mutual TLS, aka mTLS, is the configuration of two-way TLS: server authentication to the client, as usual, but also the other way around, client authentication to the server.

We now have enough understanding of the concepts to get our hands dirty.

Generating Certificates with cert-manager

A couple of root Certificate Authorities are installed in browsers by default. That's how we can browse HTTPS websites safely, trusting that https://apache.org is the site they pretend to be. The infrastructure has no pre-installed certificates, so we must start from scratch.

We need at least one root certificate. In turn, it will generate all other certificates. While it's possible to do every manually, I'll rely on cert-manager in Kubernetes. As its name implies, cert-manager is a solution to manage certificates.

Installing it with Helm is straightforward:

helm repo add jetstack https://charts.jetstack.io  #1

helm install \
  cert-manager jetstack/cert-manager \
  --namespace cert-manager \                       #2
  --create-namespace \                             #2
  --version v1.11.0 \
  --set installCRDs=true \
  --set prometheus.enabled=false                   #3

Add the charts' repository
Install the objects in a dedicated namespace
Don't monitor, in the scope of this post

We can make sure that everything works as expected by looking at the pods:

kubectl get pods -n cert-manager

cert-manager-cainjector-7f694c4c58-fc9bk  1/1  Running  2  (2d1h ago)  7d
cert-manager-cc4b776cf-8p2t8              1/1  Running  1  (2d1h ago)  7d
cert-manager-webhook-7cd8c769bb-494tl     1/1  Running  1  (2d1h ago)  7d

cert-manager can sign certificates from multiple sources: HashiCorp Vault, Let's Encrypt, etc. To keep things simple:

We will generate our dedicated root certificate, i.e., Self-Signed
We won't handle certificates rotation

Let's start with the following:

apiVersion: cert-manager.io/v1
kind: ClusterIssuer                           #1
metadata:
  name: selfsigned-issuer
spec:
  selfSigned: {}
---
apiVersion: v1
kind: Namespace
metadata:
  name: tls                                   #2
---
apiVersion: cert-manager.io/v1
kind: Certificate                             #3
metadata:
  name: selfsigned-ca
  namespace: tls
spec:
  isCA: true
  commonName: selfsigned-ca
  secretName: root-secret
  issuerRef:
    name: selfsigned-issuer
    kind: ClusterIssuer
    group: cert-manager.io
---
apiVersion: cert-manager.io/v1
kind: Issuer                                  #4
metadata:
  name: ca-issuer
  namespace: tls
spec:
  ca:
    secretName: root-secret

Certificate authority that generates certificates cluster-wide
Create a namespace for our demo
Namespaced root certificate using the cluster-wide issuer. Only used to create a namespaced issuer
Namespaced issuer. Used to create all other certificates in the post

After applying the previous manifest, we should be able to see the single certificate that we created:

kubectl get certificate -n tls

NAME            READY   SECRET        AGE
selfsigned-ca   True    root-secret   7s

The certificate infrastructure is ready; let's look at Apache APISIX.

Quick Overview of a Sample Apache APISIX Architecture

Apache APISIX is an API Gateway. By default, it stores its configuration in etcd, a distributed key-value store - the same one used by Kubernetes. Note that in real-world scenarios, we should set up etcd clustering to improve the resiliency of the solution. For this post, we will limit ourselves to a single etcd instance. Apache APISIX offers an admin API via HTTP endpoints. Finally, the gateway forwards calls from the client to an upstream. Here's an overview of the architecture and the required certificates:

Apache APISIX architecture

Let's start with the foundational bricks: etcd and Apache APISIX. We need two certificates: one for etcd, in the server role, and one for Apache APISIX, as the etcd client.

Let's set up certificates from our namespaced issuer:

apiVersion: cert-manager.io/v1
kind: Certificate
metadata:
  name: etcd-server                         #1
  namespace: tls
spec:
  secretName: etcd-secret                   #2
  isCA: false
  usages:
    - client auth                           #3
    - server auth                           #3
  dnsNames:
    - etcd                                  #4
  issuerRef:
    name: ca-issuer                         #5
    kind: Issuer
---
apiVersion: cert-manager.io/v1
kind: Certificate
metadata:
  name: apisix-client                       #6
  namespace: tls
spec:
  secretName: apisix-client-secret
  isCA: false
  usages:
    - client auth
  emailAddresses:
    - apisix@apache.org                     #7
  issuerRef:
    name: ca-issuer                         #5
    kind: Issuer

Certificate for etcd
Kubernetes Secret name, see below
Usages for this certificate
Kubernetes Service name, see below
Reference the previously namespaced issuer created earlier
Certificate for Apache APISIX as a client of etcd
Mandatory attribute for clients

After applying the above manifest, we can list the certificates in the tls namespace:

kubectl get certificates -n tls

NAME              READY   SECRET                 AGE
selfsigned-ca     True    root-secret            8m59s    //1
apisix-client     True    apisix-client-secret   8m22s    //2
etcd-server       True    etcd-secret            8m54s    //2

Previously created certificate
Newly-created certificates signed by selfsigned-ca

cert-manager's Certificates

So far, we have created Certificate objects, but we haven't explained what they are. Indeed, they are simple Kubernetes CRDs provided by cert-manager. Under the cover, cert-manager creates a Kubernetes Secret from a Certificate. It manages the whole lifecycle, so deleting a Certificate deletes the bounded Secret. The secretName attribute in the above manifest sets the Secret name.

kubectl get secrets -n tls

NAME                   TYPE                DATA   AGE
apisix-client-secret   kubernetes.io/tls   3      35m
etcd-secret            kubernetes.io/tls   3      35m
root-secret            kubernetes.io/tls   3      35m

Let's look at a Secret, e.g., apisix-client-secret:

kubectl describe apisix-client-secret -n tls

Name:         apisix-client-secret
Namespace:    tls
Labels:       controller.cert-manager.io/fao=true
Annotations:  cert-manager.io/alt-names:
              cert-manager.io/certificate-name: apisix-client
              cert-manager.io/common-name:
              cert-manager.io/ip-sans:
              cert-manager.io/issuer-group:
              cert-manager.io/issuer-kind: Issuer
              cert-manager.io/issuer-name: ca-issuer
              cert-manager.io/uri-sans:

Type:  kubernetes.io/tls

Data
====
ca.crt:   1099 bytes
tls.crt:  1115 bytes
tls.key:  1679 bytes

A Secret created by a Certificate provides three attributes:

tls.crt: The certificate itself
tls.key: The private key
ca.crt: The signing certificate in the certificate chain, i.e., root-secret/tls.crt

Kubernetes encodes Secret content in base 64. To get any of the above in the plain text, one should decode it, e.g.:

 kubectl get secret etcd-secret -n tls -o jsonpath='{ .data.tls\.crt }' | base64

Configuring mTLS Between etcd and APISIX

With the certificates available, we can now configure mutual TLS between etcd and APISIX. Let's start with etcd:

apiVersion: v1
kind: Pod
metadata:
  name: etcd
  namespace: tls
  labels:
    role: config
spec:
  containers:
    - name: etcd
      image: bitnami/etcd:3.5.7
      ports:
        - containerPort: 2379
      env:
        - name: ETCD_TRUSTED_CA_FILE        #1
          value: /etc/ssl/private/ca.crt
        - name: ETCD_CERT_FILE              #2
          value: /etc/ssl/private/tls.crt
        - name: ETCD_KEY_FILE               #3
          value: /etc/ssl/private/tls.key
        - name: ETCD_ROOT_PASSWORD
          value: whatever
        - name: ETCD_CLIENT_CERT_AUTH       #4
          value: "true"
        - name: ETCD_LISTEN_CLIENT_URLS
          value: https://0.0.0.0:2379
      volumeMounts:
        - name: ssl
          mountPath: /etc/ssl/private       #5
  volumes:
    - name: ssl
      secret:
        secretName: etcd-secret             #5

Set the trusted CA
Set the certificate
Set the private key
Require clients to pass their certificate, hence ensuring mutual authentication
Mount the previously generated secret in the container for access

Now, it's Apache APISIX's turn:

apiVersion: v1
kind: ConfigMap                                            #1
metadata:
  name: apisix-config
  namespace: tls
data:
  config.yaml: >-
    apisix:
      ssl:
        ssl_trusted_certificate: /etc/ssl/certs/ca.crt     #2
    deployment:
      etcd:
        host:
          - https://etcd:2379
        tls:
          cert: /etc/ssl/certs/tls.crt                     #2
          key: /etc/ssl/certs/tls.key                      #2
      admin:
        allow_admin:
          - 0.0.0.0/0
        https_admin: true                                  #3
        admin_api_mtls:
          admin_ssl_cert: /etc/ssl/private/tls.crt         #3
          admin_ssl_cert_key: /etc/ssl/private/tls.key     #3
          admin_ssl_ca_cert: /etc/ssl/private/ca.crt       #3
---
apiVersion: v1
kind: Pod
metadata:
  name: apisix
  namespace: tls
  labels:
    role: gateway
spec:
  containers:
    - name: apisix
      image: apache/apisix:3.2.0-debian
      ports:
        - containerPort: 9443                              #4
        - containerPort: 9180                              #5
      volumeMounts:
        - name: config                                      #1
          mountPath: /usr/local/apisix/conf/config.yaml
          subPath: config.yaml
        - name: ssl                                        #6
          mountPath: /etc/ssl/private
        - name: etcd-client                                #7
          mountPath: /etc/ssl/certs
  volumes:
    - name: config
      configMap:
        name: apisix-config
    - name: ssl                                            #6,8
      secret:
        secretName: apisix-server-secret
    - name: etcd-client                                    #7,8
      secret:
        secretName: apisix-client-secret

Apache APISIX doesn't offer configuration via environment variables. We need to use a ConfigMap that mirrors the regular config.yaml file
Configure client authentication for etcd
Configure server authentication for the Admin API
Regular HTTPS port
Admin HTTPS port
Certificates for server authentication
Certificates for client authentication
Two sets of certificates are used, one for server authentication for the Admin API and regular HTTPS, and one for client authentication for etcd.

At this point, we can apply the above manifests and see the two pods communicating. When connecting, Apache APISIX sends its apisix-client certificate via HTTPS. Because an authority signs the certificate that etcd trusts, it allows the connection.

I've omitted the Service definition for brevity's sake, but you can check them in the associated GitHub repo.

NAME     READY   STATUS    RESTARTS   AGE
apisix   1/1     Running   0          179m
etcd     1/1     Running   0          179m

Client Access

Now that we've set up the basic infrastructure, we should test accessing it with a client. We will use our faithful curl, but any client that allows configuring certificates should work, e.g, httpie.

The first step is to create a dedicated certificate-key pair for the client:

apiVersion: cert-manager.io/v1
kind: Certificate
metadata:
  name: curl-client
  namespace: tls
spec:
  secretName: curl-secret
  isCA: false
  usages:
    - client auth
  emailAddresses:
    - curl@localhost.dev
  issuerRef:
    name: ca-issuer
    kind: Issuer

curl requires a path to the certificate file instead of the content. We can go around this limitation through the magic of zsh: the =( ... ) syntax allows the creation of a temporary file. If you're using another shell, you'll need to find the equivalent syntax or download the files manually.

Let's query the Admin API for all existing routes. This simple command allows checking that Apache APISIX is connected to etcd, and it can read its configuration from there.

curl --resolve 'admin:32180:127.0.0.1' https://admin:32180/apisix/admin/routes \                     #1
     --cert =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.crt }' | base64 -d) \  #2
     --key =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.key }' | base64 -d) \   #2
     --cacert =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.ca\.crt }' | base64 -d) \ #2
     -H 'X-API-KEY: edd1c9f034335f136f87ad84b625c8f1'

--resolve avoids polluting one's /etc/hosts file. curl will translate admin to localhost, but the query is sent to admin inside the Kubernetes cluster, thus using the correct Service
Get the required data inside the Secret, decode it, and use it as a temporary file

If everything works, and it should, the result should be the following:

{"total":0,"list":[]}

No routes are available so far because we have yet to create any.

TLS with Upstreams

Last but not least, we should configure TLS for upstreams. In the following, I'll use a simple nginx instance that responds with static content. Use it as an illustration for more complex upstreams.

The first step, as always, is to generate a dedicated Certificate for the upstream. I'll skip how to do it as we already created a few. I call it upstream-server and its Secret, unimaginatively, upstream-secret. We can now use the latter to secure NGINX:

apiVersion: v1
kind: ConfigMap                                           #1
metadata:
  name: nginx-config
  namespace: tls
data:
  nginx.conf: >-
    events {
      worker_connections 1024;
    }
    http {
      server {
        listen              443 ssl;
        server_name         upstream;
        ssl_certificate     /etc/ssl/private/tls.crt;     #2
        ssl_certificate_key /etc/ssl/private/tls.key;     #2

        root /www/data;
        location / {
            index index.json;
        }
      }
    }
---
apiVersion: v1
kind: Pod
metadata:
  name: upstream
  namespace: tls
  labels:
    role: upstream
spec:
  containers:
    - name: upstream
      image: nginx:1.23-alpine
      ports:
        - containerPort: 443
      volumeMounts:
        - name: config
          mountPath: /etc/nginx/nginx.conf                #1
          subPath: nginx.conf
        - name: content
          mountPath: /www/data/index.json                 #3
          subPath: index.json
        - name: ssl                                       #2
          mountPath: /etc/ssl/private
  volumes:
    - name: config
      configMap:
        name: nginx-config
    - name: ssl                                           #2
      secret:
        secretName: upstream-secret
    - name: content                                       #3
      configMap:
        name: nginx-content

NGINX doesn't allow configuration via environment variables; we need to use the ConfigMap approach
Use the key-certificate pair created via the Certificate
Some static content unimportant in the scope of this post

The next step is to create the route with the help of the Admin API. We prepared everything in the previous step; now we can use the API:

curl --resolve 'admin:32180:127.0.0.1' https://admin:32180/apisix/admin/routes/1 \
     --cert =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.crt }' | base64 -d) \     #1
     --key =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.key }' | base64 -d) \      #1
     --cacert =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.ca\.crt }' | base64 -d) \    #1
     -H 'X-API-KEY: edd1c9f034335f136f87ad84b625c8f1' -X PUT -i -d "{
        \"uri\": \"/\",
        \"upstream\": {
          \"scheme\": \"https\",                                                                        #2
          \"nodes\": {
            \"upstream:443\": 1
          },
          \"tls\": {
            \"client_cert\": \"$(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.crt }' | base64 -d)\", #3
            \"client_key\": \"$(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.key }' | base64 -d)\"   #3
          }
        }
     }"

Client auth for Admin API, as above
Use HTTPS for the upstream
Configure key-certificate pair for the route. Apache APISIX stores the data in etcd and will use them when you call the route. Alternatively, you can keep the pair as a dedicated object and use the newly-created reference (just like for upstreams). It depends on how many routes the certificate needs. For more information, check the SSL endpoint

Finally, we can check it works as expected:

curl --resolve 'upstream:32443:127.0.0.1' https://upstream:32443/ \
     --cert =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.crt }' | base64 -d) \
     --key =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.tls\.key }' | base64 -d) \
     --cacert =(kubectl get secret curl-secret -n tls -o jsonpath='{ .data.ca\.crt }' | base64 -d)

And it does:

{ "hello": "world" }

Conclusion

In this post, I've described a working Apache APISIX architecture and implemented mutual TLS between all the components: etcd and APISIX, client and APISIX, and finally, client and upstream. I hope it will help you to achieve the same.

To go further: