Deploying Envoy With a Python Flask Web App and Kubernetes

In the first post in this series, Getting started with Lyft Envoy for microservice resilience, we discussed how Envoy is a powerful proxy that can be used for managing your traffic layer. We explored Envoy a bit, dug into a bit of how it works, and promised to actually deploy a real application using Kubernetes, Postgres, Flask, and Envoy. This time around we’ll make good on that promise.

The Application

What we’re going to do in this tutorial is to deploy a very, very simple REST-based user service: it can create users, read information about a user, and process simple logins. Obviously this isn’t terribly interesting by itself, but it brings several real-world concerns together:

• It requires persistent storage, so we’ll have to tackle that early.
• It will let us explore scaling the different pieces of the application.
• It will let us explore Envoy at the edge, where the user’s client talks to our application, and,
• It will let us explore Envoy internally, brokering communications between the various parts of the application.

Since Envoy is language-agnostic, we can use anything we like for the service itself. For this series, we’ll pick on Flask, both because it’s simple and because I like Python. On the database side, we’ll use PostgreSQL – it has good Python support, and it’s easy to get running both locally and in the cloud. And we’ll manage the whole thing with Kubernetes.

Kubernetes

Kubernetes is Datawire’s go-to container orchestrator these days, mostly because it does a fairly good job of letting you use the same tools whether you’re doing local development or deploying into the cloud for production. To get rolling today, we’ll need a Kubernetes cluster in which we’ll work. Within our cluster, we’ll create deployments that run the individual pieces of our application, and then expose services provided by those deployments (and when we do that, we get to decide whether to expose the service to the world outside the cluster, or only to other cluster members).

We’ll start out using Minikube to create a simple Kubernetes cluster running locally. The existence of Minikube is one of the things I really like about Kubernetes – it gives me an environment that’s almost like running Kubernetes somewhere out in the cloud, but it’s entirely local and (with some care) it can keep working at 30,000 feet on an airplane with no WiFi.

Note, though, that I said almost like running in the cloud. In principle, Kubernetes is Kubernetes and where you’re running doesn’t matter. In reality, of course, it does matter: networking, in particular, is something that ends up varying a bit depending on how you’re running your cluster. So getting running in Minikube is a great first step, but we’ll have to be aware that things will probably break a little bit as we move into the cloud. (If you’re setting up Kubernetes on AWS, this article can help.)

Setting Up

Minikube

Of course, you’ll need Minikube installed. See https://github.com/kubernetes/minikube/releases for more here. Mac users might also consider:

brew cask install minikube

Once Minikube is installed, you’ll need to start it. Mac users may want the xhyve driver to avoiding needing to install VirtualBox:

minikube start --vm-driver xhyve

Alternately:

minikube start

will fire things up with the default driver.

Kubernetes

To be able to work with Minikube, you’ll need the Kubernetes CLI, kubectl . Instructions are at https://kubernetes.io/docs/user-guide/prereqs/ — or, on a Mac, just use

brew install kubernetes-cli

Docker

Finally, you’ll also need the Docker command-line tool, in order to build the Docker images we’ll be working with. See https://www.docker.com/products/overview#/install_the_platform for more on this — or, on a Mac, just use

brew install docker

Minikube starts a Docker daemon when it starts up, and we need to use it for our Docker image builds so that the Minikube containers can load our images. To set up your Docker command-line tools for that:

eval $(minikube docker-env)

The Application

All the code and configuration we’ll use in this demo is on GitHub at https://github.com/datawire/envoy-steps.

Grab a clone of that, and cd into it. If you’re in the right place, you’ll see a README.md and directories named postgres, usersvc, etc. Each of the directories is for a Kubernetes deployment, and each can be brought up or down independently with

sh $service/up.sh

or

sh $service/down.sh

Obviously, I’d prefer to simply include everything you need in this blog post, but between Python code, all the Kubernetes config, docs, etc, there’s just too much. So we’ll hit the highlights here, and you can look at the details to your heart’s content in your clone of the repo.

Database Matters

Our database can be really simple — we just need a single table to store our user information. We can start by writing the Flask app to check at boot time and create our table if it doesn’t exist, relying on Postgres itself to make sure that only one table ever exists. Later, as we look into multiple Postgres servers, we may need to change this — but let’s keep it simple for now.

So the only thing we really need is a way to spin up a Postgres server in our Kubernetes cluster. Fortunately, there’s a published Postgres 9.6 Docker image readily accessible, so creating the Postgres deployment is pretty easy. The relevant config file is postgres/deployment.yaml, which includes in its spec section the specifics of the image we’ll use:

spec: 
  containers: 
  - name: postgres 
    image: postgres:9.6

Given the deployment, we also need to expose the Postgres service within our cluster. That’s defined in postgres/service.yaml with highlights:

spec: 
  type: ClusterIP 
  ports: 
  - name: postgres 
    port: 5432 
  selector: 
    service: postgres

Note that we mark this with type ClusterIP, so that it can be seen only within the cluster.

To fire this up, just run:

sh postgres/up.sh

or if you’d rather, you can do it by hand:

kubectl create -f postgres/deployment.yaml
kubectl create -f postgres/service.yaml

Once that’s done (whether by hand or with our script) then kubectl get pods should show the postgres pod running:

NAME                       READY  STATUS   RESTARTS AGE
postgres-1385931004-p3szz  1/1    Running  0        5s

and kubectl get services should show its service:

NAME      CLUSTER-IP     EXTERNAL-IP  PORT(S)   AGE
postgres  10.107.246.55  <none>       5432/TCP  5s

So we now have a running Postgres server, reachable from anywhere in the cluster at postgres:5432.

The Flask App

Our Flask app is really simple: basically it just responds to PUT requests to create users, and GET requests to read users and respond to health checks. You can see it in full in the GitHub repo.

The only real gotcha is that by default, Flask will listen only on the loopback address, which will prevent any connections from outside the Flask app’s container. We set the Flask app to explicitly listen on 0.0.0.0 instead, so that we can actually speak to it from elsewhere (whether from in the cluster or outside).

To get the app running in Kubernetes, we’ll need a Docker image that contains our app. We’ll build this on top of the lyft/envoy image, since we already know we’re headed for Envoy later — thus our Dockerfile (sans comments) ends up looking like this:

FROM lyft/envoy:latest
RUN apt-get update && apt-get -q install -y 
    curl 
    python-pip 
    dnsutils
WORKDIR /application
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY service.py .
COPY entrypoint.sh .
RUN chmod +x entrypoint.sh
ENTRYPOINT [ "./entrypoint.sh" ]

We’ll build that into a Docker image, then fire up a Kubernetes deployment and service with it. The deployment, in  usersvc/deployment.yaml , looks basically the same as the one for postgres, just with a different image name:

spec:
  containers:
  - name: usersvc
    image: usersvc:step1

Likewise, usersvc/service.yaml is much like its postgres sibling, but we’re using type LoadBalancer to indicate that we want the service exposed to users outside the cluster:

spec:
  type: LoadBalancer
  ports:
  - name: usersvc
    port: 5000
    targetPort: 5000
  selector:
    service: usersvc

It may seem odd to be starting with LoadBalancer here — after all, we want to use Envoy to do load balancing, right? The point is walking before running: our first test will be to talk to our service without Envoy, and for that we need to expose the port to the outside world.

To build the Docker image and crank up the service, you can run

sh usersvc/up.sh

or once again you can do it all by hand:

docker build -t usersvc:step1 usersvc
kubectl create -f usersvc/deployment.yaml
kubectl create -f usersvc/service.yaml

(Note that we didn’t push the Docker image anywhere: one of the benefits of Minikube is that a container can always pull your images from the local Docker daemon started by Minikube. When we look at getting out into the cloud, obviously this won’t work, so expect to see some changes around this in subsequent posts.)

At this point, kubectl get pods should show both the usersvc pod and the postgres pod running:

NAME                       READY  STATUS   RESTARTS AGE
postgres-1385931004-p3szz  1/1    Running  0        5m
usersvc-1941676296-kmglv   1/1    Running  0        5s

First Test!

And now for the moment of truth: let’s see if it works without Envoy before moving on! This will require us to get the IP address and mapped port number for the usersvc service. Since we’re using Minikube, we use:

minikube service --url usersrvc

to get a neatly-formed URL to our usersvc. (Obviously, this will change when we move beyond Minikube.)

Let’s start with a basic health check using curl  from the host system, reaching into the cluster to the usersvc, which in turn is talking within the cluster to postgres:

curl $(minikube service --url usersvc)/user/health

If all goes well, the health check should return something like:

{
  "hostname": "usersvc-1941676296-kmglv",
  "msg": "user health check OK",
  "ok": true,
  "resolvedname": "172.17.0.10"
}

Next up we can try saving and retrieving a user:

curl -X PUT 
     -H "Content-Type: application/json"  
     -d '{ "fullname": "Alice", "password": "alicerules" }'  
$(minikube service --url usersvc)/user/alice

This should give us a user record for Alice, including her UUID but not her password:

{ 
  "fullname": "Alice", 
  "hostname": "usersvc-1941676296-kmglv", 
  "ok": true, 
  "resolvedname": "172.17.0.10", 
  "uuid": "44FD5687B15B4AF78753E33E6A2B033B"
}

If we repeat it for Bob, we should get much the same:

curl -X PUT  
     -H "Content-Type: application/json"  
     -d '{ "fullname": "Bob", "password": "bobrules" }'  
$(minikube service --url usersvc)/user/bob

Note, of course, that Bob should have a different UUID:

{ 
  "fullname": "Bob", 
  "hostname": "usersvc-1941676296-kmglv", 
  "ok": true, 
  "resolvedname": "172.17.0.10", 
  "uuid": "72C77A08942D4EADA61B6A0713C1624F"
}

Finally, we should be able to read both users back (again, minus passwords!) with

curl $(minikube service --url usersvc)/user/alice
curl $(minikube service --url usersvc)/user/bob

Enter Envoy

Given that all of that is working (whew!)… it’s time to stick Envoy in front of everything, so it can manage routing when we start scaling the front end. As we discussed in the previous article, this means that we have an edge Envoy and an application Envoy, each of which needs is own configuration. We’ll crank up the edge Envoy first.

Since the edge Envoy runs in its own container, we’ll need a separate Docker image for it. Here’s the Dockerfile:

FROM lyft/envoy:latest
RUN apt-get update && apt-get -q install -y 
    curl 
    dnsutils
COPY envoy.json /etc/envoy.json
CMD /usr/local/bin/envoy -c /etc/envoy.json

which is to say, we take lyft/envoy:latest, copy in our own Envoy config, and start Envoy running.

Our edge Envoy’s config is fairly simple, too, since it only needs to proxy any URL starting with /user to our usersvc. Here’s how you set up virtual_hosts for that:

"virtual_hosts": [ 
  {
    "name": "service",
    "domains": [ "*" ],
    "routes": [
      {
        "timeout_ms": 0,
        "prefix": "/user",
        "cluster": “usersvc"
      }
    ]
  }
]

and here’s the related clusters section:

"clusters": [
  {
    "name": “usersvc”,
    "type": "strict_dns",
    "lb_type": "round_robin",
    "hosts": [
      {
        "url": “tcp://usersvc:80”
      }
    ]
  }
]

Note that we’re using strict_dns, which means that we’re relying on every instance of the usersvc appearing in the DNS. We’ll find out if this actually works shortly!

As usual, you can build the Docker image and crank up the edge Envoy the simple way:

sh edge-envoy/up.sh

or the complex way:

docker build -t edge-envoy:step2 edge-envoy
kubectl create -f edge-envoy/deployment.yaml
kubectl create -f edge-envoy/service.yaml

Sadly, we can’t really test anything yet, since the edge Envoy is going to try to talk to application Envoys that aren’t running yet.

App Changes for Envoy

Once the edge Envoy is running, we need to switch our Flask app to use an application Envoy. We needn’t change the database at all, but the Flask app needs a few tweaks:

• We need to have the Dockerfile copy in an Envoy config file.
• We need to have the entrypoint.sh script start Envoy as well as the Flask app.
• While we’re at it, we can switch back to having Flask listen only on the loopback interface, and,
• We’ll switch the service from a LoadBalancer to a ClusterIP.

The effect here is that we’ll have a running Envoy through which we can talk to the Flask app — but also that Envoy will be the only way to talk to the Flask app. Trying to go direct will be blocked in the network layer.

The application Envoy’s config, while we’re at it, is very similar to the edge Envoy’s. The listeners section is actually identical, and the clusters section nearly so:

"clusters": [ 
  {
    "name": “usersvc”,
    "type": "static",
    "lb_type": "round_robin",
    "hosts": [
      {
        "url": “tcp://127.0.0.1:80”
      }
    ]
  }
]

Basically, we just use a static single-member cluster, with only localhost listed.

All the changes to the Flask side of the world can be found in the usersvc2 directory, which is literally a copy of the usersvc directory with the changes we discussed above for the Flask side of the world (and it tags its image usersvc:step2 instead of usersvc:step1). The easy way to handle things at this point is:

sh usersvc/down.sh
sh usersvc2/up.sh

to drop the old version and bring up the new.

Second Test!

Once all that is done, voilà: you should be able to retrieve Alice and Bob from before:

curl $(minikube service --url edge-envoy)/user/alice
curl $(minikube service --url edge-envoy)/user/bob

…but note that we’re using the edge-envoy service here, not the usersvc, which means that we are indeed talking through the Envoy mesh! In fact, if you try talking directly to usersvc, it will fail: that’s part of how we can be sure that Envoy is doing its job.

Scaling the Flask App

One of the promises of Envoy is helping with scaling applications. Let’s see how well it handles that by scaling up to multiple instances of our Flask app:

kubectl scale --replicas=3 deployment/usersvc

Once that’s done, kubectl get pods should show more usersvc instances running:

NAME                         READY STATUS   RESTARTS  AGE
edge-envoy-2874730579-7vrp4  1/1   Running  0         3m
postgres-1385931004-p3szz    1/1   Running  0         5m
usersvc-2016583945-h7hqz     1/1   Running  0         6s
usersvc-2016583945-hxvrn     1/1   Running  0         6s
usersvc-2016583945-pzq2x     1/1   Running  0         3m

and we should then be able to see curl getting routed to multiple hosts. Try running:

curl $(minikube service --url edge-envoy)/user/health

multiple times, and look at the hostname element. It should be cycling across our three usersvc nodes.

But it’s not. Uh oh. What’s going on here?

Remembering that we’re running Envoy in strict_dns mode, a good first check would be to look at the DNS. We can do this by running nslookup from inside the cluster. Specifically, we can use a usersvc pod:

kubectl exec usersvc-2016583945-h7hqz /usr/bin/nslookup usersvc

(Make sure to use one of your pod names when you run this! Just pasting the line above is extremely unlikely to work.)

Running this check, we find that only one address comes back — so Envoy’s DNS-based service discovery simply isn’t going to work. Envoy can’t round-robin among our three service instances if it never hears about two of them.

The Service Discovery Service

What’s going on here is that Kubernetes puts each service into its DNS, but it doesn’t put each service endpoint into its DNS — and we need Envoy to know about the endpoints in order to load-balance. Thankfully, Kubernetes does know the service endpoints for each service, and Envoy knows how to query a REST service for discovery information. We can make this work with a simple Python shim that bridges from the Envoy “Service Discovery Service” (SDS) to the Kubernetes API.

(There’s also the Istio project, which is digging into a more full-featured solution here. Istio is still in its very early stages, though, so we’re going to stick with the simple way here.)

Our SDS is in the usersvc-sds directory. It’s pretty straightforward: when Envoy asks it for service information, it uses the requests Python module to query the Kubernetes endpoint's API, and reformats the results for Envoy. The most bizarre bit might be the token it reads at the start: Kubernetes is polite enough to install an authentication token on every container it starts, precisely so that this sort of thing is possible.

We also need to modify the edge Envoy’s config slightly: rather than using strict_dns mode, we need sds mode. In turn, that means we have to define an sds cluster (which uses DNS to locate its server at the moment — we may have to tackle that later, too, as we scale the SDS out!):

"cluster_manager": { 
  "sds": {
    "cluster": {
      "name": "usersvc-sds",
      "connect_timeout_ms": 250,
      "type": "strict_dns",
      "lb_type": "round_robin",
      "hosts": [
        {
          "url": "tcp://usersvc-sds:5000"
        }
      ]
    },
    "refresh_delay_ms": 15000 
  },
  "clusters": [
    {
      "name": "usersvc",
      "connect_timeout_ms": 250,
      "type": "sds",
      "service_name": "usersvc",
      "lb_type": "round_robin",
      "features": "http2"
    }
  ]
}

Look carefully: the sds cluster is not defined inside the clusters dictionary, but as a peer of clusters. Its value is a cluster definition, though. Once the sds cluster is defined, you can simply say "type": "sds" in a service cluster definition, and delete any hosts array for that cluster.

The edge-envoy2 directory has everything set up for an edge Envoy running this config. So let’s crank up the SDS, then down the old edge Envoy and fire up the new:

sh usersvc-sds/up.sh
sh edge-envoy/down.sh
sh edge-envoy2/up.sh

Now, repeating our health check really should show you round-robining around the hosts.
But, of course, asking for the details of user Alice should always give the same results, no matter which host does the database lookup:

curl $(minikube service --url edge-envoy)/user/alice

If you repeat that a few times, the host information should change, but the user information should not.

Up Next

We have everything working, including using Envoy to handle round-robining traffic between our several Flask apps. With kubectl scale, we can easily change the number of instances of Flask apps we’re running. And, as you probably noticed, bringing Envoy in once our app was running wasn’t hard. It’s a pretty promising way to add a lot of flexibility without a lot of pain.

Next up: Google Container Engine and AWS. One of the promises of Kubernetes is being able to easily get stuff deployed in multiple environments, so we’re going to see whether that actually works.