Skip to content

Latest commit

 

History

History
385 lines (281 loc) · 10.9 KB

lesson4.md

File metadata and controls

385 lines (281 loc) · 10.9 KB
Raw

Further developing libraries

When starting out with Jsonnet, it is very likely that you'll with existing code and dependencies. Developing on these can seem confusing and tedious, however there are several techniques that can help us iterate at different velocities.

Objectives

  • Wrap and extend a dependency locally
  • Developing on upstream libraries

Lesson

Wrapping libraries

As shown in the refactoring exercise, let's initialize a new project with k8s-libsonnet:

$ jb init

$ jb install github.com/jsonnet-libs/k8s-libsonnet/1.23@main

$ mkdir lib

$ echo "(import 'github.com/jsonnet-libs/k8s-libsonnet/1.23/main.libsonnet')" > lib/k.libsonnet

And install privatebin-libsonnet:

$ jb install github.com/Duologic/privatebin-libsonnet

local privatebin = import 'github.com/Duologic/privatebin-libsonnet/main.libsonnet';

{
  privatebin: privatebin.new(),
}

// example1/example1.jsonnet

This is relatively simple web application that exposes itself on port 8080.

local k = import 'ksonnet-util/kausal.libsonnet';

{
  new(
    name='privatebin',
    image='privatebin/nginx-fpm-alpine:1.3.5',
  ):: {

    local container = k.core.v1.container,
    container::
      container.new('privatebin', image)
      + container.withPorts([
        k.core.v1.containerPort.newNamed(name='http', containerPort=8080),
      ])
      + k.util.resourcesRequests('50m', '100Mi')
      + k.util.resourcesLimits('150m', '300Mi')
      + container.livenessProbe.httpGet.withPath('/')
      + container.livenessProbe.httpGet.withPort('http')
      + container.readinessProbe.httpGet.withPath('/')
      + container.readinessProbe.httpGet.withPort('http')
    ,

    local deployment = k.apps.v1.deployment,
    deployment: deployment.new(name, 1, [self.container]),

    service: k.util.serviceFor(self.deployment),
  },
}

// example1/vendor/github.com/Duologic/privatebin-libsonnet/main.libsonnet

privatebin-libsonnet is relatively simple, it already exposes the container separate from the deployment and we can pick the name and image when initializing it with new().

One hurdle: the library does not expose a function to change the port. We could go about and create a change request upstream, but considering this is a very trivial change it might not be worth the time.

So, let's try to deal with it locally.

ksonnet-util

The ksonnet-util library contains a set of util functions developed around the now deprecated ksonnet-lib. It also functions as a compatibility layer towards k8s-libsonnet. Many of the util functions have since made it into k8s-libsonnet with the aim to make this library obsolete over time.

For the purpose of this exercise we want to run this application for different teams and each team should be able to change the port to their liking.

Create a local library in lib/privatebin/:

local privatebin = import 'github.com/Duologic/privatebin-libsonnet/main.libsonnet';
local k = import 'k.libsonnet';

privatebin
{
  withPort(port): {
    container+:
      k.core.v1.container.withPorts([
        k.core.v1.containerPort.newNamed(
          name='http',
          containerPort=port
        ),
      ]),
  },
}

// example1/lib/privatebin/main.libsonnet

Here we import the vendored library and extend it. We are extending the privatebin library here extended with the withPort() function that will change the ports of the container.

Have a look at the container and containerPort documentation for the details on each.

local privatebin = import 'privatebin/main.libsonnet';

{
  privatebin:
    privatebin.new('backend')
    + privatebin.withPort(9000),
}

// example1/example2.jsonnet

To use the new library, we need to change the import to match privatebin/main.libsonnet, JSONNET_PATH will expand it to lib/privatebin/main.libsonnet.

This makes the withPort() function available and now teams can set their own port.

In this example, the backend team has named its privatebin backend, this will generate a deployment/service with that name. As the name is quite generic, this may cause conflicts.

Let's add a suffix to prevent the naming conflict:

local privatebin = import 'github.com/Duologic/privatebin-libsonnet/main.libsonnet';
local k = import 'k.libsonnet';

privatebin
{
  new(team):
    super.new(team + '-privatebin'),

  withPort(port): {
    container+:
      k.core.v1.container.withPorts([
        k.core.v1.containerPort.newNamed(
          name='http',
          containerPort=port
        ),
      ]),
  },
}

// example2/lib/privatebin/main.libsonnet

Here we call super.new(), this means it will use the new() function defined in privatebin.

The deployment/service will now be suffixed with -privatebin, ie. backend-privatebin.

Use case: replace Pentagon with External Secrets Operator

Wrapping libraries is a powerful concept, it can be used to manipulate or even replace whole systems that are used across a code base.

local pentagon = import 'github.com/grafana/jsonnet-libs/pentagon/pentagon.libsonnet';

pentagon
{
  _config+:: {
    local this = self,
    pentagon+: {
      vault_address:
        if this.cluster_name == 'dev'
        then 'vault-dev.example.com'
        else 'vault-prod.example.com',
    },
  },
}

// usecase-pentagon/lib/pentagon/example1.libsonnet
local pentagon = import 'pentagon/example1.libsonnet';

{
  pentagon: pentagon {
    _config+:: {
      cluster_name: 'dev',
      namespace: 'app1',
    },
    pentagon_mappings: [
      pentagon.pentagonKVMapping('path/to/secret', 'k8sSecretName'),
    ],
  },
}

// usecase-pentagon/example1.jsonnet

For synchronizing secrets between Vault and Kubernetes, Grafana Labs used a fork of Pentagon. This process was a deployment per namespace with each team managing their own deployments. To facilitate a consistent connection configuration we wrapped the pentagon library. Teams could create a Vault->Kubernetes mapping with the shortcut function pentagonKVmapping to populate the pentagon_mappings array (which gets turned into a configMap).

As a deployment per namespace accumulated quite a bit of resources, we opted to replace it with External Secrets Operator. This means we'd go from many deployments and configMaps (at least one of each per namespace) to a single operator deployment per cluster, a secretStore per namespace and many externalSecret objects.

local externalSecrets = import 'external-secrets-libsonnet/main.libsonnet';
local pentagon = import 'github.com/grafana/jsonnet-libs/pentagon/pentagon.libsonnet';

pentagon
{
  local this = self,

  // Remove/hide resources
  _config+:: {},
  deployment:: {},
  config_map:: {},
  rbac:: {},
  cluster_role:: {},
  cluster_role_binding:: {},

  local externalSecret = externalSecrets.nogroup.v1beta1.externalSecret,
  externalSecrets: std.sort([
    local mapping = this.pentagon_mappings_map[m];

    externalSecret.new(mapping.secretName)
    + externalSecret.spec.secretStoreRef.withName('vault-backend')
    + externalSecret.spec.secretStoreRef.withKind('SecretStore')
    + externalSecret.spec.target.withName(mapping.secretName)
    + externalSecret.spec.withDataFrom([
      {
        extract: {
          key: mapping.vaultPath,
        },
      },
    ])

    for m in std.objectFields(this.pentagon_mappings_map)
  ], function(e) if std.objectHasAll(e, 'idx') then e.idx else 0),
}

// usecase-pentagon/lib/pentagon/example2.libsonnet

The operator deployment and secretStore objects are managed centrally, only thing left to do is replace the secret mappings everywhere. With more than 2500 mappings, this would have been a hell of a refactoring job. Fortunately we had the wrapped library in place and we could transform the mappings array into externalSecret objects transparently.

As the wrapped library is aware of the cluster it was being deployed too, we were able gradually roll this out across the fleet.

Finally with this in place, we informed each team on how to refactor their code to use External Secrets directly, allowing them to work on it at their own pace.

Developing on vendored libraries

As it gives immediate feedback, it often happens that a vendored library is developed alongside the project that is using it. However any invocation of jb install will remove changes from vendor/, which makes it a little bit more challenging. Let's have a look at the different options.

Simply edit files in vendor/

This is the easiest except the risk of loosing changes is highest. For testing small changes this is probably safe, it gives immediate feedback whether the changes match expectations. The small changes should be pushed upstream early on so they don't get lost.

Clone dependency in vendor/

Similar but a bit more elaborate, it is possible to git clone the dependency straight into vendor/. It faces the same risk as above but allows for a shorter loop to push changes upstream.

Use local reference

$ jb install ../../lesson3/example2/lib/webserver

{
  "version": 1,
  "dependencies": [
    {
      "source": {
        "local": {
          "directory": "../../lesson3/example2/lib/webserver"
        }
      },
      "version": ""
    }
  ],
  "legacyImports": true
}

// example3/jsonnetfile.json
.
├── jsonnetfile.json
├── jsonnetfile.lock.json
└── vendor
    └── webserver -> ../../../lesson3/example2/lib/webserver

This installs a local library relative to the project root with a symlink. Changes made in vendor/ or on the real location are unaffected by jb install however it changes jsonnetfile.json to something that can't be shared.

Symlink in lib/

$ ln -s ../../../lesson3/example2/lib/webserver lib/

.
└── lib
    └── webserver -> ../../../lesson3/example2/lib/webserver

By relying on the import order, a symlink in lib/ could be made. With lib/ being matched before vendor/, it will be used first. This approach is unaffected by jb install and doesn't change jsonnetfile.json.

Conclusion

Wrapping libraries locally by leveraging the extensible nature of Jsonnet can be very useful. It can alter default behavior that is more suitable for the project.

Developing directly on vendored libraries on the other hand is quite clumsy, the techniques described require a bit of pragmatism to be useful. jsonnet-bundler could benefit from a feature to make this easier.