Add Kubernetes example

This example starts a fairly complete Kubernetes cluster and showcases
perfectly reproducible remote execution via the local remote execution
toolchain containers.

This example uses a three-layer setup process:

1. The infra layer is a kind cluster with Cilium and MetalLB. This layer
   is built to be easily swappable with more "production grade"
   clusters.
2. The operations layer deploys a few standard applications that are not
   inherently required for NativeLink, but are solid deployments that
   one would likely want running in a cluster. This includes monitoring
   and handling image availability.
3. The application layer is a straightforward `kubectl apply -k .` which
   deploys a NativeLink CAS, Worker and Scheduler. This deployment
   differs from the Docker Compose setup in that it does not make use of
   any system paths and doesn't allow visibility "outside" of the node
   itself. That is, it's a hard requirement that the worker image is
   self-contained. Storage is fully ephemeral in this example and a
   `kubectl delete -k .` will destroy the cache for quick iterations and
   cache testing.

.bazelrc

@@ -68,3 +68,6 @@ build:lre --define=EXECUTOR=remote
 
 # See: https://github.com/bazelbuild/bazel/issues/19714#issuecomment-1745604978
 build:lre --repo_env=BAZEL_DO_NOT_DETECT_CPP_TOOLCHAIN=1
+
+# Allow user-side customization.
+try-import %workspace%/.bazelrc.user

.github/workflows/lre.yaml

@@ -38,3 +38,62 @@ jobs:
            --config=lre \
            --verbose_failures \
            //local-remote-execution/examples:hello_lre"
+
+  remote:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-22.04]
+    name: Remote / ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    steps:
+      - name: Checkout
+        uses: >- # v4.1.1
+          actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11
+
+      - name: Install Nix
+        uses: >- #v7
+          DeterminateSystems/nix-installer-action@5620eb4af6b562c53e4d4628c0b6e4f9d9ae8612
+
+      - name: Cache Nix derivations
+        uses: >- # Custom commit, last pinned at 2023-11-17.
+          DeterminateSystems/magic-nix-cache-action@a04e6275a6bea232cd04fc6f3cbf20d4cb02a3e1
+
+      - name: Start Kubernetes cluster
+        run: >
+          nix develop --impure --command
+          bash -c "cd deployment-examples/kubernetes \
+            && ./00_infra.sh \
+            && ./01_operations.sh \
+            && ./02_application.sh"
+
+      - name: Get gateway IPs
+        id: gateway-ips
+        run: |
+          echo "cache_ip=$(kubectl get gtw cache -o=jsonpath='{.status.addresses[0].value}')" >> "$GITHUB_ENV"
+          echo "scheduler_ip=$(kubectl get gtw scheduler -o=jsonpath='{.status.addresses[0].value}')" >> "$GITHUB_ENV"
+
+      - name: Print cluster state
+        run: |
+          kubectl get svc -A
+          kubectl get pod -A
+          kubectl get svc -A
+          kubectl get deployments -A
+          kubectl describe gtw
+          echo "cas"
+          kubectl logs -l app=nativelink-cas
+          echo "scheduler"
+          kubectl logs -l app=nativelink-scheduler
+          echo "worker"
+          kubectl logs -l app=nativelink-worker
+
+      - name: Build hello_lre with LRE toolchain.
+        run: >
+          nix develop --impure --command
+          bash -c "bazel run \
+            --config=lre \
+            --remote_instance_name=main \
+            --remote_cache=grpc://$cache_ip:50051 \
+            --remote_executor=grpc://$scheduler_ip:50052 \
+            --verbose_failures \
+            //local-remote-execution/examples:hello_lre"

.gitignore

@@ -11,3 +11,4 @@ __pycache__
 .DS_Store
 .pre-commit-config.yaml
 result
+.bazelrc.user

deployment-examples/kubernetes/.gitignore

@@ -0,0 +1 @@
+worker.json

deployment-examples/kubernetes/00_infra.sh

@@ -0,0 +1,127 @@
+# This script sets up a local development cluster. It's roughly equivalent to
+# a managed K8s setup.
+
+# For ease of development and to save disk space we pipe a local container
+# registry through to kind.
+#
+# See https://kind.sigs.k8s.io/docs/user/local-registry/.
+
+reg_name='kind-registry'
+reg_port='5001'
+if [ "$(docker inspect -f '{{.State.Running}}' "${reg_name}" 2>/dev/null || true)" != 'true' ]; then
+  docker run \
+    -d --restart=always -p "127.0.0.1:${reg_port}:5000" --network bridge --name "${reg_name}" \
+    registry:2
+fi
+
+# Start a basic cluster. We use cilium's CNI and eBPF kube-proxy replacement.
+
+cat <<EOF |  kind create cluster --config -
+---
+kind: Cluster
+apiVersion: kind.x-k8s.io/v1alpha4
+nodes:
+  - role: control-plane
+  - role: worker
+  - role: worker
+networking:
+  disableDefaultCNI: true
+  kubeProxyMode: none
+containerdConfigPatches:
+  - |-
+    [plugins."io.containerd.grpc.v1.cri".registry]
+      config_path = "/etc/containerd/certs.d"
+EOF
+
+# Enable the registry on the nodes.
+
+REGISTRY_DIR="/etc/containerd/certs.d/localhost:${reg_port}"
+for node in $(kind get nodes); do
+  docker exec "${node}" mkdir -p "${REGISTRY_DIR}"
+  cat <<EOF | docker exec -i "${node}" cp /dev/stdin "${REGISTRY_DIR}/hosts.toml"
+[host."http://${reg_name}:5000"]
+EOF
+done
+
+# Connect the registry to the cluster network.
+
+if [ "$(docker inspect -f='{{json .NetworkSettings.Networks.kind}}' "${reg_name}")" = 'null' ]; then
+  docker network connect "kind" "${reg_name}"
+fi
+
+# Advertise the registry location.
+
+cat <<EOF | kubectl apply -f -
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: local-registry-hosting
+  namespace: kube-public
+data:
+  localRegistryHosting.v1: |
+    host: "localhost:${reg_port}"
+    help: "https://kind.sigs.k8s.io/docs/user/local-registry/"
+EOF
+
+# Prepare Gateway API CRDs. These MUST be available before we start cilium.
+
+kubectl apply -f https://github.com/kubernetes-sigs/gateway-api/releases/download/v1.0.0/experimental-install.yaml
+
+kubectl wait --for condition=Established crd/gatewayclasses.gateway.networking.k8s.io
+kubectl wait --for condition=Established crd/gateways.gateway.networking.k8s.io
+kubectl wait --for condition=Established crd/httproutes.gateway.networking.k8s.io
+kubectl wait --for condition=Established crd/tlsroutes.gateway.networking.k8s.io
+kubectl wait --for condition=Established crd/grpcroutes.gateway.networking.k8s.io
+kubectl wait --for condition=Established crd/referencegrants.gateway.networking.k8s.io
+
+# Start cilium.
+
+helm repo add cilium https://helm.cilium.io
+
+helm upgrade \
+    --install cilium cilium/cilium \
+    --version 1.15.0-pre.3 \
+    --namespace kube-system \
+    --set k8sServiceHost=kind-control-plane \
+    --set k8sServicePort=6443 \
+    --set kubeProxyReplacement=strict \
+    --set gatewayAPI.enabled=true \
+    --wait
+
+# Set up MetalLB. Kind's nodes are containers running on the local docker
+# network. We reuse that network for LB-IPAM so that LoadBalancers are available
+# via "real" local IPs.
+
+KIND_NET_CIDR=$(docker network inspect kind -f '{{(index .IPAM.Config 0).Subnet}}')
+METALLB_IP_START=$(echo ${KIND_NET_CIDR} | sed "[email protected]/[email protected]@")
+METALLB_IP_END=$(echo ${KIND_NET_CIDR} | sed "[email protected]/[email protected]@")
+METALLB_IP_RANGE="${METALLB_IP_START}-${METALLB_IP_END}"
+
+helm install --namespace metallb-system --create-namespace \
+  --repo https://metallb.github.io/metallb metallb metallb \
+  --version 0.13.12 \
+  --wait
+
+cat <<EOF | kubectl apply -f -
+---
+apiVersion: metallb.io/v1beta1
+kind: L2Advertisement
+metadata:
+  name: l2-ip
+  namespace: metallb-system
+spec:
+  ipAddressPools:
+    - default-pool
+---
+apiVersion: metallb.io/v1beta1
+kind: IPAddressPool
+metadata:
+  name: default-pool
+  namespace: metallb-system
+spec:
+  addresses:
+    - ${METALLB_IP_RANGE}
+EOF
+
+# At this point we have a similar setup to the one that we'd get with a cloud
+# provider. Move on to `01_operations.sh` for the cluster setup.

deployment-examples/kubernetes/01_operations.sh

@@ -0,0 +1,28 @@
+# This script configures a cluster with a few standard deployments.
+
+# TODO(aaronmondal): Add Grafana, OpenTelemetry and the various other standard
+#                    deployments one would expect in a cluster.
+
+kubectl apply -f gateway.yaml
+
+IMAGE_TAG=$(nix eval .#image.imageTag --raw)
+
+$(nix build .#image --print-build-logs --verbose) \
+    && ./result \
+    | skopeo \
+      copy \
+      --dest-tls-verify=false \
+      docker-archive:/dev/stdin \
+      docker://localhost:5001/nativelink:local
+
+IMAGE_TAG=$(nix eval .#lre.imageTag --raw)
+
+echo $IMAGE_TAG
+
+$(nix build .#lre --print-build-logs --verbose) \
+    && ./result \
+    | skopeo \
+      copy \
+      --dest-tls-verify=false \
+      docker-archive:/dev/stdin \
+      docker://localhost:5001/nativelink-toolchain:local

deployment-examples/kubernetes/02_application.sh

@@ -0,0 +1,15 @@
+# Get the nix derivation hash from the toolchain container, change the
+# `TOOLCHAIN_TAG` variable in the `worker.json.template` to that hash and apply
+# the configuration.
+
+KUSTOMIZE_DIR=$(git rev-parse --show-toplevel)/deployment-examples/kubernetes
+
+sed "s/__NATIVELINK_TOOLCHAIN_TAG__/$(nix eval .#lre.imageTag --raw)/g" \
+  "$KUSTOMIZE_DIR/worker.json.template" \
+  > "$KUSTOMIZE_DIR/worker.json"
+
+kubectl apply -k "$KUSTOMIZE_DIR"
+
+kubectl rollout status deploy/nativelink-cas
+kubectl rollout status deploy/nativelink-scheduler
+kubectl rollout status deploy/nativelink-worker

deployment-examples/kubernetes/03_delete_application.sh

@@ -0,0 +1,11 @@
+# Get the nix derivation hash from the toolchain container, change the
+# `TOOLCHAIN_TAG` variable in the `worker.json.template` to that hash and delete
+# the configuration.
+
+KUSTOMIZE_DIR=$(git rev-parse --show-toplevel)/deployment-examples/kubernetes
+
+sed "s/__NATIVELINK_TOOLCHAIN_TAG__/$(nix eval .#lre.imageTag --raw)/g" \
+  "$KUSTOMIZE_DIR/worker.json.template" \
+  > "$KUSTOMIZE_DIR/worker.json"
+
+kubectl delete -k "$KUSTOMIZE_DIR"

deployment-examples/kubernetes/README.md

@@ -0,0 +1,76 @@
+# Kubernetes example
+
+This deployment sets up a 3-container deployment with separate CAS, scheduler
+and worker. Don't use this example deployment in production. It's insecure.
+
+In this example we're using `kind` to set up the cluster and `cilium` with
+`metallb` to provide a `LoadBalancer` and `GatewayController`.
+
+First set up a local development cluster:
+
+```bash
+./00_infra.sh
+```
+
+Next start a few standard deployments. This part also builds the remote
+execution containers and makes them available to the cluster:
+
+```bash
+./01_operations.sh
+```
+
+Finally deploy NativeLink:
+
+```bash
+./02_application.sh
+```
+
+> [!TIP]
+> You can use `./03_delete_application.sh` to remove just the nativelink
+> deployments but leave the rest of the cluster intact.
+
+This demo setup creates two gateways to expose the `cas` and `scheduler`
+deployments via your local docker network:
+
+```bash
+CACHE=$(kubectl get gtw cache -o=jsonpath='{.status.addresses[0].value}')
+SCHEDULER=$(kubectl get gtw scheduler -o=jsonpath='{.status.addresses[0].value}')
+
+echo "Cache IP: $CACHE"
+echo "Scheduler IP: $SCHEDULER"
+
+# Prints something like:
+#
+# Cache IP: 172.20.255.200
+# Scheduler IP: 172.20.255.201
+```
+
+You can now pass these IPs to your bazel invocation to use the remote cache and
+executor:
+
+```bash
+bazel test \
+    --config=lre \
+    --remote_instance_name=main \
+    --remote_cache=grpc://$CACHE:50051
+    --remote_executore=grpc://$SCHEDULER:50052
+    //:dummy_test
+```
+
+> [!TIP]
+> You can add these flags to a to a `.bazelrc.user` file in the workspace root.
+> Note that you'll need to pass in explicit IPs as this file can't resolve
+> environment variables:
+> ```bash
+> # .bazelrc.user
+> build --config=lre
+> build --remote_instance_name=main
+> build --remote_cache=grpc://172.20.255.200:50051
+> build --remote_executor=grpc://172.20.255.201:50052
+> ```
+
+When you're done testing, delete the cluster:
+
+```bash
+kind delete cluster
+```