Machine Learning/LiftWing

Lift Wing

A scalable machine learning model serving infrastructure on Kubernetes using KServe.

• Phabricator MVP Task: https://phabricator.wikimedia.org/T272917

Stack

Software	Version
k8s	v1.16.5
Istio	v1.9.5
Knative	v0.18.1
KServe	v0.7.0

Istio

Istio is a service-mesh where we can run our ML-services. It is installed using the istioctl package, which has been added to the WMF APT repository (https://wikitech.wikimedia.org/wiki/APT_repository, debian buster). See: https://apt-browser.toolforge.org/buster-wikimedia/main/ , we are currently running Istio 1.9.5 (istioctl: 1.9.5-1)

Knative

We use Knative Serving for running serverless containers on k8s using Istio. It also allows for various deployment strategies like canary, blue-green, A/B tests, etc.

Charts

• Knative Serving CRDs: https://gerrit.wikimedia.org/r/plugins/gitiles/operations/deployment-charts/+/refs/heads/master/charts/knative-serving-crds/
• Knative Serving: https://gerrit.wikimedia.org/r/plugins/gitiles/operations/deployment-charts/+/refs/heads/master/charts/knative-serving/

Images

• Webhook: https://docker-registry.wikimedia.org/knative-serving-webhook/tags/
• Queue: https://docker-registry.wikimedia.org/knative-serving-queue/tags/
• Controller: https://docker-registry.wikimedia.org/knative-serving-controller/tags/
• Autoscaler: https://docker-registry.wikimedia.org/knative-serving-autoscaler/tags/
• Activator: https://docker-registry.wikimedia.org/knative-serving-activator/tags/
• Net-istio webhook: https://docker-registry.wikimedia.org/knative-net-istio-webhook/tags/
• Net-istio controller: https://docker-registry.wikimedia.org/knative-net-istio-controller/tags/

KServe

We use KServe for it's custom InferenceService resource. It enables us to expose our ML models as asynchronous micro-services.

Charts

• KServe: https://gerrit.wikimedia.org/r/plugins/gitiles/operations/deployment-charts/+/refs/heads/master/charts/kserve/
• InferenceService: https://gerrit.wikimedia.org/r/plugins/gitiles/operations/deployment-charts/+/refs/heads/master/charts/kserve-inference/

Images

• KServe agent: https://docker-registry.wikimedia.org/kserve-agent/tags/
• Kserve controller: https://docker-registry.wikimedia.org/kserve-controller/tags/
• KServe storage-initializer: https://docker-registry.wikimedia.org/kserve-storage-initializer/tags/

Hosts

eqiad

• ml-serve1001-4

codfw

• ml-serve2001-4

Components

Monitoring

• KServe: https://grafana.wikimedia.org/d/Rvs1p4K7k/kserve
• Knative Serving: https://grafana.wikimedia.org/d/c6GYmqdnz/knative-serving

Serving

We host our Machine Learning models as Inference Services (isvcs), which are asynchronous micro-services that can transform raw feature data and make predictions. Each inference service has production images that are published in the WMF Docker Registry via the Deployment Pipeline. These images are then used for an isvc configuration in our ml-services helmfile in the operations/deployment-charts repo.

• Model Deployment Guide: Machine Learning/LiftWing/Deploy
• Inference Service Docs: Machine_Learning/LiftWing/Inference Services

Storage

We store model binary files in Swift, which is an open-source s3-compatible object store that is widely-used across the WMF. The model files are downloaded by the storage-initializer (init:container) when an Inference Service pod is created. The storage-initializer then mounts the model binary in the pod at /mnt/models/ and can be loaded by the predictor container.

• Model Upload info: Machine_Learning/LiftWing/Deploy#How_to_upload_a_model_to_Swift

Development

We are developing inference services on the ML Sandbox using our own WMF KServe images & charts.

We previously used multiple sandbox clusters running MiniKF.

Services

We are serving ML models as Inference Services, which are containerized applications. The code is currently hosted on gerrit: https://gerrit.wikimedia.org/g/machinelearning/liftwing/inference-services