Roadmap changes (kubeflow#554)

* Updated roadmap for 0.3 and 0.4 * more changes * blah * merged * typo
magdalenakuhn17 · Nov 20, 2019 · 42a6abd · 42a6abd
1 parent 9d89afb
commit 42a6abd
Showing 1 changed file with 58 additions and 43 deletions.
diff --git a/ROADMAP.md b/ROADMAP.md
@@ -1,56 +1,47 @@
-# KF Serving 2019 Roadmap
-## Q3 2019
-### v0.2 Integrate with the ML Ecosystem (ETA: August 15, 2019)
-Objective: "Continue to simplify the user experience by deeply integrating with the Kubeflow Ecosystem."
-* Kubeflow Integration
-    * Prepare KFServing to release v0.2 and v0.3 alongside Kubeflow v0.7.
-    * Integrate with `kfctl generate` and `kfctl apply`.
-    * Deploy as a [Kubernetes Application](https://github.com/kubernetes-sigs/application).
-    * Integrate with Kubeflow Pipelines to enable model deployment from a Pipeline.
-    * Integrate with Fairing to enable model deployment from a Notebook.
-    * Achieve 20% End-to-End Test Coverage of Supported Features. (See v0.3 for 80%).
-    * Support PVCs to enable integration with on-prem Kubeflow installations.
-    * Document Installation for various cloud providers (GCP, IBM Cloud, Azure, AWS).
+# KF Serving Roadmap
+## 2019 - 2020
 
-Objective: "Empower users to deeply understand their predictions and validate KFServing's static graph architecture."
-* Explainability
-    * Deploy a predictor and explainer, powered by Alibi.
-    * Deploy a predictor and explainer, powered by user specified explainer container.
+### v0.4 Performance (ETA: Jan 31, 2019)
+Objective: "Prevent performance regressions across a known set of representative models."
+* Automated Performance Tests 
+    * Define a set of Models to test covering a wide array of usecases and frameworks.
+    * Publish performance results over time to enable regression tracking.
 
-Objective: "Increase coverage of ML frameworks to support previously unsupported customer workloads."
-* Frameworks
-    * Deploy a ONNX model
-    * Explore supporting other model serialization mechanisms for certain frameworks (e.g. saving PyTorch models with dill)
+Objective: "Enable users to deploy latency sensitive models with KFServing."
+* High Performance Dataplane
+    * Enable support for GRPC or similar.
+    * Continue to support existing HTTP Dataplane.
+
+Objective: "Reduce Total Cost of Ownership when deploying multiple underutilized models."
+* GPU Sharing 
+    * Reduce TCO by enabling models of the same framework and version to be co-hosted in a single model server.
+
+### v0.3 Stability (ETA: Dec 15, 2019)
+Objective: "Improve practices around dependency management." 
+* Migrate to Kubebuilder 2.0.
+    * Use Go Modules.
+    * Stop Vendoring dependencies.
+    * Avoid the extremely heavy dependency on Tensorflow.
+* Migrate to Kubernetes 1.15.
+    * Enable LabelSelectors for the Pod Mutation Webhook.
 
-### v0.3 Performance and Stability (ETA: September 1, 2019)
-Objective: "Prevent feature regressions with 80% end-to-end test coverage against a live Cluster."
+Objective: "Prevent feature regressions with greater end-to-end test coverage against a live cluster."
 * Automated End-to-End Tests
     * Execute against a Kubeflow maintained GKE Cluster.
     * Execute against a Kubeflow maintained AKS Cluster.
     * Achieve >80% Test Coverage of Supported Features.
 
-Objective: "Prevent performance regressions across a known set of representative models."
-* Automated Performance Tests 
-    * Define a set of Models to test covering a wide array of usecases and frameworks.
-    * Publish performance results in a temporally comparable way.
-
-Objective: "Improve the Serverless Experience by reducing cold starts/stops to 10 seconds on warmed models."
-* Model Caching
-    * Reduce model download time by caching models from cloud storage on Persistent Volumes.
-* Image Caching
-    * Reduce container download time by ensuring images are cached in all cloud environments.
-* Server Shutdown
-    * Ensure that all model servers shutdown within 10 seconds of not receiving traffic.
-
-Objective: "Simplify user experience with handling credentials for storage backends."
-* Secure storage mechanisms
-    * Implement a simplified user experience with storage backends protected by credentials (e.g. S3/GCS accounts with credentials)
+Objective: "Improve build and release processes to improve the developer experience and avoid regressions."
+* Improve build reliability
+    * Implement build retries.
+    * Reduce PyTorch build time.
+* Automated Image Injection for Model Servers.
+    * Implement new developer commands to deploy kfserving with local images.
+* Improve versioning of XGBoost, SKLearn, and PyTorch
+    * Replace KFServing version with the corresponding framework version.
 
 # Future 
 ## Unscheduled Work
-* Multi-Model Serving.
-    * Multiple InferenceServices share resources in the backend.
-    * GPU Sharing.
 * Flexible Inference Graphs [MLGraph CRD](https://github.com/SeldonIO/mlgraph).
     * Model Experimentation.
     * Ensembling.
@@ -66,8 +57,32 @@ Objective: "Simplify user experience with handling credentials for storage backe
     * Queue and batch requests to increase throughput.
 
 # Historical
+### v0.2 Integrate with the ML Ecosystem (Oct 31, 2019)
+Objective: "Continue to simplify the user experience by deeply integrating with the Kubeflow Ecosystem."
+* Kubeflow Integration
+    * Prepare KFServing to release v0.2 and v0.3 alongside Kubeflow v0.7.
+    * Integrate with `kfctl generate` and `kfctl apply`.
+    * Deploy as a [Kubernetes Application](https://github.com/kubernetes-sigs/application).
+    * Integrate with Kubeflow Pipelines to enable model deployment from a Pipeline.
+    * Integrate with Fairing to enable model deployment from a Notebook.
+    * Achieve 20% End-to-End Test Coverage of Supported Features. (See v0.3 for 80%).
+    * Support PVCs to enable integration with on-prem Kubeflow installations.
+    * Document Installation for various cloud providers (GCP, IBM Cloud, Azure, AWS).
+
+Objective: "Empower users to deeply understand their predictions and validate KFServing's static graph architecture."
+* Explainability
+    * Deploy a predictor and explainer, powered by Alibi.
+    * Deploy a predictor and explainer, powered by user specified explainer container.
+
+Objective: "Increase coverage of ML frameworks to support previously unsupported customer workloads."
+* Frameworks
+    * Deploy a ONNX model
+    * Explore supporting other model serialization mechanisms for certain frameworks (e.g. saving PyTorch models with dill)
+
+
+
 ## Q2 2019
-### v0.1: InferenceService Minimum Viable Product (ETA: June 30, 2019)
+### v0.1: InferenceService Minimum Viable Product (June 30, 2019)
 Objective: "Simplify the user experience and provide a low barrier to entry by minimizing the amount of YAML necessary to deploy a trained model."
 * High Level Interfaces
     * Deploy a Tensorflow model without specifying a Tensorflow Serving Technology.