Lykeion: Machine Learning as a Service

Life of a Data Scientist is not always easy. They can create fancy machine learning models to solve really complex problems. However, usually, that’s not the hardest part of the job: trouble comes when it’s time to see that model making decisions in a production environment.

Transcript

1. Lykeion Machine Learning as a Service Jose Honrado - Data Engineer @ Cabify
2. A couple of definitions about ML Feature It is an individual measurable property of a phenomenon being observed. Examples: user total drop-offs (rides), user first/last drop-off date, user language Model It is a mathematical model built using a ML algorithm based on sample data (training), in order to make predictions or decisions without being explicitly programmed to perform that task. Example: model that predicts the probability of a user ordering a Cabify again
3. The problem
4. The ideal scenario
5. Lykeion objectives • Autonomy. Data scientists should productionize ML models by themselves. • Scalability. Allow deploying models and computing features at scale. • Traceability & Debuggability. Ensure that model trainings are reproducible. • Correctness & Precision. Fix the gap between training and predicting data (features).
6. Cabify’s machine learning platform
7. Lykeion architecture overview Features Cloud Bigtable BigQuery GraphQL API Kubernetes Engine Elastic API Kubernetes Engine Models Model Repository API Kubernetes Engine Feature Manifests Model Storage Model Management API Kubernetes Engine Feature generation Cloud Dataflow Train Eval Model Serving API Kubernetes Engine Model 1 Model N Google Product Lykeion service
8. Features Architecture Contracts Computation Serving
9. Lykeion overview - Features Features Cloud Bigtable BigQuery GraphQL API Kubernetes Engine Elastic API Kubernetes Engine Feature Manifests Feature generation Cloud Dataflow
10. Feature manifests
11. Feature generation process
12. Feature generation job
13. Feature serving API
14. Models Architecture Contracts Workflow Serving
15. Lykeion Overview - Models Models Model Repository API Kubernetes Engine Model Storage Model Management API Kubernetes Engine Train Eval Model Serving API Kubernetes Engine Model 1 Model N
16. Lykeion model concepts • Family: rider_p_alive • Segment: es_madrid, es_sevilla, br_sao_paulo... • Version: 1.4.3
17. Contract for models Open source framework to manage the ML lifecycle. Including: • MLflow projects • MLflow models • MLflow tracking
18. Model productioning workflow
19. Model serving API
20. Lykeion overview Lykeion Features Cloud Bigtable BigQuery GraphQL API Kubernetes Engine Elastic API Kubernetes Engine Models Model Repository API Kubernetes Engine Feature Manifests Model Storage CLI tool Feature generation Cloud Dataflow Train Eval Model Management API Kubernetes Engine Model Serving API Kubernetes Engine Model 1 Model N
21. Model productioning workflow (lykectl) 1. Create model family: lykectl families create --alias forward --name Forward \ --description "Forward test model" --authors "aristotle@cabify.com" 2. Create model: lykectl models create forward --segment default --version 1.0 \ --description "Version description" \ --source-url "git@git-provider.com:projects.git#mlprojects/forward@<SHA1>" \ --input-features "rider.age,rider.drop_offs_count" 3. Train/deploy model: lykectl models train|deploy forward --segment default --version 1.0
22. Conclusions • Good acceptance by Data Science team. ◦ They were part of the development as stakeholders. • Way faster time to market of ML models. • Easy access to model predictions via API. • Several teams consuming features via API.
23. Thanks! That’s it