Introducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
4 likes2,627 views
The document discusses the challenges of deploying machine learning models in production and introduces Kubeflow as a solution, highlighting its design and core components. It features contributions from various experts, providing insights on model training, serving, and pipelines while noting the integration of tools like TensorFlow and Apache Spark. The document also addresses potential downsides of using Kubeflow, such as overhead and the active development nature of the platform.
![Trevor Grant
● From Chicago
● Preferred pronouns he/him
● Various odd jobs around IBM (data scientist/evangelist/janitor)
● PMC Apache Mahout, Streams, Community Development
● Apache Roadshow Chicago [1]
● IoT Track at Apache Con North America [2]
● Blog: rawkintrevo.org
● Twitter: @rawkintrevo
● Github: rawkintrevo
● Shameless self promotion: rawkintrevo.org/shameless-self-promotion/
[1] https://www.apachecon.com/chiroadshow19/
[2] https://www.apachecon.com/acna19/index.html](https://image.slidesharecdn.com/weds200pmroom124-125introducingkubefloww-190401213917/85/Introducing-Kubeflow-w-Special-Guests-Tensorflow-and-Apache-Spark-4-320.jpg)
![But what about [special foo-baz-inator] or
[special-yak-shaving-tool]?
Write a Dockerfile and build an image, use FROM so you’re not starting from
scratch.
FROM gcr.io/kubeflow-images-public/tensorflow-1.6.0-notebook-cpu
RUN pip install py-special-yak-shaving-tool
Then tell set it as a param for your training/serving job as needed:
ks param set tfjob-v1alpha2 image "my-special-image-goes-here”
Now your fortran lives forever!](https://image.slidesharecdn.com/weds200pmroom124-125introducingkubefloww-190401213917/85/Introducing-Kubeflow-w-Special-Guests-Tensorflow-and-Apache-Spark-47-320.jpg)
Introducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
- 1. Machine Learning: From Lab to Production with Kubeflow With Special Guests Tensorflow and an Apache Spark teaser
- 2. Agenda ● About Us ● Background ● Problem: Building Model is Easy, Serving models in Prod is hard. ● What Is KubeFlow? ● Kubeflow’s Design and Core Components ● …
- 3. Holden About Me Slides ● Prefered pronouns are she/her ● Developer Advocate at Google ● Apache Spark PMC / ASF member + contributor on lots of other projects ● previously IBM, Alpine, Databricks, Google, Foursquare & Amazon ● co-author of Learning Spark & High Performance Spark ● Twitter: @holdenkarau ● Slide share http://www.slideshare.net/hkarau ● Code review livestreams: https://www.twitch.tv/holdenkarau / https://www.youtube.com/user/holdenkarau ● Talk Videos http://bit.ly/holdenSparkVideos ● Talk feedback: http://bit.ly/holdenTalkFeedback ● Organizing data track @ IT Next AMS - CFP Open!
- 4. Trevor Grant ● From Chicago ● Preferred pronouns he/him ● Various odd jobs around IBM (data scientist/evangelist/janitor) ● PMC Apache Mahout, Streams, Community Development ● Apache Roadshow Chicago [1] ● IoT Track at Apache Con North America [2] ● Blog: rawkintrevo.org ● Twitter: @rawkintrevo ● Github: rawkintrevo ● Shameless self promotion: rawkintrevo.org/shameless-self-promotion/ [1] https://www.apachecon.com/chiroadshow19/ [2] https://www.apachecon.com/acna19/index.html
- 6. Kubeflow Salesman: **Slaps roof of Kubeflow** THIS BAD BOY CAN FIT SO MANY BUZZWORDS IN IT
- 7. Kubeflow Salesman: **Slaps roof of Kubeflow** THIS BAD BOY CAN FIT SO MANY BUZZWORDS IN IT
- 8. Background Things we thought you might know.
- 9. History of Predictive Analytics Photo: Numerology Sign Photo: Akash Kataruka Photo: Hans Splinter
- 11. What is Statistics? Machine Learning?
- 12. What is Statistics? Machine Learning? A.I. (Artificial Intelligence)
- 13. What is Statistics? Machine Learning? A.I. (Artificial Intelligence) Photo: Andreas Kretschmer Model Training Photo: Helen Harrop Model Serving
- 14. What is Statistics? Machine Learning? A.I. (Artificial Intelligence)
- 15. What is Statistics? Machine Learning? A.I. (Artificial Intelligence)
- 16. From don’t know shit to know your shit* Verses # of GPUs required *Holden’s gut feelings after coffee NeedGPUs Knowledge of Shit
- 18. Kubeflow: Dev on Laptop Deploy in Cloud
- 20. So what is Kubeflow?
- 23. What is Kubeflow? VIK hotels group
- 25. The (many) kinds of models you can train ● All your favourite Python libraries* (in Jupyter) ○ Different options to parallelize, with more coming (for now MPI or Beam ish) ● PyTorch ● Tensorflow (along with hyper param tuning with katib) ● mxnet ● etc. Add-ons: ● H2o: CI is failing but you know, it's Wednesday ● And more!
- 26. But don't forget about data prep friends! ● For now options are a little limited, but other tools like Apache Spark are in the works ○ https://github.com/kubeflow/kubeflow/pull/1467 ● Now with Spark!* ● Pachyderm ● pandas? ● shell scripts? ● Tensorflow Transform in local mode Python 2 only…... ● Yeah I guess we did forget about our dataprep friends *Where now == what's in master as of Feb 14th 2019
- 27. Model persistence/deployment/CI ● You need to save your model somewhere ● Your favourite cloud storage provider goes here ○ Why invest in model management when I can make directory? ● ModelDB ● WeaveFlux ● Pachyderm
- 28. Model Serving Python Flask TensorFlow Model Server Openvino NVIDIA Inference Server Seldon Core ● Routers for A/B tests and multi-armed bandits. ● Supports lots of libraries (Python/Spark/R/etc) ● Monitoring / Security If you can put it in a Docker container, you can use it.
- 29. So you want to use this?
- 30. What’s Next?! Step away from keyboard Think about type(s) of model Look at components directory and see what’s a fit tool wise Don’t know? Choose jupyter deal with the details live Can’t find it? ^^ New Cat Content!!! ^._.^
- 31. What about just tensorflow?* ks registry add kubeflow github.com/kubeflow/kubeflow/tree/${VERSION}/kubeflow ks pkg install kubeflow/core@${VERSION} ks pkg install kubeflow/tf-serving@${VERSION} ks pkg install kubeflow/tf-job@${VERSION}
- 32. Getting the chef's recommend pairing: kfctl.sh init my_awesome_project --platform {none, gcp, minikube} cd my_awesome_project kfctl.sh generate platform && kfctl.sh apply platform kfctl.sh generate k8s && kfctl.sh apply k8s # Add spark cd ks_ap && ks pkg install kubeflow/spark Douglas O'Brien
- 33. Connect to the Kubeflow Web UI kubectl port-forward svc/ambassador -n kubeflow 8080:80 & # Or use IAP, but that's… another story
- 34. The chef's recommend pairing is: ● Jupyter Hub ● TF Job & TF Serving ● PyTorch ● Katib (Hyper parameter tuning) ● Ambassador (makes it easier to access the UIs) ● Pipelines (Argo + Magic) chicoblue
- 35. Click-to-deploy: get started hella fast* on GCP https://deploy.kubeflow.cloud
- 40. ro Hasegawa
- 41. What are those pipelines? “Kubeflow Pipelines is a platform for building and deploying portable, scalable machine learning (ML) workflows based on Docker containers.” - kubeflow.org Directed Acyclic Graph (DAG) of “pipeline components” (read “docker containers”) each performing a function.
- 44. Serving that job (not the only way)
- 45. When you don't know anything or know a lot about ML: Hyper Parameter Tuning ● Katib ○ Does not depend on a specific ML tool (e.g. not just TF) ○ Supports a few different search algorithms ○ e.g. "What should I set my L1 regularization too? Idk let's ask the computer" ● Great way to accidently overfit too!* (*if you're not careful) ● As respective cloud providers, we are happy to rent you a lot of resources ● Seriously, mention our names in the sales call. We're both going for promo (and that shit is hard)
- 46. Katib Screenshot
- 47. But what about [special foo-baz-inator] or [special-yak-shaving-tool]? Write a Dockerfile and build an image, use FROM so you’re not starting from scratch. FROM gcr.io/kubeflow-images-public/tensorflow-1.6.0-notebook-cpu RUN pip install py-special-yak-shaving-tool Then tell set it as a param for your training/serving job as needed: ks param set tfjob-v1alpha2 image "my-special-image-goes-here” Now your fortran lives forever!
- 48. Live streamed demos (recorded on YouTube) ● Kubeflow intro https://codelabs.developers.google.com/codelabs/kubeflow-introductio n/index.html & streamed http://bit.ly/kfIntroStream ● Kubeflow E2E with Github issue summurizationhttps://codelabs.developers.google.com/codelabs/cloud- kubeflow-e2e-gis/ & streamed http://bit.ly/kfGHStream ● You can tell they were live streamed by how poorly went, I promise no video editing has occurred.
- 49. Limited & Optional: Workshop Demo ● We are doing a workshop @ Strata SF and we'd love to trick offer the option to do a self-guided pre-ffer you an exciting opportunity to try a self-guided version free of charge and provide us feedback. We have some demo accts. ● What you will might learn: ○ Installing Kubeflow ○ Setting up a project ○ Deploying that project to GCP / Azure / IBM ○ Monkeying around with a project and still having it work ● Please come and talk to us after. Holden is wearing a shark dress. ● We'll be around to answer your questions fionasjournal
- 50. Why you shouldn't use this?
- 51. Downsides to Kubeflow ● Lot's of overhead versus doing it locally ● Active development (look it's 0.4) ● 3 talks on Kubeflow can give you 3 different toolsets