#IA Track - Technology and Tools
Google Cloud Platform allows to build, to train and to serve ML models in serverless thanks to a dedicated service: AI-Platform. This service can do all, but, on the orediction serving part, there is others Google Cloud Platform services which can offer different characteristic and could be more interesting for some use cases
- Big Query for batch prediction and based on Structured data
- Cloud Run for online prediction, which brings Knative compatibility and thus portability through different operator and event on premise!
During this talk, use cases, demos and implementation examples will be described based on Tensorflow already trained model. A comparaisons will be performed in term of deployment process, serving performances, team organisation and skills required, pricing and trade-off between portability and efficiency.
5. What is Serverless ?
Operational
Model
Programming
Model
No Server
Management
Fully Managed
Security
Pay only for
usage
Service-based Event-driven Portable
7. What is AI-Platform on GCP ?
Explore
data
Build
model
Train
model
Serve
model
Prepare
data
AI Platform
AI Platform
Cloud
Datalab
Cloud
Dataproc
Cloud
Dataprep
Cloud
Dataflow
BigQuery
Cloud Data
Fusion Datastudio AI platform
10. Portable solution
Any language
Any library
Any binary
Ecosystem of base images
Industry standard
.js .rb .go
.py .sh …
0 1 0
1 0 0
1 1 1
Containers
17. ● Common API and runtime environment for
serving workloads
● Implements learnings from Google and over 50
companies contributing
● Portability of experiences, tooling, and
workloads between Knative environments - you
can even run serverless on-prem
https://knative.dev
Portability based on Knative
18. Cloud Run
Fully managed, deploy your
workloads and don’t see the
cluster.
Cloud Run on Anthos
Deploy into Anthos, run
serverless side-by-side with
your existing workloads.
Knative Everywhere
Use the same APIs and
tooling anywhere you run
Kubernetes with Knative.
One experience, where you want it
19. - 1 vCPU, 2Gb of memory
- Scale to 0
- Pay-as-you-use
- Portable serverless container
Cloud Run summary
21. What is batch prediction process ?
Extract data
to files
Run batch
model
Store results
Load results in database
$ $ $ $
BigQuery Storage AI Platform Storage
23. SELECT data as input FROM `dataset.my_data`
SELECT * FROM ML.PREDICT(
MODEL `model.my_model`, (
Use model in a query
INSERT INTO result.my_results
)
)
24. BigQuery Storage AI Platform Storage
Batch prediction with BQML
$ $ $ $
Extract data
to files
Run batch
model
Store results
Load results in database
27. Batch prediction
Structured data
Compute intensive - GPU
AI-Platform
Data reachable with BQ
(un)structured data
No data engineering
BigQuery
Not only Tensorflow model
Focus on the last part, for serving a trainded model
Prediction on data
Inference on image/video/NLP
Make and use by Google. V2.0 in September
Tensorflow Lite and JS to run on mobile and browser
Focus on the last part, for serving a trainded model
Prediction on data
Inference on image/video/NLP
Small : cheaper and scale to 0. Quick deployment (about 60s)
Beta: powerfull, customizable, expensive: bigger and don’t scale to 0
Sticky to GCP + load only your model, no binary allowed