MLOps World 2022
Vendor Overview
There were two large groups of vendors.
- Vendors selling ML pipelines or integrated ML Ops platforms
- Vendor selling model monitoring systems aka observability.
Cloud Representation
This conference was heavily AWS based. AWS was the most represented of the big three. They had their own booth. I saw no Azure representation ( but some Microsoft Research folk).
I also see very little Google representation, one talk. IBM was also present. Most vendors had AWS solutions, or were hosted on AWS. Azure and GCP compatibility and offerings were scarce.
Sessions
- June 7th 10am
First session Best practice
Lina
not very usful to me. joined late doue to hova + hopin tehcnical difficulties.
Automated Machine Learning & Tuning with FLAML
11 am
3h session.
Manually tuning hyper parameters is a large effort.
Costs a lot, tuning opportunity lost due to load.
Also cost is not one time.
The goal is to remove the hassle of tuning.
https://microsoft.github.io/FLAML/
Taking MLOps 0-60: How to Version Control, Unify Data and Management
https://pachyderm.io/ also: https://www.pachyderm.com/
github.com/pachyderm/examples/tree/master/label-studio
MLOps at Scale
Shelbee aws person
Cross org size Cross Industy
People process technology
Govern workloads
Lesson 1 : start with process.
What matter are the problem that grt in the way of deployment.
The art of the possible sessions
Outline performance expectation
"If you need 100% accuracy ML is not for you"
Q: executive sponsorship
Ml strategy must have data strategy
Need day to day access to data
Must have a data catalog
When you talk about pii and thousands of staff accessing it. You talk about liability to the model and such. But there is liability to the business.
Cost control. Iac , as code ,
Support sesff to choose correct correct jsnrance sizes.
Control , communication about budget and spending.
Have automated process for deployment.
Missing: approvals.
Isn't it "just a another software deploy"
Retraining plan in plan at the beginning
Model monitoring
Business performance
Technical performance
Lesson 2 you can scale with people issues.
You didn't mention security people.
Lots of data scientist but not data engineers and ml engineers.
"Jobs that need to be done" how about "roles and responsibilities".
Full stack unicorn : is a liability
Sick days Technical know how Business knowledge Quality of work Keeping current Mental health
Have common goals do depth are not at odd with each other.
Feature store , model registry .
Lesson 3 know you tradeoffs
Core to Business?
Effort
Lesson 4 can't just use software dev process for ml.
Experime Tatiana is key . Most dev cycles don't have space for that.
Not every experiment should go to a pipeline
But like pipelines can have sraged releases, some fast than others and you can do local test and have pipelines.
With long train times you can train with every pipeline run.
You need model lineage. Where did the model came ?
Lesson 5 start small and iterate
Think maturity model
‐---------- Arize
Fraud not fraud
Failed not failed
Feature contribution
Analysis models.
What is the contribution. Explainability
Ground truths , fast actuals Drift is proxy to performance.
Data drift, co variant drift, related to
Feature drift.
Time based drift, environment changed.
Anomaly detect, versus drift... when a cohort has changed. Between two whole distributions.
Psi pop stable index.
Use drift investigation to track back to the anomaly.
Google paper 'dataflow"
Cat boost : titanic survival model
---
Feature engineering
Look for missing values Measure of central tendency
Measures if percentile , why not candle sticks
Biasness detection
Pandas report
Iv information value association value
Ig information gain association value
Data stability
Typically ad hic... most modelers have no domain knowledge.
The weak Ness of auto mstic Feature Gen is that you don't know where it came from. You don't know I'd it's just you data set. You don't know about drift.
Feature recoomnedn sankey visualization
Connecting attributes to features.
Checks
Feature stability v attribute stability
Encoding . Woe weight of evidence
Attributes are what go into the eda , the raw data,. Features are what you decide to treat as inputs to models. Either straight features or synthetic, generated.
Optio s:
Make added with features that don't drift. The model is more reliable and can be used from longer before Retraining.
Or you find a very predictive feature that has drift and understand you will need to make a new model soon.
Link Dump
https://github.com/microsoft/DeepSpeed
