On this podcast, Keith Trnka, Machine Learning Engineer / Engineering Leader, talked about applying engineering to the data science thought process.

Communication is key to engineering. Writing code effectively is critical. Factors like the way variables and functions are named, the way code is structured e.t.c communicates ideas and purpose between the developers.

In data science, the majority of code writing happens in the inherently messy exploration phase. This makes it difficult to picture these engineering communication assumptions with data science. It is because data science/machine learning is a complex field with a lot of terminologies and a math background that requires people to learn and adapt to that.

Two things are going on when it comes to data science. On one hand, there are unique challenges with explaining the quality of a piece of code. The other challenge is dealing with the exploration mode itself

ML Failure Modes
Machine Learning comes with a number of common failure modes. Trying to do something that doesn't make sense or isn't needed by the user/business can cause machine learning projects to fail. Another failure mode on the flip is every other piece of engineering in the product.

At the last New York Meetup, Kyle Gallatin, Senior Software Engineer at Etsy, shared the technical bits on how Etsy's ML Platform team worked to improve support for Deep Learning at scale.

Etsy's ML Platform
Etsy's Machine Learning Platform team is composed of two squads. The ML Training squad focuses on infrastructure for training and prototyping. The Model Serving squad focuses on infrastructure for serving at scale and all other services. With over 70 models in production managing, scaling and troubleshooting real-time workloads are more challenging.

Scaling challenges include;
- Custom feature transforms from frameworks causes huge latency spikes
- Deep Learning models have different ideal infrastructure workload settings
- For experiments, latency and cost were treated before deploying to production, which means surprise in latency and cost leads to delay in experiments.

Etsy's Solutions
Etsy built a package solution for the model called Caliper. Caliper enables the engineers to get early latency and cost feedback about a model, before putting it in production.

Distributed tracing was also added, for more system wide observability. This enables them to get spans and see how long different portions of their system are contributing to latency.

An ML problem is defined by inputs, outputs, and the objective function that guides the learning process. Most general types of ML tasks are either classification or regression. However, a regression model can easily be framed as a classification model and vice versa. Within classification, there are more subtypes of binary, multiclass, and multilabel.

Within classification problems, the fewer classes there are to classify, the simpler the problem. With only two possible classes in a binary classification problem, it is typically the least challenging type of classification problem. Multiclass classification involves problems with more than two classes.

Multilabel Classification
For a multilabel classification problem, an example can belong to multiple classes. There are two major approaches to multilabel classification problems. The first is to treat it as you would a multiclass classification. The second approach is to turn it into a set of binary classification problems.

Handling multilabel classification tasks are usually the most challenging for companies. The number of classes an example can have varies from example to example, this makes it difficult for label annotation since it increases the label multiplicity problem. Also, the varying number of classes makes it hard to extract predictions from raw probability.

Jason Dunne, Senior Product Marketing Manager at Tecton, will be talking about the new capability of Tecton's feature platform and how it will enable data engineers, ML engineers, and data scientists to improve their feature engineering workflows.

In Tecton 0.6, data teams can now develop and test features quickly with the flexibility of a Python notebook using the new notebook-driven development capability.