Many machine learning (ML) practitioners use partial dependence plots (PDP) to gain insights into model behaviors. But have you run into situations where PDPs average two groups with different behaviors and produce curves applicable to none? Are you longing for tools that help you understand detailed model behavior in a visually manageable way? Look no further! We are thrilled to share with you our newest model interpretation tools: the Nearby Individual Conditional Expectation plot and its companion, the Partial Dependence at Quantiles plot. They highlight local behaviors and hint at how much we may trust such readings. A not NICE...

Continue reading

This post is about how Yelp is transitioning from the management of Cassandra clusters in EC2 to orchestrating the same clusters in production on Kubernetes. We will start by discussing the EC2-based deployment we have used for the past few years, followed by an introduction to the Cassandra operator, its responsibilities, the core reconciliation workflow of the operator, and finally, the etcd locking we employ for cross-region coordination. Cassandra is a distributed wide-column NoSQL datastore and is used at Yelp for both primary and derived data. Yelp’s infrastructure for Cassandra has been deployed on AWS EC2 and ASG (Autoscaling Group)...

Continue reading

Engineers on Yelp’s Consumer Growth team work closely with product managers, data scientists, and designers to increase user acquisition, engagement and retention to fuel the rest of the business. The team is central to growing the number of active users on the Yelp platform. In-App Update, introduced in Android Pie, is a feature that allows Android users to update the app by showing a prompt while they’re using the app and keeps users in the app as the update is happening in the background. We thought it would be a valuable feature for consumers on Yelp and began researching and...

Continue reading

The relational database of choice at Yelp is MySQL and it powers much of the Yelp app and yelp.com. MySQL does not include a native high-availability solution for the replacement of a primary server, which is a single point of failure. This is a tradeoff of its dedication to ensuring consistency. Replacing a primary server is sometimes necessary due to planned or unplanned events, like an operating system upgrade, a database crash or hardware failure. This requires pausing data modifications to the database while the server is restarted or replaced and can mean minutes of downtime. Pausing data modifications means...

Continue reading

Jupyter notebooks are a key tool that powers Yelp data. It allows us to do ad hoc development interactively and analyze data with visualization support. As a result, we rely on Jupyter to build models, create features, run Spark jobs for big data analysis, etc. Since notebooks play a crucial role in our business processes, it is really important for us to ensure the notebook output is reproducible. In this blog post, we’ll introduce our notebook archive and sharing service called Folium and its key integrations with our Jupyterhub that enable notebook reproducibility and improve ML engineering developer velocity. Folium...

Continue reading

At Yelp we process terabytes of streaming data a day using Apache Flink to power a wide range of applications: ETL pipelines, push notifications, bot filtering, sessionization and more. We run hundreds and hundreds of Flink jobs, so routine operations like deployments, restarts, and savepoints don’t take thousands of hours of developers’ time, which would be the case without the right degree of automation. The latest addition to our toolshed is a new stream processing platform built on top of PaaSTA, Yelp’s Platform As A Service. Sitting at its core, a Kubernetes operator automatically watches over the deployment and the...

Continue reading

At Yelp, new web pages and app screens are powered by GraphQL for fetching data. This blog post describes the Dream Query – a pattern our feature teams use when refactoring or creating new pages. (Check out our previous blog post to see how we dynamically codegen DataLoaders to implement the server layer!) Scoping a new feature with GraphQL Let’s jump in with an example! Imagine your team is tasked with creating the new version of the “Header component” for the website (we’ll use the Yelp.com website in our example). You may receive a design mock that looks like this:...

Continue reading

At Yelp, usage of Prometheus, the open-source monitoring system and time series database, is blossoming. Yelp is initially focusing on onboarding infrastructure services to be monitored via Prometheus, one such service being Apache Kafka. This blogpost discusses some of the performance issues we initially encountered while monitoring Kafka with Prometheus, and how we solved them by contributing back to the Prometheus community. Kafka at Yelp primer Kafka is an integral part of Yelp’s infrastructure, clusters are varied in size and often contain several thousand topics. By default, Kafka exposes a lot of metrics that can be collected, most of which...

Continue reading