Empowering Pinterest Data Scientists and Machine Learning Engineers with PySpark

1,189
Pinterest
Pinterest is a social bookmarking site where users collect and share photos of their favorite events, interests and hobbies. One of the fastest growing social networks online, Pinterest is the third-largest such network behind only Facebook and Twitter.

Data scientists and machine learning engineers at Pinterest found themselves hitting major challenges with existing tools. Hive and Presto were readily accessible tools for large scale data transformations, but complex logic is difficult to write in SQL. Some engineers wrote complex logics in Cascading or Scala Spark jobs, but these have a steep learning curve and take significantly more time to learn and build jobs. Furthermore, data scientists and machine learning engineers often trained models in a small-scale notebook environment, but they lacked the tools to perform large-scale inference.

To combat these challenges, we, (machine learning and data processing platform engineers), built and productionized PySpark infrastructure. The PySpark infrastructure gives our users the following capabilities:

  • Writing logic using the familiar Python language and libraries, in isolated environments that allow experimenting with new packages.
  • Rapid prototyping from our JupyterHub deployment, enabling users to interactively try out feature transformations, model ideas, and data processing jobs.
  • Integration with our internal workflow system, so that users can easily productionize their PySpark applications as scheduled workflows.

PySpark on Kubernetes as a minimum viable product (MVP)

We first built an MVP PySpark infrastructure on Pinterest Kubernetes infrastructure with Spark Standalone Mode and tested with users for feedback.

Figure 1. An overview of the MVP architecture

The infrastructure consists of Kubernetes pods carrying out different tasks:

  • Spark Master managing cluster resources
  • Workers — where Spark executors are spawned
  • Jupyter servers assigned to each user

When users launch PySpark applications from those Jupyter servers, Spark drivers are created in the same pod as Jupyter and the requested executors in worker pods.

This architecture enabled our users to experience the power of PySpark for the first time. Data scientists were able to quickly grasp Python UDFs, transform features, and perform batch inference of TensorFlow models with terabytes of data.

This architecture, however, had some limitations:

  • Jupyter notebook and PySpark driver share resources since they are in the same pod.
  • Driver’s port and address are hard-coded in the config.
  • Users can launch only one PySpark application per assigned Jupyter server.
  • Python dependency per user/team is difficult.
  • Resource management is limited to FIFO approach across all the users (no queue defined).

As the demand for PySpark grew, we worked on a production-grade PySpark infrastructure based on Yarn, Livy, and Sparkmagic.

Production-grade PySpark infrastructure

Figure 2: An overview of the production architecture

In this architecture, each Spark application runs on the YARN cluster. We use Apache Livy to proxy between our internal JupyterHub, the Spark application and the YARN cluster. On Jupyter, Sparkmagic provides a PySpark kernel that forwards the PySpark code to a running Spark application. Conda provides isolated Python environments for each application.

With this architecture, we offer two development approaches.

Interactive development:

  1. A user creates a conda environment zip containing Python packages they need, if any.
  2. From JupyterHub, they create a notebook with PySpark kernel from Sparkmagic.
  3. In the notebook, they declare resources required, conda environment, and other configuration. Livy launches a Spark application on the YARN cluster.
  4. Sparkmagic ships the user’s Jupyter cells (via Livy) to the PySpark application. Livy proxies results back to the Jupyter notebook.

See the attached picture (see Appendix) for a full annotated example of a Jupyter notebook.

Non-interactive development (ad-hoc and production workflow runs):

  1. A Pinterest-internal Job Submission Service acts as the gateway to the YARN cluster.
  2. In development, the user’s local Python code base is packaged into an archive and submitted to launch a PySpark application in YARN.
  3. In scheduled production runs, the production build’s archive is submitted instead.

Benefits

This infrastructure offers us the following benefits:

  1. No resources sharing between Jupyter notebook and PySpark drivers
  2. No hard-coded drivers’ ports and addresses
  3. Users can launch many PySpark applications
  4. Efficient resource allocation and isolation with aggressive dynamic allocation for high resource utilization
  5. Python dependency per user is supported
  6. Resource accountable
  7. Dr. Elephant for PySpark Job analyses

Technical details

Pinterest JupyterHub Integration: (benefits #1,2,3)

We made the Sparkmagic kernel available in Jupyter. When the kernel is selected, a config managed by ZooKeeper is loaded with all necessary dependencies.

We set up Apache Livy, which provides a REST API proxy from Jupyter to the YARN cluster and PySpark applications.

A YARN cluster: (benefit #4)

  • Efficient resource allocation and isolation. We define a queue structure with Fair Scheduler to ensure dedicated resources and preemptable under certain conditions (e.g. after waiting for at least 10 minutes) but a portion of non-preemptable resources will be held for queues with minResource being set. Scheduler and resource manager logs are to manage cluster resources.
  • Aggressive Dynamic allocation policy for high resource utilization. We set the policy where a PySpark application holds at most a certain amount of executors and automatically releases resources once they don’t need. This policy makes sure resources are recycled faster, leading to a better resource utilization.

Python Dependency Management: (benefit #5)

Users can try various Python libraries (e.g. different ML frameworks) without asking platform engineers to install them. To that end, we created a Jenkins job to package a conda environment based on a requirement file, and archive it as a zip file on S3. PySpark applications launched with “ — archives” to broadcast zip file to driver along with all executors, and reset both “PYSPARKPYTHON” (for driver) as well as “spark.yarn.appMasterEnv.PYSPARKPYTHON” (for executors). That way, each application runs under in an isolated Python environment with all libraries needed.

Integrating with Pinterest-internal Job Submission Service (JSS): (benefit #6)

To productionize PySpark applications, users leverage the internal workflow system to schedule. We provided a workflow template to integrate with job submission interfaces to specify code location, parameters, and a Python environment artifact to use.

Self-service job performance analysis: (benefit #7)

We forked the open-sourced Dr. Elephant, and added new heuristics to analyze application’s configuration with various kinds of runtime metrics (executor, job, stage, …). This service provides tuning suggestions and offers guidelines on how to write a spark job properly. The service alleviates users’ debugging-and-troubleshooting pain, boosting the velocity. Moreover, it avoids resource waste and improves cluster stability. Below is an example of the performance analysis.

Figure 3: An overview of Dr. Elephant

Impacts

PySpark is now being used throughout our Product Analytics and Data Science, and Ads teams for a wide range of use cases.

  • Training: users can train models with mllib or any Python machine learning frameworks (e.g. TensorFlow) iteratively with any size of data.
  • Inference: users can test and productionize their Python codes for inferences without depending on platform engineers.
  • Ad-hoc analyses: users can perform various ad-hoc analyses as needed.

Moreover, our users now have the freedom to explore various Python dependencies and use Python UDF for large scale data.

Acknowledgement

We thank David Liu (EM, Machine Learning Platform team), Ang Zhang (EM, Data Processing Platform team), Tais (our TPM), Pinterest Product Analytics and Data Science organization (Sarthak Shah, Grace Huang, Minli Zhang, Dan Lee, Ladi Ositelu), Compute-Platform team (Harry Zhang, June Liu), Data Processing Platform team (Zaheen Aziz), Jupyter team (Prasun Ghosh — Tech Lead) for their support and the collaborations.

Appendix — An example of our use-case (Appendix):

Below is an example of how our users train a model, and run inference logic at scale from their Jupyter notebook with PySpark. We leave explanations in each cell.

Pinterest
Pinterest is a social bookmarking site where users collect and share photos of their favorite events, interests and hobbies. One of the fastest growing social networks online, Pinterest is the third-largest such network behind only Facebook and Twitter.
Tools mentioned in article
Open jobs at Pinterest
Senior Staff Machine Learning Enginee...
San Francisco, CA

About Pinterest:

Millions of people across the world come to Pinterest to find new ideas every day. It’s where they get inspiration, dream about new possibilities and plan for what matters most. Our mission is to help those people find their inspiration and create a life they love. As a Pinterest employee, you’ll be challenged to take on work that upholds this mission and pushes Pinterest forward. You’ll grow as a person and leader in your field, all the while helping users make their lives better in the positive corner of the internet.

Homefeed is a discovery platform at Pinterest that helps users find and explore their personal interests. We work with some of the largest datasets in the world, tailoring over billions of unique content to 330M+ users. Our content ranges across all categories like home decor, fashion, food, DIY, technology, travel, automotive, and much more. Our dataset is rich with textual and visual content and has nice graph properties — harnessing these signals at scale is a significant challenge. The Homefeed ranking team focuses on the machine learning model that predicts how likely a user will interact with a certain piece of content, as well as leveraging those individual prediction scores for holistic optimization to present users with a feed of diverse content.

What you’ll do:

  • Work on state-of-the-art large-scale applied machine learning projects
  • Improve relevance and the user experience on Homefeed
  • Re-architect our deep learning models to improve their capacity and enable more use cases
  • Collaborate with other teams to build/incorporate various signals to machine learning models
  • Collaborate with other teams to extend our machine learning based solutions to other use cases

What we’re looking for:

  • Passionate about applied machine learning and deep learning
  • 8+ years experience applying machine learning methods in settings like recommender systems, search, user modeling, image recognition, graph representation learning, natural language processing

#L1-EA2

EPM Lead Developer, Adaptive Planning...
San Francisco, CA

About Pinterest: 

Millions of people across the world come to Pinterest to find new ideas every day. It’s where they get inspiration, dream about new possibilities and plan for what matters most. Our mission is to help those people find their inspiration and create a life they love. As a Pinterest employee, you’ll be challenged to take on work that upholds this mission and pushes Pinterest forward. You’ll grow as a person and leader in your field, all the while helping users make their lives better in the positive corner of the internet.

The EPM technology team at Pinterest is looking for a senior EPM architect who has at least four years of technical experience in Workday Adaptive Planning. You will be the solutions architect who oversees technical design of the complete EPM ecosystem with emphasis on Adaptive Financial and Workforce planning. The right candidate will also need to have hands-on development experience with Adaptive Planning and related technologies. The role is in IT but will work very closely with FP&A and the greater Finance/Accounting teams. Experience with Tableau suite of tools is a plus.

What you'll do: 

  • Together with the EPM Technology team, you will own Adaptive Planning and all related services
  • Oversee architecture of existing Adaptive Planning solution and make suggestions for improvements
  • Solution and lead Adaptive Planning enhancement projects from beginning to end
  • Help EPM Technology team gain deeper understanding of Adaptive Planning and train the team on Adaptive Planning best practices
  • Establish strong relationship with Finance users and leadership to drive EPM roadmap for Adaptive Planning and related technologies
  • Help establish EPM Center of Excellence at Pinterest

What we're looking for: 

  • Hands-on design and build experience with all Adaptive Planning technologies: standard sheets, cube sheets, all dimensions, reporting, integration framework, security, dashboarding and OfficeConnect
  • Strong in application design, data integration and application project lifecycle
  • Comfortable working side-by-side with business
  • Ability to translate business requirements to technical requirements
  • Strong understanding in all three financial statements and the different enterprise planning cycles
  • Familiar with Tableau suite of tools

 

Software Engineer, Shopping Discovery
San Francisco, CA

About Pinterest:

Millions of people across the world come to Pinterest to find new ideas every day. It’s where they get inspiration, dream about new possibilities and plan for what matters most. Our mission is to help those people find their inspiration and create a life they love. As a Pinterest employee, you’ll be challenged to take on work that upholds this mission and pushes Pinterest forward. You’ll grow as a person and leader in your field, all the while helping users make their lives better in the positive corner of the internet.

Shopping is at the core of Pinterest’s mission to help people create a life they love. The shopping discovery team at Pinterest is inventing a brand new, more visual and personalized shopping experience for 350M+ users worldwide. The team is responsible for delivering mid-funnel shopping experience on shopping surfaces like Product Detail Page, Shopping Search, Shopping on Board etc. You'll be responsible for optimizing the whole page layout by appropriately selecting and slotting the UI templates and recommendation modules optimizing towards a shopping metric. As an engineer of the team you will be working on the most cutting edge recommendation algorithms to develop diverse types of shopping recommendations that will be displayed across different shopping surfaces on Pinterest.

What you'll do: 

  • Develop large scale shopping recommendation algorithms
  • Build data pipelines to do data analysis and collect training data
  • Train deep learning models to improve quality and engagement of shopping recommenders
  • Work on backend and infrastructure to build, deploy and serve machine learning models
  • Develop algorithms to optimize the whole page layout of the shopping surfaces
  • Drive the roadmap for next generation of shopping recommenders

What we're looking for: 

  • 5+ years working experience in the area of applied Machine Learning
  • Interest or experience working on a large-scale search, recommendation and ranking problems
  • Interest and experience in doing full stack ML, including backend and ML infrastructure
  • Experience with big data technologies MapReduce/Hadoop/Hive/Presto/Spark
  • Expert in Java, C++ or Python

#LI-LP2

Backend Engineer, Ads Indexing Platform
San Francisco, CA

About Pinterest:

Millions of people across the world come to Pinterest to find new ideas every day. It’s where they get inspiration, dream about new possibilities and plan for what matters most. Our mission is to help those people find their inspiration and create a life they love. As a Pinterest employee, you’ll be challenged to take on work that upholds this mission and pushes Pinterest forward. You’ll grow as a person and leader in your field, all the while helping users make their lives better in the positive corner of the internet.

Pinterest is one of the fastest growing online advertising platforms and our continued success depends on the reliability, performance, and scalability of the Ads indexing  systems. You’ll help us build the world class Ads indexing in-house solutions that will deliver world class performance and scale to 100x our current size. You'll join a small, early-stage team, working on multiple critical functional areas and lay the foundation for Pinterest’s business success. 

What you'll do:

  • Own and innovate the core functional areas of the Ads indexing systems
  • Re-architect core Ads indexing services/components to achieve greater scalability, freshness, performance, efficiency, and reliability
  • Apply distributed systems principles to build next-generation Ads indexing services/components
  • Develop a low latency incremental indexing pipeline to empower various of Pinterest products with the fresh Shopping Catalog data

What we're looking for:

  • 3+ years of industry experience with distributed systems, data infrastructure, and systems programming
  • Experience in solving complex scaling, latency, or performance problems in high-volume distributed systems
  • Proficiency in at least one of the systems languages (Java, C++)
  • Experience in building and owning critical user-facing backend serving systems

#LI-GK1

Verified by
Security Engineer
Tech Lead, Big Data Platform
Software Engineer
Talent Brand Manager
Sourcer
Software Engineer
You may also like