Awesome
You Don't Need a Bigger Boat
An end-to-end (Metaflow-based) implementation of an intent prediction (and session recommendation) flow for kids who can't MLOps good and wanna learn to do other stuff good too.
After few months of iterations, this project is now stable. Quick Links:
- Our MLOps blog series is completed;
- A new open source repo has been released, showing a simplified version of many of the concepts in this project, to provide a gentler entry point into modern MLOps pipelines;
- A second open source repo has been released in collaboration with Outerbounds and NVIDIA, showing a Merlin-focused version of many of the concepts in this project.
If you find this project (or its sister repositories above) useful, please add a star to help us spread the word!
Philosophical Motivations
There are plenty of tutorials and blog posts around the Internet on data pipelines and tooling. However:
- they (for good pedagogical reasons) tend to focus on one tool / step at a time, leaving us to wonder how the rest of the pipeline works;
- they (for good pedagogical reasons) tend to work in a toy-world fashion, leaving us to wonder what would happen when a real dataset and a real-world problem enter the scene.
This repository (and soon-to-be-drafted written tutorial) aims to fill these gaps. In particular:
- we provide open-source working code that glues together what we believe are some of the best tools in the ecosystem, going all the way from raw data to a deployed endpoint serving predictions;
- we run the pipeline under a realistic load for companies at "reasonable scale", leveraging a huge open dataset we released in 2021; moreover, we train a model for a real-world use case, and show how to monitor it after deployment.
The repo may also be seen as a (very opinionated) introduction to modern, PaaS-like pipelines (as also discussed here); while there is obviously room for disagreement over tool X or tool Y, we believe the general principles to be sound for companies at "reasonable scale": in-between bare-bone infrastructure for Tech Giants, and ready-made solutions for low-code/simple scenarios, there is a world of exciting machine learning at scale for sophisticated practitioners who don't want to waste their time managing cloud resources.
Note #1: while the code is provided as an end-to-end solution, we may sacrifice some terseness for clarity / pedagogical reasons.
Note #2: when we say the pipeline is an "end-to-end solution", we mean it - it goes from millions of raw events to a working endpoint that you can ping. As such, there are many moving pieces and it may take a while to understand how all the pieces fit together: this is not meant to be a recipe for building a small ML-powered feature, but a template for building an entire AI company (at least, the beginning of one) - as such, the learning curve is a bit steeper, but you will be rewarded with a ML stack tried and tested at unicorn scale.
Note #3: starting June 2022, a new repo is available, showcasing how to join dataOps and MLOps in a simplified, yet realistic environment: check it out as a gentler introduction to the same concepts!
Overview
The repo shows how several (mostly open-source) tools can be effectively combined together to run data pipelines at scale with very small teams. The project now features:
- Metaflow for ML DAGs
- Snowflake as a data warehouse solution (Alternatives: Redshift)
- Prefect as a general orchestrator (Alternatives: Airflow, or even Step Functions on AWS)
- dbt for data transformation (Alternatives: ?)
- Great Expectations for data quality (Alternatives: dbt-expectations plugin)
- Weights&Biases for experiment tracking (Alternatives: Comet, Neptune)
- Sagemaker / Lambda for model serving (Alternatives: many)
The following picture from Recsys gives a quick overview of a similar pipeline:
We provide two versions of the pipeline, depending on the sophistication of the setup:
- a Metaflow-only version, which runs from static data files to Sagemaker as a single Flow, and can be run from a Metaflow-enabled laptop without much additional setup;
- a data warehouse version, which runs in a more realistic setup, reading data from Snowflake and using an external orchestrator to run the steps. In this setup, the downside is that a Snowflake and a Prefect Cloud accounts are required (nonetheless, both are veasy to get); the upside is that the pipeline reflects almost perfectly a real setup, and Metaflow can be used specifically for the ML part of the process.
The parallelism between the two scenarios should be pretty clear by looking at the two projects: if you are familiarizing with all the tools for the first time, we suggest you to start from the Metaflow version and then move to the full-scale one when all the pieces of the puzzle are well understood.
Note: if you are new to Metaflow, we recommend you to go through the official installation and this stand-alone tutorial first.
Relevant Material
If you want to know more, you can give a look at the following material:
- "MLOps without much Ops",
TDS Blog Series
, September 2021; - "ML Ops at Reasonable Scale" (video),
Stanford MLSys
, July 2021; - "You Do Not Need a Bigger Boat: Recommendations at Reasonable Scale in a (Mostly) Serverless and Open Stack" pre-print,
RecSys 2021
.
Setup
General Prerequisites (do this first!)
Irrespectively of the flow you wish to run, some general tools need to be in place: Metaflow of course, as the heart of our ML practice, but also data and AWS users/roles. Please go through the general items below before tackling the flow-specific instructions.
After you finish the prerequisites below, you can run the flow you desire: each folder - remote
and local
- contains a specific README which should allow you to quickly run the project end-to-end: please refer to that documentation for flow-specific instructions.
Dataset
The project leverages the open dataset from the 2021 Coveo Data Challenge: the dataset can be downloaded directly from here (refer to the full README for terms and conditions). Data is freely available under a research-friendly license - for background information on the dataset, the use cases and relevant work in the ML literature, please refer to the accompanying paper.
Once you download and unzip the dataset in a local folder of your choice
(the zip contains 3 files, browsing_train.csv
, search_train.csv
, sku_to_content.csv
),
write down their location as an absolute path (e.g. /Users/jacopo/Documents/data/train/browsing_train.csv
):
both projects need to know where the dataset is.
AWS
Both projects - remote
and local
- use AWS services extensively - and by design: this ties back to our philosophy of PaaS-whenever-possible, and play nicely with our core adoption of Metaflow. While you can setup your users in many functionally equivalent ways, note that if you want to run the pipeline from ingestion to serving you need to be comfortable with the following AWS interactions:
- Metaflow stack (see below): we assume you installed the Metaflow stack and can run it with an AWS profile of your choice;
- Serverless stack (see below): we assume you can run
serverless deploy
in your AWS stack; - Sagemaker user: we assume you have an AWS user with permissions to manage Sagemaker endpoints (it may be totally distinct from any other Metaflow user).
Serverless
We wrap Sagemaker predictions in a serverless REST endpoint provided by AWS Lambda and API Gateway. To manage the lambda stack we use Serverless as a wrapper around AWS infrastructure.
Metaflow
Metaflow: Configuration
If you have an AWS profile configured with a metaflow-friendly user, and you created metaflow stack with CloudFormation, you can run the following command with the resources created by CloudFormation to set up metaflow on AWS:
metaflow configure aws --profile metaflow
Remember to use METAFLOW_PROFILE=metaflow
to use this profile when running a flow. Once
you completed the setup, you can run flow_playground.py
to test the AWS setup is working as expected (in particular, GPU batch jobs can run correctly). To run the flow with the custom profile created, you should do:
METAFLOW_PROFILE=metaflow python flow_playground.py run
Metaflow: Tips & Tricks
- Parallelism Safe Guard
- The flag
--max-workers
should be used to limit the maximum number of parallel steps - For example
METAFLOW_PROFILE=metaflow python flow_playground.py run --max-workers 8
limits the maximum number of parallel tasks to 8
- The flag
- Environment Variables in AWS Batch
- The
@environment
decorator is used in conjunction with@batch
to pass environment variables to AWS Batch, which will not directly have access to env variables on your local machine - In the
local
example, we use@environemnt
to pass the Weights & Biases API Key (amongst other things)
- The
- Resuming Flows
- Resuming flows is useful during development to avoid re-running compute/time intensive steps such as data preparation
METAFLOW_PROFILE=metaflow python flow_playground.py resume <STEP_NAME> --origin-run-id <RUN_ID>
- Local-Only execution
- It may sometimes be useful to debug locally (i.e to avoid Batch startup latency), we introduce a wrapper
enable_decorator
around the@batch
decorator which enables or disables a decorator's functionality - We use this in conjunction with an environment variable
EN_BATCH
to toggle the functionality of all@batch
decorators.
- It may sometimes be useful to debug locally (i.e to avoid Batch startup latency), we introduce a wrapper
FAQ
-
Both projects deal with data that has already been ingested/transmitted to the pipeline, but are silent on data collection. Any serverless option there as well?
Yes. In e-commerce use cases, for example, pixel tracking is standard (e.g. Google Analytics), so a serverless
/collect
endpoint can be used to get front-end data. In January 2022, we released a new blog post and open-source repository describing in detail a principled and serverless approach to this problem. -
What is missing / could be added if I wanted to collaborate on this project?
Few obvious things that are missing are: i) add GitHub actions for CI/CD; ii) standardize AWS permissions (as now most all commands work when launched as admin users). Want to join us? Please reach out!
Contributors
- Jacopo Tagliabue: general design and Metaflow fan boy;
- Patrick John Chia: local flow and baseline model;
- Luca Bigon: general engineering and infra optimization;
- Andrew Sutcliffe: remote flow;
- Leopoldo Garcia Vargas: QA and tests.
How to Cite our Work
If you find our principles, code or data useful, please cite our work:
Paper (RecSys2021)
@inproceedings{10.1145/3460231.3474604,
author = {Tagliabue, Jacopo},
title = {You Do Not Need a Bigger Boat: Recommendations at Reasonable Scale in a (Mostly) Serverless and Open Stack},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3460231.3474604},
doi = {10.1145/3460231.3474604},
series = {RecSys '21}
}
Data
@inproceedings{CoveoSIGIR2021,
author = {Tagliabue, Jacopo and Greco, Ciro and Roy, Jean-Francis and Bianchi, Federico and Cassani, Giovanni and Yu, Bingqing and Chia, Patrick John},
title = {SIGIR 2021 E-Commerce Workshop Data Challenge},
year = {2021},
booktitle = {SIGIR eCom 2021}
}