You Don't Need a Bigger Boat
An end-to-end (Metaflow-based) implementation of an intent prediction flow for kids who can't MLOps good and wanna learn to do other stuff good too.
This is a WIP - check back often for updates.
Philosophical Motivations
There is 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; 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.
Overview
The repo shows how several (mostly open-source) tools can be effectively combined together to run data pipelines. The project current features:
- Metaflow for ML DAGs (Alternatives: Luigi (?))
- Snowflake as a data warehouse solution (Alternatives: Redshift)
- Prefect as a general orchestrator (Alternatives: Airflow)
- dbt for data transformation (Alternatives: ?)
- Great Expectations for data quality (Alternatives: dbt-expectations plugin)
- Weights&Biases for experiment tracking (Alternatives: Comet)
- Gantry for ML monitoring (Alternatives: Aporia)
- Sagemaker / Lambda for model serving (Alternatives: many)
The following picture from our Recsys paper (forthcoming) gives a quick overview of such a 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 (see below) 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.
Relevant Material
If you want to know more, you can give a look at the following material:
- "Serverless MLOps for Reasonable Companies" (video),
Data Science Meetup
, June 2021; - "You Do Not Need a Bigger Boat: Recommendations at Reasonable Scale in a (Mostly) Serverless and Open Stack" (preprint),
RecSys 2021
.
TBC
Status Update
July 2021
End-2-end flow working for remote
and local
projects; started standardizing Prefect agents with Docker and adding other services (monitoring, feature store etc.).
TO-DOs:
- dockerize the local flow;
- write-up all of this as a blog post;
- improve code / readability / docs, add potentially some more pics and some videos;
- providing an orchestrator-free version, by using step functions to manage the steps;
- finish feature store and gantry integration;
- add Github Action flow;
- continue improving the DAG card project.
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 (check back often for updates).
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).
TBC
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.
TBC
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 and drop it in a pure PaaS pipeline with Firehose and Snowpipe, for example. While a bit out-dated for some details, we championed exactly this approach a while ago: if you want to know more, you can start from this Medium post and old code.
TBC
How to Cite our Work
If you find our principles, code or data useful, please cite our work:
Paper (forthcoming in 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}
}