May 10, 2021 in #personal · 5 min read
After I left Viaduct, I spent some time thinking about bridging the gap between machine learning research and production. MLOps, or the practice of unifying ML development and ML system operations, is a diverse field and can be roughly broken down into the following subcategories:
- Streamlining the industry model and ML pipeline development experience
- Making results of data science and ML workflows accessible to other stakeholders
- Deploying components of ML pipelines to production
- Maintaining ML pipelines in production
I’m mainly interested in solving problems around production ML pipelines. As an entrepreneur-in-residence at Amplify Partners, I’m currently building the open-source tooling that I wish I had around deploying and maintaining models in production.
What I’m working on
I’m excited to share a dev release of
mltrace, a coarse-grained lineage and tracing tool for ML pipelines. It’s designed for collaborative teams working on production ML pipelines, which are composed of many different components that run and update at different cadences. It makes it easy to trace a prediction or model’s output back to its most upstream, raw data file.
You’ll notice, in this release, that the tool doesn’t actually have anything ML-specific aspects to it. It’s intentionally designed for complex pipelines with many chained data transformations. ML pipelines are instances of complex data pipelines.
Logging and tracing is only a first step. The roadmap includes more features, specific to ML:
- Displaying whether components of pipelines are “stale” (ex: you need to rerun a component such as model training)
- Prometheus integrations to monitor component output distributions
- Causal analysis for ML bugs — if you flag several outputs as mispredicted, which component runs were common in producing these outputs? Which component is most likely to be the biggest culprit in an issue?
- Support for finer-grained lineage (at the record level)
Why is MLOps uniquely hard?
Although I don’t like using buzz words, “Software 2.0,” or the act of writing code to write code, is fundamentally a paradigm shift in software development for a couple of reasons:
- Model development, data exploration, and continual learning produce many more artifacts to log, version, and manage
- Most ML pipeline bugs are silent failures, not compiler or runtime errors
Debugging ML pipelines pre and post deployment is uncharted territory for many ML practitioners. Although data is actually more important than models, understanding how data impacts model performance is hard, thus motivating the roadmap for
Where I’m coming from
My academic interests revolve more heavily around deploying components of ML pipelines to production and keeping these pipelines in production. Even when companies deploy to production, they face significant challenges in staying there. In my experience, production ML performance can deviate significantly from offline benchmarks over time or experience a phenomenon known as performance drift. This isn’t just because of the ML model — most of my bugs were actually caused by some unanticipated data or ETL issue.
On a more personal note, while I was at Viaduct and working on my own ML projects, I was frustrated by a general lack of open-source tools around deploying and keeping ML pipelines in production. I tediously rolled my own terrible monitoring solution involving a daily executed Spark job that dumped metrics to a daily-partitioned table and a dashboard to depict a line graph of these metrics over time. Some problems with this solution include:
- This “online” metric was computed in Spark, while the “offline” training and evaluation metrics were computed in Python.
- If the
HivePartitionSensoron model output tables timed out, metrics weren’t computed.
- Monitoring dashboards could be vastly different for different tasks. It wasn’t scalable to build or onboard people onto.
- For any day on the metric graph, it was nearly impossible to retroactively figure out what data and code was involved in producing that metric. If the monitoring solution alerted me that something was wrong, I had no idea where to look. It felt somewhat useless to have monitoring without tracing and lineage.
For many domains, particularly those involving time-series data or user behavior, ML solutions are not viable until we have the infrastructure to easily continually train, maintain, and debug ML pipelines for even a single prediction task. ML pipelines need to reflect the most recent state of the world, and this is especially challenging when the state of the world changes frequently.
If you have any ideas or comments, feel free to email me at email@example.com. I’d love to chat with people experiencing similar problems.
Shreya Shankar likes systems and machine learning.