Hi all! In our organization we are a small team (i...
# ask-metaflow
g
Hi all! In our organization we are a small team (inside a big company) and we are looking for alternatives for managing our ML projects (by combining one or several open source tools). We have been considering ClearML, ZenML, Kedro+MLFlow... We want some tool to manage and organize our pipelines, but we also want to have some tool were we can manage (visualize, inspect lineage and things like that) our artifacts (say, predictions, models, etc) and also experiments. We would like to have also some kind of data versioning mechanism. I am very new to Metaflow, but I guess it covers more or less the same as Kedro, but it's not a complete MLOps solution, right? To rephrase my question, for those of you who work on premise (as in our case) and are using Metaflow, do you combine it with other tools (i.e. bentoML for deployment, dvc for data versioning, MLFlow for experimenting...), or esentially metaflow have all your needs covered? And I would like to read some feedback about the tool, are you happy with it? From what I have read in other places and the documentation it looks really nice, with a clean API, but I am willing to read from end users as well.
v
hi Jaime! Metaflow covers • data/model/artifact tracking and versioning via Metaflow artifacts • custom visualizations via
@card
• access to compute via
@kubernetes
and
@batch
and local compute • reproducibility and dependency management via
@pypi
and
@conda
and internal code packaging • secure access to data via
@secrets
and
metaflow.S3
• production workflow deployments via
step-functions
and
argo-workflows
(and
airflow
) • batch inference and deployments, integrated to surrounding systems via
@trigger
• organization of A/B tests and larger projects via namespaces and
@project
here are a few other tools that many folks use to complement Metaflow: • if you want nice visualizations to track your model training etc. out of the box, you can use e.g. the Metaflow integration in Weights and Biases or CometML, or you can use an open-source tool for that • if you need real-time inference, you can use a tool like NVIDIA Triton or BentoML or Seldon.
regarding model deployment in general, take a look at this article
happy to dive deeper into any of these topics if you have further questions 🙂
g
Thanks a lot ville for your detailed information! It's very valuable for us. Definetely we will explore the tool further. We don't have further questions yet, now it's time for us to do some "homework" and test it in our environments. But yes, it looks very promising. Thanks again!!!
🙌 1