Kubeflow

What do you like best about Kubeflow?

I like the portability of it, which makes easier to work with any kubernete clusters whether it's on single computer or in cloud.

What do you dislike about Kubeflow?

It was difficult to setup initially we had to keep dedicated team members to setup it.

What problems is Kubeflow solving and how is that benefiting you?

It is very helpful for when it comes to simplifying ML workflows after implementing Kuberflow the efficiency of workflow has been increased.

What do you like best about Kubeflow?

1. It uses Kubernetes as a backend.

2. It adheres to follow best practices of Mlops & containerization.

3. Once a workflow is properly defined then it becomes very easy to automate it.

4. It does a great python sdk to design pipeline.

5. The Front end/UI to use Kubeflow pipeline is awesome.

6. It also displayed all the logs.

What do you dislike about Kubeflow?

1. Initial steep learning curve as it involves lot of variety of concepts under one roof.

2. So the user must have knowledge apart from usual ML stuffs about Docker/Container tech, kubernetes.

3. Even the initial setup process is not so initiative.

4. Based on what material is available on its docs, it seems setting it up is comparatively easy on GCP (in fact I have use it only on GCP)

Recommendations to others considering Kubeflow:

1. If you are already using Kubernetes then adding Kubeflow to your stack will supercharge your workflows.

2. You'll have to adapt Microservice approach which will definitely provide you benefits in the longs run.

3. But be prepared for the initial steep learning curve and not so easy setup process.

What problems is Kubeflow solving and how is that benefiting you?

1. One-stop shop for orchestrating any workflow using Kubernetes.

2. We used Kubernets as backend already prior to Kubeflow and not all ML engineers were comfortable to use it. Kubeflow solved this problem as it too uses kubernets as backend but also provided a nice initiative UI to control workflows.

3. We mostly use Kubeflow for all our Computer Vision use case.

4. It involves training, inference and even internal serving. For external clients, we had in-house developed serving infra.

5. After adapting to Kubeflow, we had to also adapt the MIcroservice approach, which was blessings in disguise.

What do you like best about Kubeflow?

Automates flow of production machine learning. Kubeflow can be easily integrated with kubernetes on a lot of different cloud providers, such as Amazon web service (using Elastic Kubernetes Service), or with Google cloud (with Google Kubernetes Engine). It has API interface in different languages, espically easy to integrate with python and docker containers. Which helps users to build their own rerunnable and plugable machine learning pipelines.

What do you dislike about Kubeflow?

No easy integration with terraform and integration with domain name servers on Amazon web service. Which means that deploying kubeflow can be difficult dependent on what existing infrastructure looks like. If companies already have existing models to integrate with kubeflow that does not use containers, it could cost extra effort to implement them as Kubeflow is best used with docker containers and run on kubernetes.

Recommendations to others considering Kubeflow:

Kubeflow is one of the technologies that works best with kubernetes and one of the newer machine learning technologies that supports pipelines building which traditionally has been difficult in the field of machine learning.

What problems is Kubeflow solving and how is that benefiting you?

Production machine learning problems can be solved with kubeflow as well as pipeline building. The benefits to kubeflow are ease of use, one centralised UI and ease of integration with docker technologies. For data scientist who do not want to write a lot of code, Kubeflow provides a nice way to run and rerun experiments, train models, publish models as well as managing pipelines.

What do you like best about Kubeflow?

Scability, portability and distribute. The all-in-one feature of Kubeflow has made team easy to use and have saved lot amount of time .This is easy to use for new learner.

What do you dislike about Kubeflow?

There was a need of CI/CD feature to the team. On Kubeflow couldn't find the feature of CI/CD.

Recommendations to others considering Kubeflow:

Earlier, had used Airflow in combination with other softwares to solve the same purpose. However, with Kubeflow the life has become much easy for it's rich feature for develpoment and deployment of ML models.

What problems is Kubeflow solving and how is that benefiting you?

Automation of ML models. For development of ML workflow system. Also for creating ML system with all it's components. This has saved a lot of time and energy for model architect and model developers.

What do you like best about Kubeflow?

1. The kubeflow is based on kubernetes, it makes the scaling of models and load balancer quite easy

2. The pipelines are very elegant and make the stages very clear

What do you dislike about Kubeflow?

1. The documents of kubeflow is incomplete and some examples of source codes ( especially for docker images ) are difficult to find

2. There are no simple examples of data passing in different stages in the pipelines

3. The learning curve of DSL is high for data scientists

Recommendations to others considering Kubeflow:

Kubeflow is a great platform for model deployment and there is some learning curve for data scientist.

What problems is Kubeflow solving and how is that benefiting you?

Our team is exploring different platforms to deploy mode for production and want to find the most suitable platform

The most benefits for kubeflow is it based on kubernetes, it makes the load balancer quite easy