AWS Sagemaker is a suite of machine learning tools. Within SageMaker, a diverse array of products is available, intended to encompass various stages of the machine learning lifecycle.
We will construct a machine learning model capable of real-time operation. This model will undergo periodic training and subsequently be deployed as a real-time endpoint. To ensure the safety of our users from unfavorable results, the initial iteration of the new version will handle just 20% of the incoming traffic. As we monitor its performance, we will gradually transition the remaining traffic. The data for training this model will be sourced from our analytical database.
Developing this solution within Sagemaker entails the utilization of a variety of tools and requires a significant investment of resources from ML teams.
Here is a brief outline of the services we will need to setup:
The process above will allow us to automate our data processing, and model training.
Some additional challenges are to be expected when delivering code to production:
Constructing the aforementioned system bears resemblance to a conventional CI/CD platform, distinguished by its unique data and model attributes. This solution can be established on AWS by leveraging the following steps:
Creating a platform capable of accommodating one or two models necessitates a year and a half of dedicated engineering effort, amounting to an initial investment of approximately $330,000, all prior to reaping initial benefits.
To scale this platform and broaden its scope to cater to numerous models and additional usage scenarios, an additional three and a half full-time engineers are indispensable, elevating the overall setup cost to roughly $1,000,000.
Even with a conservative estimate, sustaining and up keeping the platform requires no less than half of the engineering team's resources, equaling an extra 2.5 full-time engineers, or a yearly maintenance expense of $500,000.
These costs do not encompass the opportunity cost incurred by waiting for the platform's construction, any licensing charges, or various human resource expenses tied to recruiting, supervising, and substituting engineers.
Collectively, the expense of developing and perpetually managing an ML platform with the capacity to accommodate a multitude of models can be substantial. It is recommended to thoughtfully evaluate requisites before embarking on this endeavor.
Qwak is built to support model productionization in an easy and cost effective way. Creating data pipelines can be done directly through the UI/CLI by creating a new feature set.
Qwak Automation empowers users to initiate both model building and deployment. This process is directly executable from the model code residing within a GitHub repository. With each build, an immutable and readily deployable model object is automatically generated. Automatic inclusion of model infrastructure and data monitoring is inherent in the system. This encompasses the ability for users to devise alerts and automated processes. Within Qwak, diverse deployment configurations are supported, including A/B testing, variable traffic allocation, and gradual rollouts.
The Qwak machine learning platform was conceived with the intention of streamlining the process for data teams to construct and implement models into production. Employing an end-to-end platform yields a diminished Total Cost of Ownership when compared to solutions reliant on numerous managed services and tools.
By leveraging an all-encompassing platform, the need for Integration Efforts is curtailed. The effort and expenses linked to constructing and maintaining integrations among various tools can be extensive. An integrated platform obviates the necessity for intricate integrations, thus conserving development resources. Additionally, inherent automation capabilities further alleviate the necessity for manual interventions, thereby diminishing labor costs.
Overall, the reduced complexity and improved productivity of an end-to-end platform contribute to a more cost-effective machine learning solution, making it a preferable choice for many organizations.
.png)
.png)
.png)