Multi-cloud with Terraform

Wed 12 Apr 2023 - 100 Words

Production Development flow for Python with Containers

Mon 27 Mar 2023 - 100 Words

Multi-model deployment, efficient?

Mon 17 Oct 2022 - 300 Words

Multi-model or Single-model Deployment Let's say we have a website selling different products, and we'd like to use Machine Learning to decide which of the many products are shown to customers at a given time. So, we go back and analyse customer behaviour, identify drivers of the metrics we care about, e.g. sales profit, and train a model we think best predicts success for our website. Then what? In order to make the predictions available to the website, we decide to wrap the model in an API, deploy the API with the model data containerized on our server infrastructure, and have the front-end services call our new model service before rendering the storefront view.

Premature Optimization, safely

Sun 16 Oct 2022 - 200 Words

Premature Optimization is a BAD Idea, unless … TODO (leading example for premature optimization) In general premature optimization is when an aspect of an implementation is optimized before the broader feature has been evaluated. Premature optimization can lead to throw-away code, technical debt, or locked-in designs. Books about software engineering practice (which ones?) teach to avoid premature optimization. In reality there are situations that call for premature optimization, e.g. when the runtime performance of an implementation is so poor that a prototype would fail to run.

Throttling in Service Stack

Sat 1 Jan 2022 - 500 Words

Throttling in Service Stack Introduction Microservice or service oriented architectures break complex functionalities down into basic services. Each service can be vended out of a separate physical cell of the infrastructure, and scaled separately. Elasticity and responsibility to scale is delegated down to the cell. With that comes the risk that the cell is not scaled on time and incoming requests exhaust resources leading to either a temporary unavailability or lasting resource exhaustion, e.

Online Training Custom Models with AWS SageMaker

Fri 4 Jun 2021 - 200 Words

Online Learning For the purpose of this article, online learning should mean that with each batch of input data a model is updated. This is to contrast offline learning where a model is trained from its unitialized origin every time. Online learning is a fundamental concept of reinforcement learning where a model updates its beliefs based on recent observations on its environment. Custom Training Images in AWS SageMaker AWS SageMaker offers an integrated environment for exploration, training, evaluation, visualization, hypter parameter optimization.

Writing Integration Tests for Infrastructure Code

Mon 1 Jan 0001 - 100 Words