Skip to content
This repository has been archived by the owner on Sep 21, 2020. It is now read-only.

Latest commit

 

History

History
39 lines (27 loc) · 3.04 KB

PRINCIPLES.md

File metadata and controls

39 lines (27 loc) · 3.04 KB
Raw

Principles

This document describes various high-level prinicples, goals, and rules of thumb for gpii-infra.

Design principles

  • Favor pushing implementation "down the stack". The more we act like a "regular" Exekube project, the more we benefit from upstream improvements. Hence, favor Terraform code over in-line shell scripts over Ruby/Rake wrapper code.
    • Here is a notable exception, where moving a calculation up to Ruby/Rake resulted in much simpler Terraform code.

Operational principles

  • Avoid deploying changes near the end of your work day.
  • Avoid deploying on Friday or prior to a holiday.
  • Create JIRA tickets for issues (mainly CI/CD) we encounter and fix - to track the work done and to have data for future reference.

Kubernetes resource requests and limits

  • Use resource requests and limits for components that run in Kubernetes.
    • Among their many benefits, Stackdriver needs resource requests and limits to generate meaningful Kubernetes dashboards.
  • Memory: use the same value for Request and Limit. This makes scheduling in the cluster more predictable.
  • CPU: favor using the same value for Request and Limit as this makes scheduling in the cluster more predictable.
    • However, to ease running resource-intensive components (CouchDB, Flowmanager, Preferences) in clusters with fewer resources (dev environments), you may set a component's CPU Request to a smaller value of the component's CPU Limit (e.g. Request = 0.5 * Limit).
  • To determine initial values for a component's Requests and Limits:
    • Observe the component at idle and under some load, e.g. by running Locust tests against the component or environment.
    • Add a buffer for safety, e.g. 1.5 * observed value.
  • Favor specifying default Requests and Limits as far "down the stack" as possible, i.e. favor Chart defaults over Terraform module defaults over environment-specific settings.

Docker images

We should be well-aware of Docker images we use in our infrastructrue, as deploying an untrusted, potentially malicious, image poses a significant security threat.

As a rule of thumb: official Docker curated images (https://docs.docker.com/docker-hub/official_images/) or images directly published by trusted OSS projects are acceptable, otherwise we should build the images ourselves.

Fully-specified version vs "floating" patch version

  • In general, and especially for critical components (e.g. couchdb), prefer to use a fully specified version (e.g. 2.3.0) for maximum predictability.
    • We have observed that even a so-called "fully-specified" version (e.g. couchdb:2.3.0) can change (Docker tags are mutable).
  • For stable utility components (e.g. alpine), prefer using a "floating" patch version (e.g. 3.9, not 3.9.4). This allows us to quickly take advantage of security updates with little risk (it is unlikely that alpine:3.9.5 breaks something that worked in alpine:3.9.4).