Stitch AR workshop notes: 20180503

stitcha-ar-workshop-notes.md

• “stitcha” name is complementary name to builda
• Goals
  • Reduce the number of asset requests on a page
  • Question the benefits of reducing the number of asset requests versus reducing the bytes for CSS (removing unused CSS)
• Traffic might have opinions about what the path is (/ac/)
• Why Google App Engine?
  • Similar pathway as Photos stack read pathway (GAE, GCP)
  • Has a lot of nice features for deploying
  • However, Security at point had concerns with GAE
• Caching is at the level of CDN
• In asset clusters, assets are time stamped and are removed as we need space
• buildaCSS skips files if they are unchanged
• There are internal apps that use pinned versions of the Toolkit CSS
• Toolkit asset could be static whereas the rest of assets would be dynamic
• Request URI limit is pretty large so we don’t think we’ll hit it. Also when compiling, we can truncate bytes/chars.
• Unresolved Qs: Why do this? Why Node? Why not modPHP in asset clusters?
  • Having a looser architecture sounds compelling (maybe for moving to the cloud?)
  • All assets should still exist in the asset cluster so if Stitcha fails, we can still use the asset cluster
• Probably leave the service dumb and remove Toolkit at the Neu-level
• This introduces a new thing that fails. Any weird cases where one component failing?
  • If Stitcha fails, we can do something at the proxy layer so that it redirects straight to the asset cluster. But, either can returns a 404.
• Scaling?
  • There is manual scaling, auto scaling, health checks API. It can scale up and down.
  • Curious, how fast can it scale up?
• Reverse fear: Right now, we only push pre-bundled files. With this, we need to push more standalone files. Can we run out of space on the - asset cluster? Our current solution doesn’t sound sustainable.
  • If there additional # files but same storage space, how does that work?
    • We can compress singular large files down smaller than if it was spaced out in more # of files
• We pay per request at GAE/GCP, consider what our SLA is
• Are product engs going to run Stitcha on their VMs?
  • Maybe. We can have it off in dev.
  • If we want DEV to mirror PROD, we can turn it on via chef.
• Debugging in production and source maps?
  • We do not have that support
• Will Stitcha work with “new thing” (ex. Webpack)?
  • At the moment, not thinking about it.
  • Need to talk about how they will intersect
• Initially, thought about doing this for CSS first, then JS. But that was before Webpack talks
• Right now we use Neu to get engineers have best practices. Willing to rip out Stitcha once we have best practices engrained in our culture - and if the “new thing” is a better solution.
• How do product teams implement new Neu Views and test the assets for both experiences?
  • Unsure. Homepage team is down to prototype that
  • Will have a config flag for transitioning
  • Stitcha is optional and not required
  • Could use Toolkit as a test for both experiences
• Call getCSS files, aggregate and de-dupe them, builds the URL, and puts it in the HTML. The product engineer will not need to know these - details.
  • Right now, there are no exclusions from getCSS.
  • Do we need more controls here for optimal caching?
  • Do we want breakpoint/media-query specific files?
    • Artifact of our existing system which will require a refactor
• Browser-level caching? Is there a performance regression?
  • All of these is predicated on being on Toolkit/Neu
  • Might be resolved with modularizing work of Toolkit
    • Breakdown Toolkit into components
    • Determine core components and keep them dynamic so that it is highly cach-able
    • Less frequency used and page-specific CSS should be the parts that are dynamic
    • This work might be done before/simultaneously with Stitcha
  • Without Stitcha, we can’t answer the question b/c we need to experiment
• Who owns pager / alerts for asset clusters? Stitcha?
  • Might be web platform
  • Line between FES and Web-Platform are getting blurry
  • Maybe pool together
• Can we be prone to attacks to overload our systems and cost us money?
  • Proxies in front of stitch should throttle this
• Current state?
  • Right now, it is in GAE sandbox
  • Still need to test load.
  • Need to figure out deployment process for prod
  • Then experiment!
• Do we need to educate the company and have open AR?
  • Maybe after load / Proof of concept testing