Wikimedia Cloud Services team/EnhancementProposals/2020 Network refresh/2021-02-24-checkin

2021-02-24 WMCS network checkin

agenda

• https://phabricator.wikimedia.org/T271058
  • buy new boxes? WMCS to decide

• https://phabricator.wikimedia.org/T209011

• procurement of hardware for new edge network setup:
  • codfw cloudsw https://phabricator.wikimedia.org/T272348 (needs discussion)

• Cross-Realm traffic guidelines document (previously known as Production Cloud services relationship)
  • https://docs.google.com/document/d/1aCKi6zsDGftUWeCIj79B0XHWYAbbyKT1HuWHQleLV2E/edit?usp=sharing
  • https://wikitech.wikimedia.org/wiki/Wikimedia_Cloud_Services_team/EnhancementProposals/Production_Cloud_services_relationship

notes

• Faidon: on cloudnet1004 HW issues: https://phabricator.wikimedia.org/T271058 feel free to request new boxes in case we need to

• No more exceptions for dumps. Two exceptions dropped! ;-)

• CloudNAT https://phabricator.wikimedia.org/T209011
  • Faidon / SRE ready to help
  • WMCS will look again and reach out

• procurement of hardware for new edge network setup:
  • codfw cloudsw https://phabricator.wikimedia.org/T272348 (needs discussion)
  • No existing underrun for this
  • Concern about time and what the need is for? (testing? to match prod?)
  • Ok to be Q4 or later
  • Purpose: Dallas is mirrored setup of production deployment in eqiad. Used as testbed for making changes. Want to ensure Dallas matches the changes in eqiad with cloudgw boxes
  • This would not be be a part of any "second" region for cloud.
  • This sounds potentially expensive. Can we reduce the need for testing replica? Can it be virtualized?
  • scale is different, it's a minimal representation
  • For SRE, "testing" in production, use incremental changes. No real testbed exists. Encourage to find ways to do this.
  • Second region could remove need for this

• Cross-Realm traffic guidelines
  • Proposal to publish the 3 common use cases, with the last complex case as a DRAFT
  • Request to create a list of services, and current thoughts on how they might map
    • Want to avoid forcing decisions before we understand them
  • Would be useful to understand the scope of the problems and services, suggesting a wikitable
  • End goal is to be able to craft new services utilizing the new document

• • Hard to map services to cases, but could map use cases to cases
  • Last case is like a loose HV, manaully managed with puppet, not controlled by neutron

• • "we can map flows more than we can map services"
  • Toolforge is a composition from many services, like NFS