What went well

• Coordination between us and Lahore
  • Timezone shift = follow the sun development
• Only one bug on release day

• Released on time (after updated schedule)

  • Touched over 500 files with no regressions and no new defects

• Bug bash was successful and offers a model for the future
• Diverse team - TNL, platform, devops - enabled us to understand the side effects of what we were doing
• Estimation overall was pretty good 
  • Initially TNL team estimated 4.5 months
  • After adding resources, able to accomplish in 2.5 months
    • Work was well suited for parallelizing, particularly addressing test failures
• Ned's test output categorization tool was very helpful!
• Pairing on transactional changes worked well in terms of peer review back and forth
• Improved a lot of things along the way
  • Requirements management - everything off of tags or hashes
• Rollback plan was written before we pushed to production

What went poorly / what can we improve

• We didn't finish before Django 1.4 was EOL'd (JE, FP)
  • We didn't start early enough
• Upgrading from 1.4 to 1.8 was a large jump (UK)
  • Jump included switching migration frameworks, the transaction framework had changed, etc - everything touching data changed
    • We didn't expect migrations to be such a huge hassle
• We have a lot of use cases for the platform and considering all of them created a lot of work
  • Some of this is not finished - we need to document the Cypress to Dogwood upgrade (potentially engineering work also)
• The Python toolchain sucks (FP, NB, JE, UK, MA, BB)
  • Dependency management sucks
    • Lots of pip quirks, including when there are multiple references to a dependency, last one wins
  • Similar dependency issue with satellite repos referencing each other caused issues and could continue to cause issues
    • Particularly an issue with devstack which does incremental updates
• Bug bash found lots of existing bugs (JE)
• We tried to write down what we were going to do, but we ended up winging it a lot anyway (NB, FP, BB)
  • Merge back to master plan
  • Release plan
  • Should have picked a consistent branch name
    • Didn't understand the variety of git usages across repos
• Some confusion resulted in duplicated work between developers (NB)
  • Lots of people in the same codebase
  • Wasn't a huge drag
• Rollback plan came together late
• Load testing should have been reviewed more thoroughly

Python toolchain sucks

• We mean pip, virtualenv, setuptools, etc
• We need to understand it better
• We need to put discipline around and and set policies about how its used
  • Standardization about how we write requirements.txt files
  • Need to clean up what we have - too much copypasta
• We should push for changes to pip that we think make sense
• Investigate other tools
• Upgrade pip after upgrading to Python 2.7.10, which gives us pip-tools
• We need an owner for pushing work forward on dependency management
  • Ned has volunteered along with Feanil and Brian
  • Think about what we need from a tool in order to do the next one better
• What group does this fall to? Release/deployment, coding standards, new?

Wrote things down but ended up winging it

• There's a difference between writing things down and figuring out what needs to be written down
• The satellite merge plan was not fleshed out well enough
• Huge number of repositories made documenting all the different scenarios difficult
• The fact that satellites have references to other satellites caused issues
• Making it up as you go isn't necessarily a problem, so long as that decision is made consciously
  • Need to write it down as you go if we start winging it
• Muzaffar's wiki page on satellite repos was very useful
  • We should add relationships between repos
• Need to designate a scribe when documenting plans as a group, and ask the question "who is in charge?"
  • Who is the keeper of history?
  • Write notes each day on what was done each day
  • This improved over the course of the project - nightly notes were sent
    • This came together when the work started to converge
  • Documentation can be overdone - can't document every little issue, otherwise we'd be overwhelmed

Action items

•  Brian Beggs Finish converting bug bash tickets
• Feanil Patel (Deactivated) Create or join working group regarding deployment & dependency management toolchain
  • Ticket to adopt better dependency management tools - pipdeptree or find or build our own
  • Document the best practices learned while doing this work, particularly wrt requirements and dependency management
  • Standardize release branching process across repos (though there are some that are forks that complicate this)
• Can we align devstack more closely with production by fixing pip or some other strategy?
• Joel Barciauskas (Deactivated) What is our overall Django upgrade strategy? Document the shell of a plan for how to tackle large upgrades like this based on our experience
  • As part of documentation of best practices, consider a "scribe" role
    
• Feanil Patel (Deactivated) Rename the table outside devops "the crisis table" and announce it at the eng all hands