• PackagingWG
• 2020-04-28-pip

Resolver Dev Syncup (28 April 2020)

Participants:

• Paul
• Pradyun
• Tzu-Ping

Agenda:

• pip 20.1 release
• test failures

  • https://trello.com/b/tydA30lP/new-resolver-for-pip-fix-the-failing-tests (maybe do not include in public notes?)

• 20.1 and resolver work
  • Merge the existing resolver PRs into a single PR so we can sort out the merge conflicts in advance. Paul will do this.
    • Paul will take a look at his monolithic merged PRs branch's failures.
• Visualization work
  • Update on work / discussion on the visualization design
  • Graphviz!!!
  • Paul: Looks good.
  • Tzu-Ping: Do we expect a random user to understand this graph?
    • Pradyun: no?
  • Q: Where does this fit in?
    • Pradyun: beside resolvelib. maybe?

  • https://github.com/sarugaku/resolvelib/issues/46#issuecomment-620106093

(bringing other discussions / past discussions)

• pip upgrade package [package ...]

  • TP: Definitely better than pip install --upgrade—if we can ever convince the users :p

• pip command to open project URL ( pypi.org/project/<name>/<version> ) to let the user have easy access to metadata.

  • TP: Think about this again, the setup.py I want to read is almost never the frontmost code on GitHub (that’s master, I want a specific version). A more viable solution would be for Warehouse to implement a “view content in sdist” page, and we link to that.

  • TP: Wiew wheel metadata is less useful because most people don’t understand Core Metadata format anyway
  • TODO: TP files a feature request to Warehouse
• Paul: We should try to not distract ourselves away from the tedious work of fixing the test failures though.
• When pip install fails, dump a “postmortem” in a file instead of cluttering the console
  • TP thinks this is a good idea, but we need to keep CI in mind (where the log file is not accessible, only the console)
  • NPM and Yarn both dump an error file on failures, but they still puts up verbose messages in the console
  • Pradyun: things to log:
    • choice made
    • discarding choice (backtracking)
    • note dependencies as discovered
    • I'm doing a download now! (TP thinks this is unnecessary)
  • Paul: More thinking about dumping the "state of the resolver" at the time of the exception for analysis
    • Dependency graph, in some serialised (machine readable) form.
      • TP It can be difficult to decide “which” graph to dump on multiple backtracking. Maybe dump the graph on each backtracking?
    • Discarded candidates, maybe, if resolvelib still has a record of them.