Here is a paper describing google3 https://research.google/pubs/why-google-stores-billions-of-l..., which includes the policy that βTo prevent dependency conflicts, as outlined earlier, it is important that only one version of an open source project be available at any given time.β More documentation on The One Version Rule for third_party: https://opensource.google/documentation/reference/thirdparty.... This rule may make it more difficult to upgrade a third_party library frequently since the maintainer of the third_party library has to test all the uses of the library from all teams, rather than each team deciding when to upgrade.
I used to maintain the internal (single) version of numpy and scipy and honestly I think the one-version rule was awesome. It greatly reduced noise and churn. It meant far less breakage, and it was easy to find and fix all the breaking tests when I was updating to a new version. I guess the only downside is sometimes an older version of numpy would prevent deepmind from making some sort of ML advance until I upgraded to get some new feature they wanted.
