Upgrading

Bundler 4

In order to prepare for Bundler 4, you can easily configure Bundler 2.7 to behave exactly like Bundler 4 will behave. To do so, set the environment variable BUNDLE_SIMULATE_VERSION to 4. Alternatively, you can use bundle config and enable “Bundler 4 mode” either globally through bundle config set --global simulate_version 4, or locally through bundle config set --local simulate_version 4. From now on in this document we will assume that all three of these configuration options are available, but will only mention bundle config set <option> <value>.

The following is a summary of the changes that we plan to introduce in Bundler 4, and why we will be making those changes. Some of them should be well known already by existing users, because we have been printing deprecation messages for years, but some of them are defaults that will be switched in Bundler 4 and needs some heads up.

Running just bundle will print help usage

We’re changing this default to make Bundler more friendly for new users. We do understand that long time users already know how Bundler works and find useful that just bundle defaults to bundle install. Those users can keep the existing default by configuring

bundle config default_cli_command install

Flags passed to bundle install that relied on being remembered across invocations will be removed

In particular, the --clean, --deployment, --frozen, --no-prune, --path, --shebang, --system, --without, and --with options to bundle install.

Remembering CLI options has been a source of historical confusion and bug reports, not only for beginners but also for experienced users. A CLI tool should not behave differently across exactly the same invocations unless explicitly configured to do so. This is what configuration is about after all, and things should never be silently configured without the user knowing about it.

The problem with changing this behavior is that very common workflows are relying on it. For example, when you run bundle install --without development:test in production, those flags are persisted in the app’s configuration file and further bundle invocations will happily ignore development and test gems. This magic will disappear from bundler 4, and you will explicitly need to configure it, either through environment variables, application configuration, or machine configuration. For example, with bundle config set --local without development test.

Bundler will include checksums in the lockfile by default

We shipped this security feature recently and we believe it’s time to turn it on by default, so that everyone benefits from the extra security assurances by default.

Strict source pinning in Gemfile is enforced by default

In bundler 4, the source for every dependency will be unambiguously defined, and Bundler will refuse to run otherwise.

• Multiple global Gemfile sources will no longer be supported.

  Instead of something like this:

  source "https://main_source"
  source "https://another_source"
  
  gem "dependency1"
  gem "dependency2"
  

  do something like this:

  source "https://main_source"
  
  gem "dependency1"
  
  source "https://another_source" do
    gem "dependency2"
  end
  

• Global path and git sources will no longer be supported.

  Instead of something like this:

  path "/my/path/with/gems"
  git "https://my_git_repo_with_gems"
  
  gem "dependency1"
  gem "dependency2"
  

  do something like this:

  gem "dependency1", path: "/my/path/with/gems"
  gem "dependency2", git: "https://my_git_repo_with_gems"
  

  or use the block forms if you have multiple gems for each source and you want to be a bit DRYer:

  path "/my/path/with/gems" do
    # gem "dependency1"
    # ...
    # gem "dependencyn"
  end
  
  git "https://my_git_repo_with_gems" do
    # gem "dependency1"
    # ...
    # gem "dependencyn"
  end
  

Notable CLI changes

• bundle viz will be removed and extracted to a plugin.

  This is the only bundler command requiring external dependencies, both an OS dependency (the graphviz package) and a gem dependency (the ruby-graphviz gem). Removing these dependencies will make development easier and it was also seen by the bundler team as an opportunity to develop a bundler plugin that it’s officially maintained by the bundler team, and that users can take as a reference to develop their own plugins. The plugin will contain the same code as the old core command, the only difference being that the command is now implemented as bundle graph which is much easier to understand. However, the details of the plugin are under discussion. See #3333.

• The bundle install command will no longer accept a --binstubs flag.

  The --binstubs option has been removed from bundle install and replaced with the bundle binstubs command. The --binstubs flag would create binstubs for all executables present inside the gems in the project. This was hardly useful since most users will only use a subset of all the binstubs available to them. Also, it would force the introduction of a bunch of most likely unused files into source control. Because of this, binstubs now must be created and checked into version control individually.

• The bundle inject command will be replaced with bundle add

  We believe the new command fits the user’s mental model better and it supports a wider set of use cases. The interface supported by bundle inject works exactly the same in bundle add, so it should be easy to migrate to the new command.

Other notable changes

• Git and Path gems will be included in vendor/cache by default

  We’re unsure why these gems were treated specially so we’ll start caching them normally.

• Bundler will use cached local data if available when network issues are found during resolution.

  Just trying to provide a more resilient behavior here.

• Bundler.clean_env, Bundler.with_clean_env, Bundler.clean_system, and Bundler.clean_exec will be removed

  All of these helpers ultimately use Bundler.clean_env under the hood, which makes sure all bundler-related environment are removed inside the block it yields.

  After quite a lot user reports, we noticed that users don’t usually want this but instead want the bundler environment as it was before the current process was started. Thus, Bundler.with_original_env, Bundler.original_system, and Bundler.original_exec were born. They all use the new Bundler.original_env under the hood.

  There’s however some specific cases where the good old Bundler.clean_env behavior can be useful. For example, when testing Rails generators, you really want an environment where bundler is out of the picture. This is why we decided to keep the old behavior under a new more clear name, because we figured the word “clean” was too ambiguous. So we have introduced Bundler.unbundled_env, Bundler.with_unbundled_env, Bundler.unbundled_system, and Bundler.unbundled_exec.

• Bundler.environment is deprecated in favor of Bundler.load.

  We’re not sure how people might be using this directly but we have removed the Bundler::Environment class which was instantiated by Bundler.environment since we realized the Bundler::Runtime class was the same thing. During the transition Bundler.environment will delegate to Bundler.load, which holds the reference to the Bundler::Environment.

• Deployment helpers for vlad and capistrano are being removed.

  These are natural deprecations since the vlad tool has had no activity for years whereas capistrano 3 has built-in Bundler integration in the form of the capistrano-bundler gem, and everyone using Capistrano 3 should be already using that instead. If for some reason, you are still using Capistrano 2, feel free to copy the Capistrano tasks out of the Bundler 2 file lib/bundler/deployment.rb and put them into your app.

  In general, we don’t want to maintain integrations for every deployment system out there, so that’s why we are removing these.