Skip to content

Latest commit

 

History

History
149 lines (121 loc) · 7.98 KB

2019-02-12-execution-transitions.md

File metadata and controls

149 lines (121 loc) · 7.98 KB
Raw
created last updated status reviewers title authors discussion thread
2019-02-12
2019-03-22
Approved
gregce
Execution Transitions
katre

Abstract

Historically, Bazel has had a single configuration used for targets that need to be executed on the host. This is called the host configuration, and is created when Bazel begins analysis. Every dependency that is built for the host platform shares the same host configuration.

This fails to work properly when actions can be executed on platforms different from the host platform (for example, using remote execution). In these cases, Bazel needs a transition to an exec configuration, and not every target built to be used on an execution platform should have the same configuration.

See the related design for Toolchain Transitions.

Proposal

In order to properly transition to a configuration where the previous execution platform is the current target platform, the execution platform must be available to the transition. Note that there is no flag to directly set the execution platform (the closest flag available is --extra_execution_platforms, which adds available execution platforms, but does not select one).

This new transition needs to be usable on any attribute, regardless of whether the rule involved is providing a toolchain itself. As an example, the genrule rule has an attribute named tools, which should be built in the context of the execution platform, instead of the current host platform.

Because of this, the execution transition will be a single-instanced implementation of PatchTransition, similarly to how HostTransition.INSTANCE is implemented. This will require special handling in DependencyResolver to pass in the correct execution platform.

For native rules, accessing the new transition will involve changing transitions from .cfg(HostTransitions.INSTANCE) to .cfg(ExecTransition.INSTANCE).

For Starlark rules, a new argument to the existing cfg parameter will be added, allowing rules to declare attributes with cfg = "exec" instead of cfg = "host".

Implementation

Sidebar: Transition Factories

Currently, Bazel's ConfigurationTransition interface can be implemented by code which wants to transition from one configuration to another. However, most configuration transitions are static instances, created once for each attribute. Execution transitions need to be specific to each configured target, since they need to know the execution platform that was selected for that target.

The solution for this is a standard computer science approach: we replace static transition instances with higher-level Transition Factories. Transitions which require extra information at creation time can be created by factories which have access to the needed data. This is actually how the currently implemented SplitTransitionProvider and RuleTransitionFactory work, except in an ad hoc and unrelated manner.

In addition to the new Transition Factory interface, additional helper classes will be added such as wrappers (to convert a ConfigurationTransition to a factory) and a new ComposingTransitionFactory (analogous to the existing ComposingTransition.

Execution Transition

To implement the new execution transition, using the result of toolchain resolution, a new transition factory will be created and used. This factory will access the UnloadedToolchainContext present in the rule, and pass the selected execution platform to the newly created ConfigurationTransition instance that it creates.

In the new transition, the following steps will take place:

  1. For all fragments, the FragmentOptions.getHost method will be called.
  2. The previous execution platform's label will be copied to the --platforms flag. This will replace any value of --platforms set by getHost().
  3. After other transitions have been applied, the (in process) flag migration process will then update related legacy flags, such as --cpu and --crosstool_top.

The getHost method will be called to allow for flags to be set to reasonable default values for non-target builds. For example, users expect that when they set both --compilation_mode and --host_compilation_mode, any tools (which are built for the execution platform) will use the latter flag. This allows, for example, building output targets with debugging code, while using optimized builds for tools that are invoked during the build.

Since the selected execution platform is strictly a function of the required toolchains, the target platform, and the available execution platforms and toolchains, there are no worries that this will add any non-determinism to the build. However, the new data added (the selected execution platform label) will cause the configuration key to change, as the configuration itself is different. This could be avoided if there were an out-of-band way to pass information into a transition, or a way to use transition factories instead of static transitions when defining attribute configurations.

Bazel queries need to ensure that the --nohost_deps flag also ignores execution dependencies.

Other implementation possibilities

Instead of using a transition factory which is set unconditionally during toolchain resolution, DependencyResolver could be updated to detect that the new ExecutionTransition.INSTANCE transition is being used, and directly inject the execution platform at that point. This would simplify the configuration changes, at the cost of making the implementation less flexible. In particular, it would require more work to identify an execution transition that is part of a composed transition.

Backwards Compatibility

As this is a new transition, with a new access system, there are no backwards compatibility worries. See the migration plan for Toolchain Transitions, however, for some caveats for Starlark rules.

Over time, we should expect to replace existing uses of the host transition (both native and in Starlark) with the exec transition. This migration can be handled usecase-by-usecase, and will be addressed separately.