Project: /_project.yaml Book: /_book.yaml
The parallel evaluation and incrementality model of Bazel.
The data model consists of the following items:
SkyValue. Also called nodes. SkyValues are immutable objects that contain all the data built over the course of the build and the inputs of the build. Examples are: input files, output files, targets and configured targets.SkyKey. A short immutable name to reference a SkyValue, for example, FILECONTENTS:/tmp/foo or PACKAGE://foo.SkyFunction. Builds nodes based on their keys and dependent nodes.Skyframe. Code name for the incremental evaluation framework Bazel is based on.A build consists of evaluating the node that represents the build request (this is the state we are striving for, but there is a lot of legacy code in the way). First its SkyFunction is found and called with the key of the top-level SkyKey. The function then requests the evaluation of the nodes it needs to evaluate the top-level node, which in turn result in other function invocations, and so on, until the leaf nodes are reached (which are usually nodes representing input files in the file system). Finally, we end up with the value of the top-level SkyValue, some side effects (such as output files in the file system) and a directed acyclic graph of the dependencies between the nodes that were involved in the build.
A SkyFunction can request SkyKeys in multiple passes if it cannot tell in advance all of the nodes it needs to do its job. A simple example is evaluating an input file node that turns out to be a symlink: the function tries to read the file, realizes that it’s a symlink, and thus fetches the file system node representing the target of the symlink. But that itself can be a symlink, in which case the original function will need to fetch its target, too.
The functions are represented in the code by the interface SkyFunction and the services provided to it by an interface called SkyFunction.Environment. These are the things functions can do:
env.getValue. If the node is available, its value is returned, otherwise, null is returned and the function itself is expected to return null. In the latter case, the dependent node is evaluated, and then the original node builder is invoked again, but this time the same env.getValue call will return a non-null value.env.getValues(). This does essentially the same, except that the dependent nodes are evaluated in parallel.SkyFunction implementations should not access data in any other way than requesting dependencies (such as by directly reading the file system), because that results in Bazel not registering the data dependency on the file that was read, thus resulting in incorrect incremental builds.
Once a function has enough data to do its job, it should return a non-null value indicating completion.
This evaluation strategy has a number of benefits:
Since functions can only access input data by depending on other nodes, Bazel can build up a complete data flow graph from the input files to the output files, and use this information to only rebuild those nodes that actually need to be rebuilt: the reverse transitive closure of the set of changed input files.
In particular, two possible incrementality strategies exist: the bottom-up one and the top-down one. Which one is optimal depends on how the dependency graph looks like.
During bottom-up invalidation, after a graph is built and the set of changed inputs is known, all the nodes are invalidated that transitively depend on changed files. This is optimal if we know that the same top-level node will be built again. Note that bottom-up invalidation requires running stat() on all input files of the previous build to determine if they were changed. This can be improved by using inotify or a similar mechanism to learn about changed files.
During top-down invalidation, the transitive closure of the top-level node is checked and only those nodes are kept whose transitive closure is clean. This is better if we know that the current node graph is large, but we only need a small subset of it in the next build: bottom-up invalidation would invalidate the larger graph of the first build, unlike top-down invalidation, which just walks the small graph of second build.
We currently only do bottom-up invalidation.
To get further incrementality, we use change pruning: if a node is invalidated, but upon rebuild, it is discovered that its new value is the same as its old value, the nodes that were invalidated due to a change in this node are “resurrected”.
This is useful, for example, if one changes a comment in a C++ file: then the .o file generated from it will be the same, thus, we don’t need to call the linker again.
The main limitation of this model is that the invalidation of a node is an all-or-nothing affair: when a dependency changes, the dependent node is always rebuilt from scratch, even if a better algorithm would exist that would mutate the old value of the node based on the changes. A few examples where this would be useful:
.class file changes in a .jar, we could theoretically modify the .jar file instead of building it from scratch again.The reason why Bazel currently does not support these things in a principled way (we have some measure of support for incremental linking, but it’s not implemented within Skyframe) is twofold: we only had limited performance gains and it was hard to guarantee that the result of the mutation is the same as that of a clean rebuild would be, and Google values builds that are bit-for-bit repeatable.
Until now, we could always achieve good enough performance by simply decomposing an expensive build step and achieving partial re-evaluation that way: it splits all the classes in an app into multiple groups and does dexing on them separately. This way, if classes in a group do not change, the dexing does not have to be redone.
This is a rough overview of some of the SkyFunction implementations Bazel uses to perform a build:
lstat(). For existent files, we also compute additional information in order to detect changes to the file. This is the lowest level node in the Skyframe graph and has no dependencies.FileStateValue and any symlinks that need to be resolved (such as the FileValue for a/b needs the resolved path of a and the resolved path of a/b). The distinction between FileStateValue is important because in some cases (for example, evaluating file system globs (such as srcs=glob(["*/*.java"])) the contents of the file are not actually needed.readdir(). Depends on the associated FileValue associated with the directory.FileValue of the associated BUILD file, and also transitively on any DirectoryListingValue that is used to resolve the globs in the package (the data structure representing the contents of a BUILD file internally)PackageValue the corresponding target is in, the ConfiguredTargetValues of direct dependencies, and a special node representing the build configuration.FileValue of the associated node, for output artifacts, it depends on the ActionExecutionValue of whatever action generates the artifact.ArtifactValues of its input files. The action it executes is currently contained within its sky key, which is contrary to the concept that sky keys should be small. We are working on solving this discrepancy (note that ActionExecutionValue and ArtifactValue are unused if we do not run the execution phase on Skyframe).