Rationale
One of the goals of Bazel is to create a build system where build target inputs and outputs are fully specified and therefore precisely known to the build system. This allows a more accurate analysis and determination of out-of-date build artifacts in the build system's dependency graph. Making the dependency graph analysis more deterministic leads to potential improvements in build times by avoiding re-executing unnecessary build targets. Build reliability is improved by avoiding errors where build targets might depend on out-of-date input artifacts. To achieve a more accurate dependency graph analysis, Bazel uses content digests rather than file-based timestamps. File timestamps are commonly used to detect changes in tools like Make or Apache Ant. Timestamps can be problematic when builds are distributed across multiple hosts due to issues with clock synchronization. Another Bazel's goal is to enable distributed and parallel builds on a remote cloud infrastructure. It is also designed to scale up to very large build repositories which may not be practical to download to an individual developer's work machine. Bazel provides tooling which helps developers to create bit-identical reproducible build outputs. Bazel's implemented rules avoid typical pitfalls such as embedding timestamps in generated outputs to ensure content digest matches. This, in turn, allows the build system to reliably cache (Starlark language
Bazel is extensible with its custom Starlark programming language. Starlark uses a syntax that is a subset of the syntax of the Python programming language. However, it doesn't implement many of Python's language features, such as the ability to mutate collections or access the file I/O, in order to avoid extensions that could create side-effects or create build outputs not known to the build system itself. Such side-effects could potentially lead to incorrect analysis of the build dependency graph. Bazel was designed as a multi-language build system. Many commonly used build systems are designed with a preference for a specific programming language. Examples of such systems include Ant and Maven for Java, Leiningen for Clojure, sbt for Scala, etc. In a multi-language project, combining separate build systems and achieving the build speed and correctness benefits described above can be difficult and problematic. Bazel also provides sandboxed build execution. This can be used to ensure all build dependencies have been properly specified and the build does not depend, for example, on libraries installed only locally on a developer's work computer. This helps to ensure that builds remain portable and can be executed in other (remote) environments. Build systems most similar to Bazel are Pants, Buck, and Please.Sandbox
One of the key features that differentiate Bazel and similar systems from earlier build systems is the use of a ''sandbox'' for compilation steps. When Bazel performs a separate compilation, it creates a new directory and fills it with symlinks to the explicit input dependencies for the rule. For languages like C/C++, this provides a significant safety net for the inclusion of header files: it ensures that the developer is aware of the files that are used in compilation, and it prevents the unexpected inclusion of a similarly named header file from another including directory. This sandbox approach leads to issues with common build tools, resulting in a number of workarounds required to correctly compile code under different architectures. For example, when performing separate compilation for Mac/Darwin architectures, the compiler writes the input paths into SO and OSO symbols in the Mach-O binary, which can be seen with a command likenm -a mybinary , grep SO
. These paths are needed for finding symbols during debugging. As a result, builds in Bazel must correct the compiled objects after the fact, trying to correct path-related issues that arose from the sandbox construction using flags like -fdebug-prefix-map
and -oso_prefix
, the latter having become available in Xcode 11.0. Similar handling needs to take place in linking phases, rewriting the rpath values in shared object libraries with a command like install_name_tool
.
Logo
See also
*References
External links
* {{Google FOSS Build automation Compiling tools Google software Software using the Apache license