Save the Date: BazelCon 2023 will be held October 24-25 at Google Munich! Learn more

Optimize Iteration Speed

Report an issue View source

This page describes how to optimize Bazel's build performance when running Bazel repeatedly.

Bazel's Runtime State

A Bazel invocation involves several interacting parts.

  • The bazel command line interface (CLI) is the user-facing front-end tool and receives commands from the user.

  • The CLI tool starts a Bazel server for each distinct output base. The Bazel server is generally persistent, but will shut down after some idle time so as to not waste resources.

  • The Bazel server performs the loading and analysis steps for a given command (build, run, cquery, etc.), in which it constructs the necessary parts of the build graph in memory. The resulting data structures are retained in the Bazel server as part of the analysis cache.

  • The Bazel server can also perform the action execution, or it can send actions off for remote execution if it is set up to do so. The results of action executions are also cached, namely in the action cache (or execution cache, which may be either local or remote, and it may be shared among Bazel servers).

  • The result of the Bazel invocation is made available in the output tree.

Running Bazel Iteratively

In a typical developer workflow, it is common to build (or run) a piece of code repeatedly, often at a very high frequency (e.g. to resolve some compilation error or investigate a failing test). In this situation, it is important that repeated invocations of bazel have as little overhead as possible relative to the underlying, repeated action (e.g. invoking a compiler, or executing a test).

With this in mind, we take another look at Bazel's runtime state:

The analysis cache is a critical piece of data. A significant amount of time can be spent just on the loading and analysis phases of a cold run (i.e. a run just after the Bazel server was started or when the analysis cache was discarded). For a single, successful cold build (e.g. for a production release) this cost is bearable, but for repeatedly building the same target it is important that this cost be amortized and not repeated on each invocation.

The analysis cache is rather volatile. First off, it is part of the in-process state of the Bazel server, so losing the server loses the cache. But the cache is also invalidated very easily: for example, many bazel command line flags cause the cache to be discarded. This is because many flags affect the build graph (e.g. because of configurable attributes). Some flag changes can also cause the Bazel server to be restarted (e.g. changing startup options).

A good execution cache is also valuable for build performance. An execution cache can be kept locally on disk, or remotely. The cache can be shared among Bazel servers, and indeed among developers.

Avoid discarding the analysis cache

Bazel will print a warning if either the analysis cache was discarded or the server was restarted. Either of these should be avoided during iterative use:

  • Be mindful of changing bazel flags in the middle of an iterative workflow. For example, mixing a bazel build -c opt with a bazel cquery causes each command to discard the analysis cache of the other. In general, try to use a fixed set of flags for the duration of a particular workflow.

  • Losing the Bazel server loses the analysis cache. The Bazel server has a configurable idle time, after which it shuts down. You can configure this time via your bazelrc file to suit your needs. The server also restarted when startup flags change, so, again, avoid changing those flags if possible.

  • Beware that the Bazel server is killed if you press Ctrl-C repeatedly while Bazel is running. It is tempting to try to save time by interrupting a running build that is no longer needed, but only press Ctrl-C once to request a graceful end of the current invocation.

  • If you want to use multiple sets of flags from the same workspace, you can use multiple, distinct output bases, switched with the --output_base flag. Each output base gets its own Bazel server.