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.
bazelcommand line interface (CLI) is the user-facing front-end tool and receives commands from the user.
The Bazel server performs the loading and analysis steps for a given command (
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
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
bazelflags in the middle of an iterative workflow. For example, mixing a
bazel build -c optwith a
bazel cquerycauses 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_baseflag. Each output base gets its own Bazel server.
To make this condition an error rather than a warning, you can use the
--noallow_analysis_cache_discard flag (introduced in Bazel 6.4.0)