Action Graph Query (aquery)

The aquery command allows you to query for actions in your build graph. It operates on the post-analysis Configured Target Graph and exposes information about Actions, Artifacts and their relationships.

aquery is useful when you are interested in the properties of the Actions/Artifacts generated from the Configured Target Graph. For example, the actual commands run and their inputs/outputs/mnemonics.

The tool accepts several command-line options. Notably, the aquery command runs on top of a regular Bazel build and inherits the set of options available during a build.

It supports the same set of functions that is also available to traditional query but siblings, buildfiles and tests.

An example aquery output (without specific details):

$ bazel aquery 'deps(//some:label)'
action 'Writing file some_file_name'
  Mnemonic: ...
  Target: ...
  Configuration: ...
  ActionKey: ...
  Inputs: [...]
  Outputs: [...]

Basic syntax

A simple example of the syntax for aquery is as follows:

bazel aquery "aquery_function(function(//target))"

The query expression (in quotes) consists of the following:

  • aquery_function(...): functions specific to aquery. More details below.
  • function(...): the standard functions as traditional query.
  • //target is the label to the interested target. 
# aquery examples:
# Get the action graph generated while building //src/target_a
$ bazel aquery '//src/target_a'

# Get the action graph generated while building all dependencies of //src/target_a
$ bazel aquery 'deps(//src/target_a)'

# Get the action graph generated while building all dependencies of //src/target_a
# whose inputs filenames match the regex ".*cpp".
$ bazel aquery 'inputs(".*cpp", deps(//src/target_a))'

Using aquery functions

There are three aquery functions:

  • inputs: filter actions by inputs.
  • outputs: filter actions by outputs
  • mnemonic: filter actions by mnemonic

expr ::= inputs(word, expr)

The inputs operator returns the actions generated from building expr, whose input filenames match the regex provided by word.

$ bazel aquery 'inputs(".*cpp", deps(//src/target_a))'

outputs and mnemonic functions share a similar syntax.

You can also combine functions to achieve the AND operation. For example:

  $ bazel aquery 'mnemonic("Cpp.*", (inputs(".*cpp", inputs("foo.*", //src/target_a))))'

The above command would find all actions involved in building //src/target_a, whose mnemonics match "Cpp.*" and inputs match the patterns ".*cpp" and "foo.*".

Important: aquery functions can't be nested inside non-aquery functions

  • Conceptually this makes sense since the output of aquery functions is Actions, not Configured Targets.
  • An example of the syntax error produced: 
            $ bazel aquery 'deps(inputs(".*cpp", //src/target_a))'
        ERROR: aquery filter functions (inputs, outputs, mnemonic) produce actions,
        and therefore can't be the input of other function types: deps
        deps(inputs(".*cpp", //src/target_a))

Options

Build options

aquery runs on top of a regular Bazel build and thus inherits the set of options available during a build.

Aquery options

--output=(text|summary|proto|jsonproto|textproto), default=text

The text (default) and summary output formats are human-readable, select proto, textproto, or jsonproto for a machine-readable format. The proto message for the machine-readable formats is analysis_v2.ActionGraphContainer.

--include_commandline, default=true

Includes the content of the action command lines in the output (potentially large).

--include_artifacts, default=true

Includes names of the action inputs and outputs in the output (potentially large).

--include_aspects, default=true

Whether to include Aspect-generated actions in the output.

--include_param_files, default=false

Include the content of the param files used in the command (potentially large). Warning: Enabling this flag will automatically enable the --include_commandline flag.

--skyframe_state, default=false

Without performing extra analysis, dump the Action Graph from Skyframe. Note: Specifying a target with --skyframe_state is currently not supported. This flag is only available with --output=proto or --output=textproto.

Other tools and features

Querying against the state of Skyframe

Skyframe is the evaluation and incrementality model of Bazel. On each instance of Bazel server, Skyframe stores the dependency graph constructed from the previous runs of the Analysis phase.

In some cases, it is useful to query the Action Graph on Skyframe. An example use case would be:

  1. Run bazel build //target_a
  2. Run bazel build //target_b
  3. File foo.out was generated.

As a Bazel user, I want to determine if foo.out was generated from building //target_a or //target_b.

One could run bazel aquery 'outputs("foo.out", //target_a)' and bazel aquery 'outputs("foo.out", //target_b)' to figure out the action responsible for creating foo.out, and in turn the target. However, the number of different targets previously built can be larger than 2, which makes running multiple aquery commands a hassle.

As an alternative, the --skyframe_state flag can be used:

  # List all actions on Skyframe's action graph
  $ bazel aquery --output=proto --skyframe_state

  # or

  # List all actions on Skyframe's action graph, whose output matches "foo.out"
  $ bazel aquery --output=proto --skyframe_state 'outputs("foo.out")'

With --skyframe_state mode, aquery takes the content of the Action Graph that Skyframe keeps on the instance of Bazel, (optionally) performs filtering on it and outputs the content, without re-running the analysis phase.

Special considerations

Output format

--skyframe_state is currently only available for --output=proto and --output=textproto

Non-inclusion of target labels in the query expression

Currently, --skyframe_state queries the whole action graph that exists on Skyframe, regardless of the targets. Having the target label specified in the query together with --skyframe_state is considered a syntax error: 

  # WRONG: Target Included
  $ bazel aquery --output=proto --skyframe_state //target_a
  ERROR: Error while parsing '//target_a)': Specifying build target(s) [//target_a] with --skyframe_state is currently not supported.

  # WRONG: Target Included
  $ bazel aquery --output=proto --skyframe_state 'inputs(".*.java", //target_a)'
  ERROR: Error while parsing '//target_a)': Specifying build target(s) [//target_a] with --skyframe_state is currently not supported.

  # CORRECT: Without Target
  $ bazel aquery --output=proto --skyframe_state
  $ bazel aquery --output=proto --skyframe_state 'inputs(".*.java")'

Comparing aquery outputs

You can compare the outputs of two different aquery invocations using the aquery_differ tool. For instance: when you make some changes to your rule definition and want to verify that the command lines being run did not change. aquery_differ is the tool for that.

The tool is available in the bazelbuild/bazel repository. To use it, clone the repository to your local machine. An example usage: 

  $ bazel run //tools/aquery_differ -- \
  --before=/path/to/before.proto \
  --after=/path/to/after.proto \
  --input_type=proto \
  --attrs=cmdline \
  --attrs=inputs

The above command returns the difference between the before and after aquery outputs: which actions were present in one but not the other, which actions have different command line/inputs in each aquery output, ...). The result of running the above command would be: 

  Aquery output 'after' change contains an action that generates the following outputs that aquery output 'before' change doesn't:
  ...
  /list of output files/
  ...

  [cmdline]
  Difference in the action that generates the following output(s):
    /path/to/abc.out
  --- /path/to/before.proto
  +++ /path/to/after.proto
  @@ -1,3 +1,3 @@
    ...
    /cmdline diff, in unified diff format/
    ...

Command options

--before, --after: The aquery output files to be compared

--input_type=(proto|text_proto), default=proto: the format of the input files. Support is provided for proto and textproto aquery output.

--attrs=(cmdline|inputs), default=cmdline: the attributes of actions to be compared.

Aspect-on-aspect

It is possible for Aspects to be applied on top of each other. The aquery output of the action generated by these Aspects would then include the Aspect path, which is the sequence of Aspects applied to the target which generated the action.

An example of Aspect-on-Aspect:

  t0
  ^
  | <- a1
  t1
  ^
  | <- a2
  t2

Let ti be a target of rule ri, which applies an Aspect ai to its dependencies.

Assume that a2 generates an action X when applied to target t0. The text output of bazel aquery --include_aspects 'deps(//t2)' for action X would be:

  action ...
  Mnemonic: ...
  Target: //my_pkg:t0
  Configuration: ...
  AspectDescriptors: [//my_pkg:rule.bzl%a2(foo=...)
    -> //my_pkg:rule.bzl%a1(bar=...)]
  ...

This means that action X was generated by Aspect a2 applied onto a1(t0), where a1(t0) is the result of Aspect a1 applied onto target t0.

Each AspectDescriptor has the following format:

  AspectClass([param=value,...])

AspectClass could be the name of the Aspect class (for native Aspects) or bzl_file%aspect_name (for Starlark Aspects). AspectDescriptor are sorted in topological order of the dependency graph.

Linking with the JSON profile

While aquery provides information about the actions being run in a build (why they're being run, their inputs/outputs), the JSON profile tells us the timing and duration of their execution. It is possible to combine these 2 sets of information via a common denominator: an action's primary output.

To include actions' outputs in the JSON profile, generate the profile with --experimental_include_primary_output --noexperimental_slim_json_profile. Slim profiles are incompatible with the inclusion of primary outputs. An action's primary output is included by default by aquery.

We don't currently provide a canonical tool to combine these 2 data sources, but you should be able to build your own script with the above information.

Known issues

Handling shared actions

Sometimes actions are shared between configured targets. In the execution phase, those shared actions are simply considered as one and only executed once. However, aquery operates on the pre-execution, post-analysis action graph, and hence treats these like separate actions whose output Artifacts have the exact same execPath. As a result, equivalent Artifacts appear duplicated.

The list of aquery issues/planned features can be found on GitHub.

FAQs

The ActionKey remains the same even though the content of an input file changed.

In the context of aquery, the ActionKey refers to the String gotten from ActionAnalysisMetadata#getKey: 

  Returns a string encoding all of the significant behaviour of this Action that might affect the
  output. The general contract of getKey is this: if the work to be performed by the
  execution of this action changes, the key must change.

  ...

  Examples of changes that should affect the key are:

  - Changes to the BUILD file that materially affect the rule which gave rise to this Action.
  - Changes to the command-line options, environment, or other global configuration resources
      which affect the behaviour of this kind of Action (other than changes to the names of the
      input/output files, which are handled externally).
  - An upgrade to the build tools which changes the program logic of this kind of Action
      (typically this is achieved by incorporating a UUID into the key, which is changed each
      time the program logic of this action changes).
  Note the following exception: for actions that discover inputs, the key must change if any
  input names change or else action validation may falsely validate.

This excludes the changes to the content of the input files, and is not to be confused with RemoteCacheClient#ActionKey.

Updates

For any issues/feature requests, please file an issue here.