The Bazel Query Reference

Report an issue View source Nightly · 7.2 · 7.1 · 7.0 · 6.5 · 6.4

This page is the reference manual for the Bazel Query Language used when you use bazel query to analyze build dependencies. It also describes the output formats bazel query supports.

For practical use cases, see the Bazel Query How-To.

Additional query reference

In addition to query, which runs on the post-loading phase target graph, Bazel includes action graph query and configurable query.

Action graph query

The action graph query (aquery) 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, and mnemonics.

For more details, see the aquery reference.

Configurable query

Traditional Bazel query runs on the post-loading phase target graph and therefore has no concept of configurations and their related concepts. Notably, it doesn't correctly resolve select statements and instead returns all possible resolutions of selects. However, the configurable query environment, cquery, properly handles configurations but doesn't provide all of the functionality of this original query.

For more details, see the cquery reference.


How do people use bazel query? Here are typical examples:

Why does the //foo tree depend on //bar/baz? Show a path:

somepath(foo/..., //bar/baz:all)

What C++ libraries do all the foo tests depend on that the foo_bin target does not?

kind("cc_library", deps(kind(".*test rule", foo/...)) except deps(//foo:foo_bin))

Tokens: The lexical syntax

Expressions in the query language are composed of the following tokens:

  • Keywords, such as let. Keywords are the reserved words of the language, and each of them is described below. The complete set of keywords is:

  • Words, such as "foo/..." or ".*test rule" or "//bar/baz:all". If a character sequence is "quoted" (begins and ends with a single-quote ' or begins and ends with a double-quote "), it is a word. If a character sequence is not quoted, it may still be parsed as a word. Unquoted words are sequences of characters drawn from the alphabet characters A-Za-z, the numerals 0-9, and the special characters */@.-_:$~[] (asterisk, forward slash, at, period, hyphen, underscore, colon, dollar sign, tilde, left square brace, right square brace). However, unquoted words may not start with a hyphen - or asterisk * even though relative [target names][(/concepts/labels#target-names) may start with those characters.

    Unquoted words also may not include the characters plus sign + or equals sign =, even though those characters are permitted in target names. When writing code that generates query expressions, target names should be quoted.

    Quoting is necessary when writing scripts that construct Bazel query expressions from user-supplied values.

     //foo:bar+wiz    # WRONG: scanned as //foo:bar + wiz.
     //foo:bar=wiz    # WRONG: scanned as //foo:bar = wiz.
     "//foo:bar+wiz"  # OK.
     "//foo:bar=wiz"  # OK.

    Note that this quoting is in addition to any quoting that may be required by your shell, such as:

    bazel query ' "//foo:bar=wiz" '   # single-quotes for shell, double-quotes for Bazel.

    Keywords, when quoted, are treated as ordinary words. For example, some is a keyword but "some" is a word. Both foo and "foo" are words.

    However, be careful when using single or double quotes in target names. When quoting one or more target names, use only one type of quotes (either all single or all double quotes).

    The following are examples of what the Java query string will be:

      'a"'a'         # WRONG: Error message: unclosed quotation.
      "a'"a"         # WRONG: Error message: unclosed quotation.
      '"a" + 'a''    # WRONG: Error message: unexpected token 'a' after query expression '"a" + '
      "'a' + "a""    # WRONG: Error message: unexpected token 'a' after query expression ''a' + '
      "a'a"          # OK.
      'a"a'          # OK.
      '"a" + "a"'    # OK
      "'a' + 'a'"    # OK

    We chose this syntax so that quote marks aren't needed in most cases. The (unusual) ".*test rule" example needs quotes: it starts with a period and contains a space. Quoting "cc_library" is unnecessary but harmless.

  • Punctuation, such as parens (), period . and comma ,. Words containing punctuation (other than the exceptions listed above) must be quoted.

Whitespace characters outside of a quoted word are ignored.

Bazel query language concepts

The Bazel query language is a language of expressions. Every expression evaluates to a partially-ordered set of targets, or equivalently, a graph (DAG) of targets. This is the only datatype.

Set and graph refer to the same datatype, but emphasize different aspects of it, for example:

  • Set: The partial order of the targets is not interesting.
  • Graph: The partial order of targets is significant.

Cycles in the dependency graph

Build dependency graphs should be acyclic.

The algorithms used by the query language are intended for use in acyclic graphs, but are robust against cycles. The details of how cycles are treated are not specified and should not be relied upon.

Implicit dependencies

In addition to build dependencies that are defined explicitly in BUILD files, Bazel adds additional implicit dependencies to rules. For example every Java rule implicitly depends on the JavaBuilder. Implicit dependencies are established using attributes that start with $ and they cannot be overridden in BUILD files.

Per default bazel query takes implicit dependencies into account when computing the query result. This behavior can be changed with the --[no]implicit_deps option. Note that, as query does not consider configurations, potential toolchains are never considered.


Bazel query language expressions operate over the build dependency graph, which is the graph implicitly defined by all rule declarations in all BUILD files. It is important to understand that this graph is somewhat abstract, and does not constitute a complete description of how to perform all the steps of a build. In order to perform a build, a configuration is required too; see the configurations section of the User's Guide for more detail.

The result of evaluating an expression in the Bazel query language is true for all configurations, which means that it may be a conservative over-approximation, and not exactly precise. If you use the query tool to compute the set of all source files needed during a build, it may report more than are actually necessary because, for example, the query tool will include all the files needed to support message translation, even though you don't intend to use that feature in your build.

On the preservation of graph order

Operations preserve any ordering constraints inherited from their subexpressions. You can think of this as "the law of conservation of partial order". Consider an example: if you issue a query to determine the transitive closure of dependencies of a particular target, the resulting set is ordered according to the dependency graph. If you filter that set to include only the targets of file kind, the same transitive partial ordering relation holds between every pair of targets in the resulting subset - even though none of these pairs is actually directly connected in the original graph. (There are no file-file edges in the build dependency graph).

However, while all operators preserve order, some operations, such as the set operations don't introduce any ordering constraints of their own. Consider this expression:

deps(x) union y

The order of the final result set is guaranteed to preserve all the ordering constraints of its subexpressions, namely, that all the transitive dependencies of x are correctly ordered with respect to each other. However, the query guarantees nothing about the ordering of the targets in y, nor about the ordering of the targets in deps(x) relative to those in y (except for those targets in y that also happen to be in deps(x)).

Operators that introduce ordering constraints include: allpaths, deps, rdeps, somepath, and the target pattern wildcards package:*, dir/..., etc.

Sky query

Sky Query is a mode of query that operates over a specified universe scope.

Special functions available only in SkyQuery

Sky Query mode has the additional query functions allrdeps and rbuildfiles. These functions operate over the entire universe scope (which is why they don't make sense for normal Query).

Specifying a universe scope

Sky Query mode is activated by passing the following two flags: (--universe_scope or --infer_universe_scope) and --order_output=no. --universe_scope=<target_pattern1>,...,<target_patternN> tells query to preload the transitive closure of the target pattern specified by the target patterns, which can be both additive and subtractive. All queries are then evaluated in this "scope". In particular, the allrdeps and rbuildfiles operators only return results from this scope. --infer_universe_scope tells Bazel to infer a value for --universe_scope from the query expression. This inferred value is the list of unique target patterns in the query expression, but this might not be what you want. For example:

bazel query --infer_universe_scope --order_output=no "allrdeps(//my:target)"

The list of unique target patterns in this query expression is ["//my:target"], so Bazel treats this the same as the invocation:

bazel query --universe_scope=//my:target --order_output=no "allrdeps(//my:target)"

But the result of that query with --universe_scope is only //my:target; none of the reverse dependencies of //my:target are in the universe, by construction! On the other hand, consider:

bazel query --infer_universe_scope --order_output=no "tests(//a/... + b/...) intersect allrdeps(siblings(rbuildfiles(my/starlark/file.bzl)))"

This is a meaningful query invocation that is trying to compute the test targets in the tests expansion of the targets under some directories that transitively depend on targets whose definition uses a certain .bzl file. Here, --infer_universe_scope is a convenience, especially in the case where the choice of --universe_scope would otherwise require you to parse the query expression yourself.

So, for query expressions that use universe-scoped operators like allrdeps and rbuildfiles be sure to use --infer_universe_scope only if its behavior is what you want.

Sky Query has some advantages and disadvantages compared to the default query. The main disadvantage is that it cannot order its output according to graph order, and thus certain output formats are forbidden. Its advantages are that it provides two operators (allrdeps and rbuildfiles) that are not available in the default query. As well, Sky Query does its work by introspecting the Skyframe graph, rather than creating a new graph, which is what the default implementation does. Thus, there are some circumstances in which it is faster and uses less memory.

Expressions: Syntax and semantics of the grammar

This is the grammar of the Bazel query language, expressed in EBNF notation:

expr ::= word
       | let name = expr in expr
       | (expr)
       | expr intersect expr
       | expr ^ expr
       | expr union expr
       | expr + expr
       | expr except expr
       | expr - expr
       | set(word *)
       | word '(' int | word | expr ... ')'

The following sections describe each of the productions of this grammar in order.

Target patterns

expr ::= word

Syntactically, a target pattern is just a word. It's interpreted as an (unordered) set of targets. The simplest target pattern is a label, which identifies a single target (file or rule). For example, the target pattern //foo:bar evaluates to a set containing one element, the target, the bar rule.

Target patterns generalize labels to include wildcards over packages and targets. For example, foo/...:all (or just foo/...) is a target pattern that evaluates to a set containing all rules in every package recursively beneath the foo directory; bar/baz:all is a target pattern that evaluates to a set containing all the rules in the bar/baz package, but not its subpackages.

Similarly, foo/...:* is a target pattern that evaluates to a set containing all targets (rules and files) in every package recursively beneath the foo directory; bar/baz:* evaluates to a set containing all the targets in the bar/baz package, but not its subpackages.

Because the :* wildcard matches files as well as rules, it's often more useful than :all for queries. Conversely, the :all wildcard (implicit in target patterns like foo/...) is typically more useful for builds.

bazel query target patterns work the same as bazel build build targets do. For more details, see Target Patterns, or type bazel help target-syntax.

Target patterns may evaluate to a singleton set (in the case of a label), to a set containing many elements (as in the case of foo/..., which has thousands of elements) or to the empty set, if the target pattern matches no targets.

All nodes in the result of a target pattern expression are correctly ordered relative to each other according to the dependency relation. So, the result of foo:* is not just the set of targets in package foo, it is also the graph over those targets. (No guarantees are made about the relative ordering of the result nodes against other nodes.) For more details, see the graph order section.


expr ::= let name = expr1 in expr2
       | $name

The Bazel query language allows definitions of and references to variables. The result of evaluation of a let expression is the same as that of expr2, with all free occurrences of variable name replaced by the value of expr1.

For example, let v = foo/... in allpaths($v, //common) intersect $v is equivalent to the allpaths(foo/...,//common) intersect foo/....

An occurrence of a variable reference name other than in an enclosing let name = ... expression is an error. In other words, top-level query expressions cannot have free variables.

In the above grammar productions, name is like word, but with the additional constraint that it be a legal identifier in the C programming language. References to the variable must be prepended with the "$" character.

Each let expression defines only a single variable, but you can nest them.

Both target patterns and variable references consist of just a single token, a word, creating a syntactic ambiguity. However, there is no semantic ambiguity, because the subset of words that are legal variable names is disjoint from the subset of words that are legal target patterns.

Technically speaking, let expressions do not increase the expressiveness of the query language: any query expressible in the language can also be expressed without them. However, they improve the conciseness of many queries, and may also lead to more efficient query evaluation.

Parenthesized expressions

expr ::= (expr)

Parentheses associate subexpressions to force an order of evaluation. A parenthesized expression evaluates to the value of its argument.

Algebraic set operations: intersection, union, set difference

expr ::= expr intersect expr
       | expr ^ expr
       | expr union expr
       | expr + expr
       | expr except expr
       | expr - expr

These three operators compute the usual set operations over their arguments. Each operator has two forms, a nominal form, such as intersect, and a symbolic form, such as ^. Both forms are equivalent; the symbolic forms are quicker to type. (For clarity, the rest of this page uses the nominal forms.)

For example,

foo/... except foo/bar/...

evaluates to the set of targets that match foo/... but not foo/bar/....

You can write the same query as:

foo/... - foo/bar/...

The intersect (^) and union (+) operations are commutative (symmetric); except (-) is asymmetric. The parser treats all three operators as left-associative and of equal precedence, so you might want parentheses. For example, the first two of these expressions are equivalent, but the third is not:

x intersect y union z
(x intersect y) union z
x intersect (y union z)

Read targets from an external source: set

expr ::= set(word *)

The set(a b c ...) operator computes the union of a set of zero or more target patterns, separated by whitespace (no commas).

In conjunction with the Bourne shell's $(...) feature, set() provides a means of saving the results of one query in a regular text file, manipulating that text file using other programs (such as standard UNIX shell tools), and then introducing the result back into the query tool as a value for further processing. For example:

bazel query deps(//my:target) --output=label | grep ... | sed ... | awk ... > foo
bazel query "kind(cc_binary, set($(<foo)))"

In the next example,kind(cc_library, deps(//some_dir/foo:main, 5)) is computed by filtering on the maxrank values using an awk program.

bazel query 'deps(//some_dir/foo:main)' --output maxrank | awk '($1 < 5) { print $2;} ' > foo
bazel query "kind(cc_library, set($(<foo)))"

In these examples, $(<foo) is a shorthand for $(cat foo), but shell commands other than cat may be used too—such as the previous awk command.


expr ::= word '(' int | word | expr ... ')'

The query language defines several functions. The name of the function determines the number and type of arguments it requires. The following functions are available:

Transitive closure of dependencies: deps

expr ::= deps(expr)
       | deps(expr, depth)

The deps(x) operator evaluates to the graph formed by the transitive closure of dependencies of its argument set x. For example, the value of deps(//foo) is the dependency graph rooted at the single node foo, including all its dependencies. The value of deps(foo/...) is the dependency graphs whose roots are all rules in every package beneath the foo directory. In this context, 'dependencies' means only rule and file targets, therefore the BUILD and Starlark files needed to create these targets are not included here. For that you should use the buildfiles operator.

The resulting graph is ordered according to the dependency relation. For more details, see the section on graph order.

The deps operator accepts an optional second argument, which is an integer literal specifying an upper bound on the depth of the search. So deps(foo:*, 0) returns all targets in the foo package, while deps(foo:*, 1) further includes the direct prerequisites of any target in the foo package, and deps(foo:*, 2) further includes the nodes directly reachable from the nodes in deps(foo:*, 1), and so on. (These numbers correspond to the ranks shown in the minrank output format.) If the depth parameter is omitted, the search is unbounded: it computes the reflexive transitive closure of prerequisites.

Transitive closure of reverse dependencies: rdeps

expr ::= rdeps(expr, expr)
       | rdeps(expr, expr, depth)

The rdeps(u, x) operator evaluates to the reverse dependencies of the argument set x within the transitive closure of the universe set u.

The resulting graph is ordered according to the dependency relation. See the section on graph order for more details.

The rdeps operator accepts an optional third argument, which is an integer literal specifying an upper bound on the depth of the search. The resulting graph only includes nodes within a distance of the specified depth from any node in the argument set. So rdeps(//foo, //common, 1) evaluates to all nodes in the transitive closure of //foo that directly depend on //common. (These numbers correspond to the ranks shown in the minrank output format.) If the depth parameter is omitted, the search is unbounded.

Transitive closure of all reverse dependencies: allrdeps

expr ::= allrdeps(expr)
       | allrdeps(expr, depth)

The allrdeps operator behaves just like the rdeps operator, except that the "universe set" is whatever the --universe_scope flag evaluated to, instead of being separately specified. Thus, if --universe_scope=//foo/... was passed, then allrdeps(//bar) is equivalent to rdeps(//foo/..., //bar).

Direct reverse dependencies in the same package: same_pkg_direct_rdeps

expr ::= same_pkg_direct_rdeps(expr)

The same_pkg_direct_rdeps(x) operator evaluates to the full set of targets that are in the same package as a target in the argument set, and which directly depend on it.

Dealing with a target's package: siblings

expr ::= siblings(expr)

The siblings(x) operator evaluates to the full set of targets that are in the same package as a target in the argument set.

Arbitrary choice: some

expr ::= some(expr)
       | some(expr, count )

The some(x, k) operator selects at most k targets arbitrarily from its argument set x, and evaluates to a set containing only those targets. Parameter k is optional; if missing, the result will be a singleton set containing only one target arbitrarily selected. If the size of argument set x is smaller than k, the whole argument set x will be returned.

For example, the expression some(//foo:main union //bar:baz) evaluates to a singleton set containing either //foo:main or //bar:baz—though which one is not defined. The expression some(//foo:main union //bar:baz, 2) or some(//foo:main union //bar:baz, 3) returns both //foo:main and //bar:baz.

If the argument is a singleton, then some computes the identity function: some(//foo:main) is equivalent to //foo:main.

It is an error if the specified argument set is empty, as in the expression some(//foo:main intersect //bar:baz).

Path operators: somepath, allpaths

expr ::= somepath(expr, expr)
       | allpaths(expr, expr)

The somepath(S, E) and allpaths(S, E) operators compute paths between two sets of targets. Both queries accept two arguments, a set S of starting points and a set E of ending points. somepath returns the graph of nodes on some arbitrary path from a target in S to a target in E; allpaths returns the graph of nodes on all paths from any target in S to any target in E.

The resulting graphs are ordered according to the dependency relation. See the section on graph order for more details.

somepath(S1 + S2, E), one possible result.
somepath(S1 + S2, E), another possible result.
allpaths(S1 + S2, E)

Target kind filtering: kind

expr ::= kind(word, expr)

The kind(pattern, input) operator applies a filter to a set of targets, and discards those targets that are not of the expected kind. The pattern parameter specifies what kind of target to match.

For example, the kinds for the four targets defined by the BUILD file (for package p) shown below are illustrated in the table:

Code Target Kind
            name = "a",
            srcs = [""],
            outs = ["a.out"],
            cmd = "...",
//p:a genrule rule
// source file
//p:a.out generated file
//p:BUILD source file

Thus, kind("cc_.* rule", foo/...) evaluates to the set of all cc_library, cc_binary, etc, rule targets beneath foo, and kind("source file", deps(//foo)) evaluates to the set of all source files in the transitive closure of dependencies of the //foo target.

Quotation of the pattern argument is often required because without it, many regular expressions, such as source file and .*_test, are not considered words by the parser.

When matching for package group, targets ending in :all may not yield any results. Use :all-targets instead.

Target name filtering: filter

expr ::= filter(word, expr)

The filter(pattern, input) operator applies a filter to a set of targets, and discards targets whose labels (in absolute form) do not match the pattern; it evaluates to a subset of its input.

The first argument, pattern is a word containing a regular expression over target names. A filter expression evaluates to the set containing all targets x such that x is a member of the set input and the label (in absolute form, such as //foo:bar) of x contains an (unanchored) match for the regular expression pattern. Since all target names start with //, it may be used as an alternative to the ^ regular expression anchor.

This operator often provides a much faster and more robust alternative to the intersect operator. For example, in order to see all bar dependencies of the //foo:foo target, one could evaluate

deps(//foo) intersect //bar/...

This statement, however, will require parsing of all BUILD files in the bar tree, which will be slow and prone to errors in irrelevant BUILD files. An alternative would be:

filter(//bar, deps(//foo))

which would first calculate the set of //foo dependencies and then would filter only targets matching the provided pattern—in other words, targets with names containing //bar as a substring.

Another common use of the filter(pattern, expr) operator is to filter specific files by their name or extension. For example,

filter("\.cc$", deps(//foo))

will provide a list of all .cc files used to build //foo.

Rule attribute filtering: attr

expr ::= attr(word, word, expr)

The attr(name, pattern, input) operator applies a filter to a set of targets, and discards targets that aren't rules, rule targets that do not have attribute name defined or rule targets where the attribute value does not match the provided regular expression pattern; it evaluates to a subset of its input.

The first argument, name is the name of the rule attribute that should be matched against the provided regular expression pattern. The second argument, pattern is a regular expression over the attribute values. An attr expression evaluates to the set containing all targets x such that x is a member of the set input, is a rule with the defined attribute name and the attribute value contains an (unanchored) match for the regular expression pattern. If name is an optional attribute and rule does not specify it explicitly then default attribute value will be used for comparison. For example,

attr(linkshared, 0, deps(//foo))

will select all //foo dependencies that are allowed to have a linkshared attribute (such as, cc_binary rule) and have it either explicitly set to 0 or do not set it at all but default value is 0 (such as for cc_binary rules).

List-type attributes (such as srcs, data, etc) are converted to strings of the form [value<sub>1</sub>, ..., value<sub>n</sub>], starting with a [ bracket, ending with a ] bracket and using "," (comma, space) to delimit multiple values. Labels are converted to strings by using the absolute form of the label. For example, an attribute deps=[":foo", "//otherpkg:bar", "wiz"] would be converted to the string [//thispkg:foo, //otherpkg:bar, //thispkg:wiz]. Brackets are always present, so the empty list would use string value [] for matching purposes. For example,

attr("srcs", "\[\]", deps(//foo))

will select all rules among //foo dependencies that have an empty srcs attribute, while

attr("data", ".{3,}", deps(//foo))

will select all rules among //foo dependencies that specify at least one value in the data attribute (every label is at least 3 characters long due to the // and :).

To select all rules among //foo dependencies with a particular value in a list-type attribute, use

attr("tags", "[\[ ]value[,\]]", deps(//foo))

This works because the character before value will be [ or a space and the character after value will be a comma or ].

Rule visibility filtering: visible

expr ::= visible(expr, expr)

The visible(predicate, input) operator applies a filter to a set of targets, and discards targets without the required visibility.

The first argument, predicate, is a set of targets that all targets in the output must be visible to. A visible expression evaluates to the set containing all targets x such that x is a member of the set input, and for all targets y in predicate x is visible to y. For example:

visible(//foo, //bar:*)

will select all targets in the package //bar that //foo can depend on without violating visibility restrictions.

Evaluation of rule attributes of type label: labels

expr ::= labels(word, expr)

The labels(attr_name, inputs) operator returns the set of targets specified in the attribute attr_name of type "label" or "list of label" in some rule in set inputs.

For example, labels(srcs, //foo) returns the set of targets appearing in the srcs attribute of the //foo rule. If there are multiple rules with srcs attributes in the inputs set, the union of their srcs is returned.

Expand and filter test_suites: tests

expr ::= tests(expr)

The tests(x) operator returns the set of all test rules in set x, expanding any test_suite rules into the set of individual tests that they refer to, and applying filtering by tag and size.

By default, query evaluation ignores any non-test targets in all test_suite rules. This can be changed to errors with the --strict_test_suite option.

For example, the query kind(test, foo:*) lists all the *_test and test_suite rules in the foo package. All the results are (by definition) members of the foo package. In contrast, the query tests(foo:*) will return all of the individual tests that would be executed by bazel test foo:*: this may include tests belonging to other packages, that are referenced directly or indirectly via test_suite rules.

Package definition files: buildfiles

expr ::= buildfiles(expr)

The buildfiles(x) operator returns the set of files that define the packages of each target in set x; in other words, for each package, its BUILD file, plus any .bzl files it references via load. Note that this also returns the BUILD files of the packages containing these loaded files.

This operator is typically used when determining what files or packages are required to build a specified target, often in conjunction with the --output package option, below). For example,

bazel query 'buildfiles(deps(//foo))' --output package

returns the set of all packages on which //foo transitively depends.

Package definition files: rbuildfiles

expr ::= rbuildfiles(word, ...)

The rbuildfiles operator takes a comma-separated list of path fragments and returns the set of BUILD files that transitively depend on these path fragments. For instance, if //foo is a package, then rbuildfiles(foo/BUILD) will return the //foo:BUILD target. If the foo/BUILD file has load('//bar:file.bzl'... in it, then rbuildfiles(bar/file.bzl) will return the //foo:BUILD target, as well as the targets for any other BUILD files that load //bar:file.bzl

The scope of the rbuildfiles operator is the universe specified by the --universe_scope flag. Files that do not correspond directly to BUILD files and .bzl files do not affect the results. For instance, source files (like are ignored, even if they are explicitly mentioned in the BUILD file. Symlinks, however, are respected, so that if foo/BUILD is a symlink to bar/BUILD, then rbuildfiles(bar/BUILD) will include //foo:BUILD in its results.

The rbuildfiles operator is almost morally the inverse of the buildfiles operator. However, this moral inversion holds more strongly in one direction: the outputs of rbuildfiles are just like the inputs of buildfiles; the former will only contain BUILD file targets in packages, and the latter may contain such targets. In the other direction, the correspondence is weaker. The outputs of the buildfiles operator are targets corresponding to all packages and .bzl files needed by a given input. However, the inputs of the rbuildfiles operator are not those targets, but rather the path fragments that correspond to those targets.

Package definition files: loadfiles

expr ::= loadfiles(expr)

The loadfiles(x) operator returns the set of Starlark files that are needed to load the packages of each target in set x. In other words, for each package, it returns the .bzl files that are referenced from its BUILD files.

Output formats

bazel query generates a graph. You specify the content, format, and ordering by which bazel query presents this graph by means of the --output command-line option.

When running with Sky Query, only output formats that are compatible with unordered output are allowed. Specifically, graph, minrank, and maxrank output formats are forbidden.

Some of the output formats accept additional options. The name of each output option is prefixed with the output format to which it applies, so --graph:factored applies only when --output=graph is being used; it has no effect if an output format other than graph is used. Similarly, --xml:line_numbers applies only when --output=xml is being used.

On the ordering of results

Although query expressions always follow the "law of conservation of graph order", presenting the results may be done in either a dependency-ordered or unordered manner. This does not influence the targets in the result set or how the query is computed. It only affects how the results are printed to stdout. Moreover, nodes that are equivalent in the dependency order may or may not be ordered alphabetically. The --order_output flag can be used to control this behavior. (The --[no]order_results flag has a subset of the functionality of the --order_output flag and is deprecated.)

The default value of this flag is auto, which prints results in lexicographical order. However, when somepath(a,b) is used, the results will be printed in deps order instead.

When this flag is no and --output is one of build, label, label_kind, location, package, proto, or xml, the outputs will be printed in arbitrary order. This is generally the fastest option. It is not supported though when --output is one of graph, minrank or maxrank: with these formats, Bazel always prints results ordered by the dependency order or rank.

When this flag is deps, Bazel prints results in some topological order—that is, dependencies first. However, nodes that are unordered by the dependency order (because there is no path from either one to the other) may be printed in any order.

When this flag is full, Bazel prints nodes in a fully deterministic (total) order. First, all nodes are sorted alphabetically. Then, each node in the list is used as the start of a post-order depth-first search in which outgoing edges to unvisited nodes are traversed in alphabetical order of the successor nodes. Finally, nodes are printed in the reverse of the order in which they were visited.

Printing nodes in this order may be slower, so it should be used only when determinism is important.

Print the source form of targets as they would appear in BUILD

--output build

With this option, the representation of each target is as if it were hand-written in the BUILD language. All variables and function calls (such as glob, macros) are expanded, which is useful for seeing the effect of Starlark macros. Additionally, each effective rule reports a generator_name and/or generator_function) value, giving the name of the macro that was evaluated to produce the effective rule.

Although the output uses the same syntax as BUILD files, it is not guaranteed to produce a valid BUILD file.

--output label

With this option, the set of names (or labels) of each target in the resulting graph is printed, one label per line, in topological order (unless --noorder_results is specified, see notes on the ordering of results). (A topological ordering is one in which a graph node appears earlier than all of its successors.) Of course there are many possible topological orderings of a graph (reverse postorder is just one); which one is chosen is not specified.

When printing the output of a somepath query, the order in which the nodes are printed is the order of the path.

Caveat: in some corner cases, there may be two distinct targets with the same label; for example, a sh_binary rule and its sole (implicit) srcs file may both be called If the result of a query contains both of these targets, the output (in label format) will appear to contain a duplicate. When using the label_kind (see below) format, the distinction becomes clear: the two targets have the same name, but one has kind sh_binary rule and the other kind source file.

--output label_kind

Like label, this output format prints the labels of each target in the resulting graph, in topological order, but it additionally precedes the label by the kind of the target.

--output minrank --output maxrank

Like label, the minrank and maxrank output formats print the labels of each target in the resulting graph, but instead of appearing in topological order, they appear in rank order, preceded by their rank number. These are unaffected by the result ordering --[no]order_results flag (see notes on the ordering of results).

There are two variants of this format: minrank ranks each node by the length of the shortest path from a root node to it. "Root" nodes (those which have no incoming edges) are of rank 0, their successors are of rank 1, etc. (As always, edges point from a target to its prerequisites: the targets it depends upon.)

maxrank ranks each node by the length of the longest path from a root node to it. Again, "roots" have rank 0, all other nodes have a rank which is one greater than the maximum rank of all their predecessors.

All nodes in a cycle are considered of equal rank. (Most graphs are acyclic, but cycles do occur simply because BUILD files contain erroneous cycles.)

These output formats are useful for discovering how deep a graph is. If used for the result of a deps(x), rdeps(x), or allpaths query, then the rank number is equal to the length of the shortest (with minrank) or longest (with maxrank) path from x to a node in that rank. maxrank can be used to determine the longest sequence of build steps required to build a target.

For example, the graph on the left yields the outputs on the right when --output minrank and --output maxrank are specified, respectively.

Out ranked

      0 //c:c
      1 //b:b
      1 //a:a
      2 //
      2 //

      0 //c:c
      1 //b:b
      2 //a:a
      2 //
      3 //
--output location

Like label_kind, this option prints out, for each target in the result, the target's kind and label, but it is prefixed by a string describing the location of that target, as a filename and line number. The format resembles the output of grep. Thus, tools that can parse the latter (such as Emacs or vi) can also use the query output to step through a series of matches, allowing the Bazel query tool to be used as a dependency-graph-aware "grep for BUILD files".

The location information varies by target kind (see the kind operator). For rules, the location of the rule's declaration within the BUILD file is printed. For source files, the location of line 1 of the actual file is printed. For a generated file, the location of the rule that generates it is printed. (The query tool does not have sufficient information to find the actual location of the generated file, and in any case, it might not exist if a build has not yet been performed.)

--output package

This option prints the name of all packages to which some target in the result set belongs. The names are printed in lexicographical order; duplicates are excluded. Formally, this is a projection from the set of labels (package, target) onto packages.

Packages in external repositories are formatted as @repo//foo/bar while packages in the main repository are formatted as foo/bar.

In conjunction with the deps(...) query, this output option can be used to find the set of packages that must be checked out in order to build a given set of targets.

Display a graph of the result

--output graph

This option causes the query result to be printed as a directed graph in the popular AT&T GraphViz format. Typically the result is saved to a file, such as .png or .svg. (If the dot program is not installed on your workstation, you can install it using the command sudo apt-get install graphviz.) See the example section below for a sample invocation.

This output format is particularly useful for allpaths, deps, or rdeps queries, where the result includes a set of paths that cannot be easily visualized when rendered in a linear form, such as with --output label.

By default, the graph is rendered in a factored form. That is, topologically-equivalent nodes are merged together into a single node with multiple labels. This makes the graph more compact and readable, because typical result graphs contain highly repetitive patterns. For example, a java_library rule may depend on hundreds of Java source files all generated by the same genrule; in the factored graph, all these files are represented by a single node. This behavior may be disabled with the --nograph:factored option.

--graph:node_limit n

The option specifies the maximum length of the label string for a graph node in the output. Longer labels will be truncated; -1 disables truncation. Due to the factored form in which graphs are usually printed, the node labels may be very long. GraphViz cannot handle labels exceeding 1024 characters, which is the default value of this option. This option has no effect unless --output=graph is being used.


By default, graphs are displayed in factored form, as explained above. When --nograph:factored is specified, graphs are printed without factoring. This makes visualization using GraphViz impractical, but the simpler format may ease processing by other tools (such as grep). This option has no effect unless --output=graph is being used.


--output xml

This option causes the resulting targets to be printed in an XML form. The output starts with an XML header such as this

  <?xml version="1.0" encoding="UTF-8"?>
  <query version="2">

and then continues with an XML element for each target in the result graph, in topological order (unless unordered results are requested), and then finishes with a terminating


Simple entries are emitted for targets of file kind:

  <source-file name='//' .../>
  <generated-file name='//' .../>

But for rules, the XML is structured and contains definitions of all the attributes of the rule, including those whose value was not explicitly specified in the rule's BUILD file.

Additionally, the result includes rule-input and rule-output elements so that the topology of the dependency graph can be reconstructed without having to know that, for example, the elements of the srcs attribute are forward dependencies (prerequisites) and the contents of the outs attribute are backward dependencies (consumers).

rule-input elements for implicit dependencies are suppressed if --noimplicit_deps is specified.

  <rule class='cc_binary rule' name='//foo:foo' ...>
    <list name='srcs'>
      <label value='//'/>
      <label value='//'/>
    <list name='deps'>
      <label value='//common:common'/>
      <label value='//collections:collections'/>
    <list name='data'>
    <int name='linkstatic' value='0'/>
    <int name='linkshared' value='0'/>
    <list name='licenses'/>
    <list name='distribs'>
      <distribution value="INTERNAL" />
    <rule-input name="//common:common" />
    <rule-input name="//collections:collections" />
    <rule-input name="//" />
    <rule-input name="//" />

Every XML element for a target contains a name attribute, whose value is the target's label, and a location attribute, whose value is the target's location as printed by the --output location.


By default, the locations displayed in the XML output contain line numbers. When --noxml:line_numbers is specified, line numbers are not printed.


By default, XML output does not include rule attribute whose value is the default value for that kind of attribute (for example, if it were not specified in the BUILD file, or the default value was provided explicitly). This option causes such attribute values to be included in the XML output.

Regular expressions

Regular expressions in the query language use the Java regex library, so you can use the full syntax for java.util.regex.Pattern.

Querying with external repositories

If the build depends on rules from external repositories (defined in the WORKSPACE file) then query results will include these dependencies. For example, if //foo:bar depends on //external:some-lib and //external:some-lib is bound to @other-repo//baz:lib, then bazel query 'deps(//foo:bar)' will list both @other-repo//baz:lib and //external:some-lib as dependencies.

External repositories themselves are not dependencies of a build. That is, in the example above, //external:other-repo is not a dependency. It can be queried for as a member of the //external package, though, for example:

  # Querying over all members of //external returns the repository.
  bazel query 'kind(http_archive, //external:*)'

  # ...but the repository is not a dependency.
  bazel query 'kind(http_archive, deps(//foo:bar))'
  INFO: Empty results