Test encyclopedia

An Exhaustive Specification of the Test Execution Environment


The Bazel BUILD language includes rules which can be used to define automated test programs in many languages.

Tests are run using bazel test. Users may also execute test binaries directly. This is allowed but not endorsed, as such an invocation will not adhere to the mandates described below.

Tests should be hermetic: that is, they ought to access only those resources on which they have a declared dependency. If tests are not properly hermetic then they do not give historically reproducible results. This could be a significant problem for culprit finding (determining which change broke a test), release engineering auditability, and resource isolation of tests (automated testing frameworks ought not DDOS a server because some tests happen to talk to it).


The goal of this document is to formally establish the runtime environment for and expected behavior of Bazel tests. It will also impose requirements on the test runner and the build system. Our intent is to help test authors avoid relying on unspecified behavior, and thus give the testing infrastructure more freedom to make implementation changes. We will also take the opportunity to tighten up some holes which currently allow many tests to pass despite not being properly hermetic, deterministic, and reentrant.

This document is intended to be both normative and authoritative. If this specification and the implemented behavior of test runner disagree, the specification takes precedence.

Purpose of tests

The purpose of Bazel tests is to confirm some property of the source files checked into the repository. (In this document, "source files" includes test data, golden outputs, and anything else kept under version control.) One user writes a test to assert an invariant which they expect to be maintained. Other users execute the test later to check whether the invariant has been broken. If the test depends on any variables other than source files (non-hermetic), its value is diminished, because the later users cannot be sure their changes are at fault when the test stops passing.

Therefore the outcome of a test must depend only on:

  • source files on which the test has a declared dependency
  • products of the build system on which the test has a declared dependency
  • resources whose behavior is guaranteed by the test runner to remain constant

Currently, such behavior is not enforced. However, test runners reserve the right to add such enforcement in the future.

Role of the build system

Test rules are analogous to binary rules in that each must yield an executable program. For some languages, this is a stub program which combines a language-specific harness with the test code. Test rules must produce other outputs as well. In addition to the primary test executable, the test runner will need a manifest of runfiles, input files which should be made available to the test at runtime, and it may need information about the type, size, and tags of a test.

The build system may use the runfiles to deliver code as well as data. (This might be used as an optimization to make each test binary smaller by sharing files across tests, e.g. through the use of dynamic linking.) The build system should ensure that the generated executable loads these files via the runfiles image provided by the test runner, rather than hardcoded references to absolute locations in the source or output tree.

Role of the test runner

From the point of view of the test runner, each test is a program which can be invoked with execve(). There may be other ways to execute tests; for example, an IDE might allow the execution of Java tests in-process. However, the result of running the test as a standalone process must be considered authoritative. If a test process runs to completion and terminates normally with an exit code of zero, the test has passed. Any other result is considered a test failure. In particular, writing any of the strings PASS or FAIL to stdout has no significance to the test runner.

If a test takes too long to execute, exceeds some resource limit, or the test runner otherwise detects prohibited behavior, it may choose to kill the test and treat the run as a failure. The runner must not report the test as passing after sending a signal to the test process or any children thereof.

The whole test target (not individual methods or tests) is given a limited amount of time to run to completion. The time limit for a test is based on its timeout attribute according to the following table:

timeout Time Limit (sec.)

Tests which do not explicitly specify a timeout have one implied based on the test's size as follows:

size Implied timeout label

For example a "large" test with no explicit timeout setting will be allotted 900 seconds to run. A "medium" test with a timeout of "short" will be allotted 60 seconds.

Unlike timeout, the size also determines the limits for other resources like RAM, as described in Common definitions.

All combinations of size and timeout labels are legal, so an "enormous" test may be declared to have a timeout of "short". Presumably it would do some really horrible things very quickly.

Tests may return arbitrarily fast regardless of timeout. A test is not penalized for an overgenerous timeout, although a warning may be issued: you should generally set your timeout as tight as you can without incurring any flakiness.

The test timeout can be overridden with the --test_timeout bazel flag, e.g. for manually running under conditions which are known to be slow. The --test_timeout values are in seconds. For example, --test_timeout=120 will set the test timeout to two minutes.

There is also a recommended lower bound for test timeouts as follows:

timeoutTime minimum (sec.)

For example, if a "moderate" test completes in 5.5s, consider setting timeout = "short" or size = "small". Using the bazel --test_verbose_timeout_warnings command line option will show the tests whose specified size is too big.

Test sizes and timeouts are specified in the BUILD file according to the specification here. If unspecified, a test's size will default to "medium".

If the main process of a test exits, but some of its children are still running, the test runner should consider the run complete and count it as a success or failure based on the exit code observed from the main process. The test runner may kill any stray processes. Tests should not leak processes in this fashion.

Test sharding

Tests can be parallelized via test sharding. See --test_sharding_strategy and shard_count to enable test sharding. When sharding is enabled, the test runner is launched once per shard. The environment variable TEST_TOTAL_SHARDS is the number of shards, and TEST_SHARD_INDEX is the shard index, beginning at 0. Runners use this information to select which tests to run - for example, using a round-robin strategy. Not all test runners support sharding. If a runner supports sharding, it must create or update the last modified date of the file specified by TEST_SHARD_STATUS_FILE. Otherwise, Bazel assumes it does not support sharding and will not launch additional runners.

Initial conditions

When executing a test, the test runner must establish certain initial conditions.

The test runner must invoke each test with the path to the test executable in argv[0]. This path must be relative and beneath the test's current directory (which is in the runfiles tree, see below). The test runner should not pass any other arguments to a test unless the user explicitly requests it.

The initial environment block shall be composed as follows:

HOMEvalue of $TEST_TMPDIRrecommended
LD_LIBRARY_PATHcolon-separated list of directories containing shared librariesoptional
JAVA_RUNFILESvalue of $TEST_SRCDIRdeprecated
LOGNAMEvalue of $USERrequired
TEST_INFRASTRUCTURE_FAILURE_FILEabsolute path to a private file in a writable directory (This file should only be used to report failures originating from the testing infrastructure, not as a general mechanism for reporting flaky failures of tests. In this context, testing infrastructure is defined as systems or libraries that are not test-specific, but can cause test failures by malfunctioning. The first line is the name of the testing infrastructure component that caused the failure, the second one a human-readable description of the failure. Additional lines are ignored.)optional
TEST_LOGSPLITTER_OUTPUT_FILEabsolute path to a private file in a writable directory (used to write Logsplitter protobuffer log)optional
TEST_PREMATURE_EXIT_FILEabsolute path to a private file in a writable directory (used for catching calls to exit())optional
TEST_RANDOM_SEED If the --runs_per_test option is used, TEST_RANDOM_SEED is set to the run number (starting with 1) for each individual test run. optional
TEST_TARGETThe name of the target being testedoptional
TEST_SIZEThe test sizeoptional
TEST_TIMEOUTThe test timeout in secondsoptional
TEST_SHARD_INDEXshard index, if sharding is usedoptional
TEST_SHARD_STATUS_FILEpath to file to touch to indicate support for shardingoptional
TEST_SRCDIRabsolute path to the base of the runfiles treerequired
TEST_TOTAL_SHARDStotal shard count, if sharding is usedoptional
TEST_TMPDIRabsolute path to a private writable directoryrequired
TEST_WORKSPACEthe local repository's workspace nameoptional
TEST_UNDECLARED_OUTPUTS_DIRabsolute path to a private writable directory (used to write undeclared test outputs)optional
TEST_UNDECLARED_OUTPUTS_ANNOTATIONS_DIRabsolute path to a private writable directory (used to write undeclared test output annotation .part files). optional
TEST_WARNINGS_OUTPUT_FILEabsolute path to a private file in a writable directory (used to write test target warnings)optional
TESTBRIDGE_TEST_ONLYvalue of --test_filter, if specifiedoptional
USERvalue of getpwuid(getuid())->pw_namerequired
XML_OUTPUT_FILE Location of the test result XML output file. The XML schema is based on the JUnit test result schema. optional

The environment may contain additional entries. Tests should not depend on the presence, absence, or value of any environment variable not listed above.

The initial working directory shall be $TEST_SRCDIR/$TEST_WORKSPACE.

The current process id, process group id, session id, and parent process id are unspecified. The process may or may not be a process group leader or a session leader. The process may or may not have a controlling terminal. The process may have zero or more running or unreaped child processes. The process should not have multiple threads when the test code gains control.

File descriptor 0 (stdin) shall be open for reading, but what it is attached to is unspecified. Tests must not read from it. File descriptors 1 (stdout) and 2 (stderr) shall be open for writing, but what they are attached to is unspecified. It could be a terminal, a pipe, a regular file, or anything else to which characters can be written. They may share an entry in the open file table (meaning that they cannot seek independently). Tests should not inherit any other open file descriptors.

The initial umask shall be 022 or 027.

No alarm or interval timer shall be pending.

The initial mask of blocked signals shall be empty. All signals shall be set to their default action.

The initial resource limits, both soft and hard, should be set as follows:

RLIMIT_NOFILEat least 1024
RLIMIT_STACKunlimited, or 2044KB <= rlim <= 8192KB

The initial process times (as returned by times()) and resource utilization (as returned by getrusage()) are unspecified.

The initial scheduling policy and priority are unspecified.

Role of the host system

In addition to the aspects of user context under direct control of the test runner, the operating system on which tests execute must satisfy certain properties for a test run to be valid.


The root directory observed by a test may or may not be the real root directory.
/proc shall be mounted.
All build tools shall be present at the absolute paths under /usr used by a local installation.
Paths starting with /home may not be available. Tests should not access any such paths.
/tmp shall be writable, but tests should avoid using these paths.
Tests must not assume that any constant path is available for their exclusive use.
Tests must not assume that atimes are enabled for any mounted filesystem.

Users and groups

The users root, nobody, and unittest must exist. The groups root, nobody, and eng must exist.

Tests must be executed as a non-root user. The real and effective user ids must be equal; likewise for group ids. Beyond this, the current user id, group id, user name, and group name are unspecified. The set of supplementary group ids is unspecified.

The current user id and group id must have corresponding names which can be retrieved with getpwuid() and getgrgid(). The same may not be true for supplementary group ids.

The current user must have a home directory. It may not be writable. Tests must not attempt to write to it.


The hostname is unspecified. It may or may not contain a dot. Resolving the hostname must give an IP address of the current host. Resolving the hostname cut after the first dot must also work. The hostname localhost must resolve.

Other resources

Tests are granted at least one CPU core. Others may be available but this is not guaranteed. Other performance aspects of this core are not specified. You can increase the reservation to a higher number of CPU cores by adding the tag "cpu:n" (where n is a positive number) to a test rule. If a machine has less total CPU cores than requested, Bazel will still run the test. If a test uses sharding, each individual shard will reserve the number of CPU cores specified here.

Tests may create subprocesses, but not process groups or sessions.

There is a limit on the number of input files a test may consume. This limit is subject to change, but is currently in the range of tens of thousands of inputs.

Time and date

The current time and date are unspecified. The system timezone is unspecified.

X Windows may or may not be available. Tests that need an X server should start Xvfb.

Test interaction with the filesystem

All file paths specified in test environment variables point to somewhere on the local filesystem, unless otherwise specified.

Tests should create files only within the directories specified by $TEST_TMPDIR and $TEST_UNDECLARED_OUTPUTS_DIR (if set).
These directories will be initially empty.
Tests must not attempt to remove, chmod, or otherwise alter these directories.
These directories may be a symbolic links.
The filesystem type of $TEST_TMPDIR/. remains unspecified.
Tests may also write .part files to the $TEST_UNDECLARED_OUTPUTS_ANNOTATIONS_DIR to annotate undeclared output files.

In rare cases, a test may be forced to create files in /tmp. For example, path length limits for Unix domain sockets typically require creating the socket under /tmp. Bazel will be unable to track such files; the test itself must take care to be hermetic, to use unique paths to avoid colliding with other, simultaneously running tests and non-test processes, and to clean up the files it creates in /tmp.

Tests must access inputs through the runfiles mechanism, or other parts of the execution environment which are specifically intended to make input files available. Tests must not access other outputs of the build system at paths inferred from the location of their own executable.

It is unspecified whether the runfiles tree contains regular files, symbolic links, or a mixture. The runfiles tree may contain symlinks to directories. Tests should avoid using paths containing .. components within the runfiles tree.

No directory, file, or symlink within the runfiles tree (including paths which traverse symlinks) should be writable. (It follows that the initial working directory should not be writable.) Tests must not assume that any part of the runfiles is writable, or owned by the current user (i.e. chmod and chgrp may fail).

The runfiles tree (including paths which traverse symlinks) must not change during test execution. Parent directories and filesystem mounts must not change in any way which affects the result of resolving a path within the runfiles tree.

In order to catch early exit, a test may create a file at the path specified by TEST_PREMATURE_EXIT_FILE upon start and remove it upon exit. If Bazel sees the file when the test finishes, it will assume that the test exited prematurely and mark it as having failed.

Tag conventions

Some tags in the test rules have a special meaning. See also the Bazel Build Encyclopedia on the tags attribute.

exclusive run no other test at the same time
external test has an external dependency; disable test caching
large test_suite convention; suite of large tests
manual* don't include test target in wildcard target patterns like :..., :*, or :all)
medium test_suite convention; suite of medium tests
small test_suite convention; suite of small tests
smoke test_suite convention; means it should be run before committing code changes into the version control system
*Note: bazel query does not respect the manual tag.


In the following, assume there is a *_binary() rule labeled //foo/bar:unittest, with a run-time dependency on the rule labeled //deps/server:server.


The runfiles directory for a target //foo/bar:unittest is the directory $(WORKSPACE)/$(BINDIR)/foo/bar/unittest.runfiles. This path is referred to as the runfiles_dir.


The runfiles directory is declared as a compile-time dependency of the *_binary() rule. The runfiles directory itself depends on the set of BUILD files that affect the *_binary() rule or any of its compile-time or run-time dependencies. Modifying source files does not affect the structure of the runfiles directory, and thus does not trigger any rebuilding.


The runfiles directory contains the following:

  • Symlinks to run-time dependencies: each OutputFile and CommandRule that is a run-time dependency of the *_binary() rule is represented by one symlink in the runfiles directory. The name of the symlink is $(WORKSPACE)/package_name/rule_name. For example, the symlink for server would be named $(WORKSPACE)/deps/server/server, and the full path would be $(WORKSPACE)/foo/bar/unittest.runfiles/$(WORKSPACE)/deps/server/server. The destination of the symlink is the OutputFileName() of the OutputFile or CommandRule, expressed as an absolute path. Thus, the destination of the symlink might be $(WORKSPACE)/linux-dbg/deps/server/42/server.
  • Symlinks to sub-runfiles: for every *_binary() Z that is a run-time depdendency of *_binary() C, there is a second link in the runfiles directory of C to the runfiles of Z. The name of the symlink is $(WORKSPACE)/package_name/rule_name.runfiles. The target of the symlink is the runfiles directory. I.e. all subprograms share a common runfiles directory.