5.1 KiB
(lifecycle-fuzzing)=
Fuzzing
Snapshot integrity errors are very problematic when they occur and can be hard to spot and prevent. To this end, a subset of the lifecycle test suite uses a combination of fuzzing and property-based testing via the Rapid Go library to randomly generate snapshots and programs to see whether or not it is possible to trigger a snapshot integrity error.
While snapshot integrity issues often happen as part of a chain of snapshot operations (e.g. the execution of multiple steps in a deployment), the precursor to any error state will always be a valid snapshot. Thus, rather than having to generate random chains of operations, we can instead simplify the problem to generating valid starting snapshots and then executing a single random operation on them. The strategy we employ is thus as follows:
-
Generate a snapshot (snapshot.go) consisting of a random set of resources (resource.go), including appropriate providers. Resources may randomly depend on each other, and may have random properties, such as whether they are custom resources or components, pending replacement, and so on.
-
Generate a program (program.go) from the previously generated snapshot. The program may choose to register any subset (including none) of the resources in the snapshot, as well as any set of new resources before, in between and after those specified in the snapshot. Resources from the snapshot that are registered may be copied as-is or registered with different properties.
-
Generate a set of provider implementations for the program (provider.go). Provider operations such as , , etc. may be configured to fail randomly, or return one of a set of random results (e.g. an update vs a replace for
Diff), on a per-resource basis. -
Generate an operation (one of
preview,up,refreshanddestroy) and associated configuration (such as a list of--targets), known in the test suite as a plan to execute (plan.go). -
Combine the snapshot, program, providers and plan to form a fixture (fixture.go) and execute it. If the operation yields a valid snapshot, the test passes, whether the operation completes successfully or not. If an invalid snapshot is produced, the test fails and the reproducing combination of snapshot, program, providers and plan is returned for debugging.
In the event that a failing test case is found, the reproducing fixture will be pretty printed to the screen and code for a reproducing test case will be written to a file to aid in debugging. See fixture.go and reprogen.go for more details.
:::{note}
By default, reproduction test files will be written to a temporary directory. A
specific directory can be configured using the
PULUMI_LIFECYCLE_TEST_FUZZING_REPRO_DIR environment variable.
:::
Running a suite of fuzz tests
The TestFuzz test generates and tests a set of fixtures, and can be run with
the top-level test_lifecycle_fuzz target:
make test_lifecycle_fuzz
You can change how many checks Rapid performs by setting the
LIFECYCLE_TEST_FUZZ_CHECKS Make variable on the command line. For instance, to
perform 10,000 checks:
make test_lifecycle_fuzz LIFECYCLE_TEST_FUZZ_CHECKS=10000
If you use go test directly, make sure you set the
PULUMI_LIFECYCLE_TEST_FUZZ variable to something; the test will be skipped if
you do not:
(cd pkg/engine/lifecycletest; PULUMI_LIFECYCLE_TEST_FUZZ=1 go test ./... -run '^TestFuzz$')
Starting from a known state
Rather than generating entirely random scenarios, it can be useful to start from
a known snapshot, and to fuzz potential provider configurations and operations,
etc. from there. The TestFuzzFromStateFile test is provided to this end. It
will read state from a JSON file (such as that produced by a pulumi stack export command) and use this as the starting point for a fuzzing run. Use the
PULUMI_LIFECYCLE_TEST_FUZZ_FROM_STATE_FILE environment variable to specify the
file to read (if this variable is not set, the test will be skipped):
PULUMI_LIFECYCLE_TEST_FUZZ_FROM_STATE_FILE=/path/to/state.json \
make test_lifecycle_fuzz_from_state_file \
LIFECYCLE_TEST_FUZZ_CHECKS=10000
In this example we have again configured 10,000 Rapid checks using the
LIFECYCLE_TEST_FUZZ_CHECKS Make variable as before.