Troubleshooting failures

This guide provides an overview of common scenarios encountered when troubleshooting Compatibility Test for Fuchsia (CTF) test failures and walks through some reasons for breakages and ways to unblock CL submission.

CTF tests assert on interactions between software frozen on a release branch and software built on the main branch of fuchsia.git. If those assertions fail due to an incompatibility, CL submission will be blocked.

The purpose of CTF is not to ban compatibility-breaking changes, rather it is meant to make such breakages visible so that they can be appropriately addressed.


In each scenario, a real compatibility issue is introduced which causes a CTF test to fail and block CL submission. The compatibility issue can be as simple as changing the return value of an existing FIDL protocol used by a test. We assume the following context common to each scenario (see motivation for more testing scenarios):

  • A FIDL client is frozen as a ctf_fuchsia_package on F19 called echo-service-tests. This connects to a FIDL protocol and asserts on the responses.

  • generate_ctf_tests.gni contains a rule template("generate_echo-service-tests") which merges the incoming client package with a server package built on main. (see the user guide for details).

  • The client calls a method on the server called Echo which takes as input a string and returns as output a string as follows:

    auto server_proxy = connect_to_named_protocol("my_protocol.EchoService");

Suppose we want to change the behavior of echo to instead return the lowercase representation of the string it is passed. We can update our test on main to do the following:


This passes for ctf_in_development tests, but when the frozen client with the old assertion is run against this new server, the test will fail:

FAILURE: "hello" != "Hello"

CL submission is now blocked, and the path forward depends on the reason for this breakage.

Unintentional breaking change - soft transition

In this scenario, we do not intend to break compatibility. This is the case if pre-built or out-of-tree components depend on the old behavior. Running those components against a platform containing the changes to Echo will result in unexpected behavior.

In this case, the safe thing to do is a soft transition to the new behavior:

  1. Introduce a new method with the new behavior:

    protocol Echo {
     // ...
     EchoLowercase(struct {input string}) -> (string);
  2. Mark the old method as deprecated or removed (optional):

    protocol Echo {
     Echo(/* ... */) -> (string);
     // ...
  3. Implement the new method in the server.

  4. Change fuchsia.git callers to use the new method instead of the old one.

The server must support both methods until all API levels in which the old method exists are no longer supported. In the above example, this will be when F19 is unsupported, since the old Echo method is removed in F20 (due to @available(removed=20)).

Intentional breaking change - change test in release branch

In this scenario, the compatibility breakage is intentional. This may arise for several reasons:

  • We want to change the behavior of an API, and we accept the risk of pre-built or out-of-tree clients breaking. This is a true positive failure, which we will explicitly acknowledge.

  • The behavior that changed is internal to a test, and does not represent a breaking change of the SDK surface itself. This is a false positive failure, because we caught a test incompatibility rather than an SDK incompatibility.

In either case, the solution is to modify the release branch so that the test will no longer fail when run against either the pre-change or post-change code on main.

The change can be made as follows:

  1. Check out the release branch:

    fx sync-to refs/heads/releases/f19
  2. Modify the assertion so it accepts both outputs (or comment it out):

     AnyOf(Eq("Hello"), Eq("hello"))
  3. Test that your changes will work on main (see below)

  4. Commit and push changes:

    git push origin HEAD:refs/for/releases/f19
  5. Get CL reviewed and submit.

  6. Land the blocked CL.

  7. Clean up the release branch to accept only the new behavior (optional).


You can test that your changes will work when applied to main. You need two checkouts of Fuchsia, one synced to the release branch and one synced to main. Do the following:

  1. On the release branch checkout, build a new CTF bundle.

    fx set core.x64 --with-tests //sdk/ctf
    fx build
  2. On the main checkout, build the CTF release tests.

    fx set core.x64 --with-tests //sdk/ctf/release:tests
    fx build
  3. In the output directory for the release branch, copy the built CTF bundle to the main repository.

    cp -fR \
  4. Rebuild the main checkout, repave device or restart emulator, and run the tests.

  5. Revert back to the version from CIPD once tests pass.

    jiri run-hooks

Example: SDK compatibility breakage (true positive)

This CL introduced a real compatibility breakage to fuchsia.ui.policy.MediaButtonsListener protocol. A new parameter was added to the MediaButtonsEvent with correct versioning annotations, however, the behavior if that parameter is left empty changed from the previous implementation. This incompatibility was caught by the CTF test for F19:

../../src/ui/tests/conformance_input_tests/ Failure
Expected equality of these values:
    Which is: "\n    volume: 0\n    mic_mute: 0\n    pause: 0\n    camera_disable: 0\n    power: 0"
    Which is: "\n    volume: 0\n    mic_mute: 0\n    pause: 0\n    camera_disable: 0\n    power: 1"

It was determined that this change is acceptable because there are not any pre-built clients of this protocol targeting API level 19.

CLs 1044994 and 1049612 were cherry-picked into the F19 release branch using the instructions above. Once the changes were rolled into the CTF release used on main, the original CL was submitted without modification.

Example: Test harness compatibility breakage (false positive)

This CL introduced a compatibility breakage to the fuchsia.tracing.controller.Controller protocol. The StopTracing method was made flexible, and changing a method from strict to flexible is not ABI-safe.

The WLAN hw-sim CTF test crashes under this change because it asserts that tracing successfully stops during the test, even though this is unrelated to the tested WLAN protocols and is not needed for correctness. This is a false positive compatibility failure because it affects only the test harness itself.

This CL was landed to turn tracing failures into warnings rather than fatal assertions, and this change was cherry-picked to the F18 release branch. Once the change was rolled into the CTF release used on main, the original CL was submitted without modification.

Intentionally drop support for an API level for specific tests

In this scenario, we want to explicitly drop support for an API level on a per-test basis. While version_history.json provides the canonical listing of supported API levels, there are reasons we would want to drop compatibility guarantees for specific protocols on a specific API level. For example, a major refactor of a non-critical subsystem where breaking old clients is acceptable.

In this case, we can take modify the thawing process to skip tests targeting the API level we want to drop:

  1. Modify generate_ctf_tests.gni on the main branch to conditionally skip thawing the test.

    template("generate_my-service-tests") {
     forward_variables_from(invoker, [ "test_info" ])
     if (defined(invoker.api_level) && invoker.api_level == "15") {
       # Do not thaw this test for F15
       not_needed([ "test_info" ])
       group(target_name) {
     } else {
       // ...
  2. Commit and submit this CL.

The above example skips thawing the entire artifact at F15. If you do not want to skip all tests, see the previous section for how to skip individual test cases on the release branch.

Example: Major refactor to diagnostics subsystem

This CL deleted support for the DirectoryReady component event, which is no longer in use. Its primary use was to support obtaining diagnostics data from components (which is now published using the fuchsia.inspect.InspectSink) protocol.

Unfortunately, components built at F15 still published their diagnostics using DirectoryReady, and all diagnostics CTF tests of this behavior would fail on main following its removal.

Fortunately, there are no pre-built components targeting F15 for which we need to obtain diagnostics data. A CL was submitted to disable all diagnostics CTF tests originating in F15.