Flaky CI: MTP/TRX host fails when suite filters select zero tests

[ReubenBond]

Description

Orleans.Analyzers.Tests intermittently causes a filtered Microsoft.Testing.Platform CI job to fail when the active suite filter selects zero tests from that assembly. The suite continues and reports no failed tests, but the assembly exits unsuccessfully during test-host shutdown.

This has reproduced on unrelated pull requests and different operating systems/frameworks.

Occurrences

PR #10746

• Run/job: https://github.com/dotnet/orleans/actions/runs/32499721406/job/96826364553
• Environment: Functional, Windows, net8.0
• Head: d1fe65475179923ba00ba6941fc077305151851a
• PR scope: documentation only

Orleans.Analyzers.Tests.dll (net8.0|x64) Zero tests ran (28s 485ms)
Exit code: 7
Standard output: Waiting 10 seconds for foreground threads to exit...
Error output: [FATAL ERROR] Foreground threads were left running, forcing process exit

The overall run reported 501 total tests, 482 succeeded, 19 skipped, 0 failed, and 1 test-application error.

PR #10739

• Run: https://github.com/dotnet/orleans/actions/runs/32497565314
• Environment: SlowBVT, Ubuntu, net10.0
• PR scope: journaling recovery behavior and tests

Orleans.Analyzers.Tests.dll (net10.0|x64) Zero tests ran (37s 377ms)
Exit code: 134
Unhandled exception. System.TimeoutException: The operation has timed out.
   at Microsoft.Testing.Platform.Helpers.TaskExtensions.TimeoutAfterAsync(...)
   at Microsoft.Testing.Extensions.TrxReport.Abstractions.TrxProcessLifetimeHandler...

Expected behavior

A test application containing no tests which match the suite filter exits successfully and releases all MTP/TRX resources without delaying or failing the aggregate CI job.

Notes

The failures occur in the analyzer test assembly under suite filters which select no analyzer tests. The different shutdown symptoms may share test-host or TRX lifetime cleanup behavior. The failed PR #10746 job is being rerun to confirm the intermittent nature.

[ReubenBond]

Investigation update:

• The original Ubuntu run shows this is broader than Orleans.Analyzers.Tests: zero-selection hosts for Orleans.DependencyInjection.Tests and Orleans.Connections.Security.Tests failed with the same TrxProcessLifetimeHandler timeout in other matrix jobs.
• The Windows symptom (Exit code: 7, foreground threads left running) and Linux symptom (Exit code: 134, named-pipe connection timeout in BeforeTestHostProcessStartCore) both occur in the out-of-process TRX lifetime path enabled by --report-trx together with --crashdump.
• Direct zero-selection runs of Orleans.Analyzers.Tests completed host shutdown normally on both net8.0 and net10.0. The failed CI jobs also passed when rerun unchanged, confirming an intermittent orchestration/lifetime race rather than analyzer test behavior.
• Orleans already pins the current stable MTP/TRX/CrashDump release (2.3.3). A repository-local change which removes TRX or crash-dump collection would trade the flake for weaker failure diagnostics, so I have not applied that workaround.

Upstream tracking with full environment, command line, and logs: microsoft/testfx#10667.

The actionable next step is to consume an upstream MTP fix once available. Until then, rerunning the affected matrix job remains the behavior-preserving mitigation.

[ReubenBond]

Investigation findings:

• No open or merged Orleans PR currently fixes Flaky CI: MTP/TRX host fails when suite filters select zero tests #10754. PR build(test): migrate from VSTest to Microsoft.Testing.Platform #10726 introduced the MTP/xUnit v3 runner configuration which exposed the behavior.
• The Ubuntu/net10 failure was tied to the earlier inline coverage path. PR ci: report code coverage on pull requests #10733's final implementation now collects coverage only from modules identified by successful prior TRX results, so Orleans.Analyzers.Tests is not launched for Functional/SlowBVT coverage when it has no selected tests.
• The Windows/net8 occurrence did not use coverage. It failed after xUnit's 10-second foreground-thread shutdown watchdog while MTP/TRX was still completing process lifetime work. The unchanged rerun passed.
• The normal zero-test exit code (8) is already ignored in test/Directory.Build.props. Exit codes 7 and 134 indicate abnormal host shutdown and should continue to fail; retrying or ignoring them would hide genuine lifecycle defects.
• Isolated reproductions completed successfully on Windows net8.0/net10.0 and in a Linux SDK 10.0.400 container with the same no-match filter plus TRX/crash/hang options. The failure appears to require solution-scale timing or resource pressure.
• Microsoft.Testing.Platform 2.3.3 is the latest available 2.x package. xUnit 4.0 adds a configurable foreground-thread wait, but a timeout increase does not fix the TRX named-pipe timeout and the major upgrade requires broader test-suite migration.
• I inspected five failed and seventeen successful CI runs after ci: report code coverage on pull requests #10733 merged. None contained the exit 7/134 or TrxProcessLifetimeHandler signature; zero-selection analyzer hosts in unrelated provider jobs completed in roughly 0.4-0.7 seconds.

The MTP/TRX lifecycle defect is now tracked upstream at microsoft/testfx#10668 with both CI signatures and version details.

No Orleans patch is proposed yet because the available local workarounds either remove crash diagnostics, suppress abnormal exits, or require brittle project-selection duplication. The next recurrence should preserve MTP diagnostic logs and dump artifacts so the upstream named-pipe/foreground-thread race can be correlated without weakening CI failure semantics.

[ReubenBond]

Current status:

• The root cause is confirmed upstream and fixed by Fix TRX lifetime handshake race microsoft/testfx#10682. The out-of-process TRX lifetime handler previously created its named-pipe server inside a scheduled task, allowing a fast zero-selection test host to attempt the connection before the pipe existed under thread-pool contention. The fix creates and configures the server synchronously before scheduling the connection wait.
• Upstream added deterministic unit coverage and an acceptance test covering zero selected tests with --crashdump, --report-trx, and --ignore-exit-code 8 on net8.0 and net10.0. MTP/TRX lifetime handshake intermittently times out for zero-selection test applications microsoft/testfx#10667 and perf(serialization): grow memory stream buffers geometrically #10668 are closed as fixed.
• The fix is assigned to Microsoft.Testing.Platform 2.4.0. Orleans' configured public feeds still expose 2.3.3 as the newest version; an exact restore of 2.4.0-preview.26424.16 from dotnet-public fails because that preview is not published there.
• No Orleans pull request currently adopts the fix, and current main still pins the MTP/TRX/CrashDump/HangDump package set to 2.3.3.
• The latest 50 Orleans CI runs contain no recurrence of the zero-test TRX/foreground-thread signature. A recent exit-code-7 Cosmos job was a real hang with ten failed tests and a collected dump, so retaining abnormal host exits and crash diagnostics continues to work as intended.

The behavior-preserving resolution is to update the four MTP packages together when 2.4.0 is published to the repository's public feed. A local workaround or suppression would weaken host-failure detection or crash diagnostics and is not warranted.