Fix Kafka AppHost health-check producer lifetime

[David Fowler (davidfowl)]

Description

Kafka AppHost health checks currently create a new producer on every execution without disposing it, accumulating polling threads over time. Register each resource's KafkaHealthCheck as a DI-owned keyed singleton so checks reuse their producer and AppHost disposal releases it.

This preserves independent configuration for multiple Kafka resources and deferred connection-string initialization, following the existing client integration pattern. No public API, dependency, or producer-default changes are required.

Fixes: #20091

Validation

• Added public-API regression coverage for DI ownership, reuse across scopes, resource isolation, and connection-string availability. No private reflection.
• All 28 selected Kafka hosting tests passed, including broker-backed health and WaitFor readiness tests.
• Ran standalone AppHosts with two real Kafka 8.2.0 brokers each, comparing released 13.5.3 with the patched source. Both used Confluent.Kafka 2.15.0, normal resource monitoring, and 40 additional rounds through HealthCheckService. All 80 additional probe results were healthy.

Live measurement	Released 13.5.3	Patched
Total health-check executions	84	84
Distinct health-check instances	84	2
Producer polling threads after the checks	84	2
Producer polling threads after AppHost disposal	84	0

Polling threads were identified using dotnet-stack while the process remained alive after AppHost disposal. The standalone verification used no private reflection. Detailed evidence is attached to #20091. This verifies the lifetime leak, not the original report's exact CPU percentage.

Checklist

• Is this feature complete?
  • Yes. Ready to ship.
  • No. Follow-up changes expected.
• Are you including unit tests for the changes and scenario tests if relevant?
  • Yes
  • No
• Did you add public API?
  • Yes
    • If yes, did you have an API Review for it?
      • Yes
      • No
    • Did you add <remarks /> and <code /> elements on your triple slash comments?
      • Yes
      • No
  • No
• Does the change make any security assumptions or guarantees?
  • Yes
    • If yes, have you done a threat model and had a security review?
      • Yes
      • No
  • No

[github-actions]

🚀 Dogfood this PR with:

⚠️ WARNING: Do not do this without first carefully reviewing the code of this PR to satisfy yourself it is safe.

curl -fsSL https://raw.githubusercontent.com/microsoft/aspire/main/eng/scripts/get-aspire-cli-pr.sh | bash -s -- 20092

Or

• Run remotely in PowerShell:

iex "& { $(irm https://raw.githubusercontent.com/microsoft/aspire/main/eng/scripts/get-aspire-cli-pr.ps1) } 20092"

[Copilot]

Copilot review overview

Review tier: Balanced
Findings: None

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

[github-actions]

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

[github-actions]

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

[David Fowler (davidfowl)]

PR Testing Report

PR Information

• PR: Fix Kafka AppHost health-check producer lifetime #20092 - Fix Kafka AppHost health-check producer lifetime
• Head commit: 15c2c650290d5e6bd9b702f8b7c7a197b615ce4b
• Tested at: September 13, 2026, approximately 15:05-15:09 UTC
• Environment: macOS ARM64, Docker Engine 28.5.1, local isolated PR CLI install, two real Kafka 8.2.0 containers.

Artifact Version Verification

• Installed CLI: 13.6.0-pr.20092.g15c2c650, matching the PR head.
• Kafka package: Aspire.Hosting.Kafka.13.6.0-pr.20092.g15c2c650.nupkg from the downloaded PR hive.
• Package repository commit: 4848901472371aa592bec6e697abadea5a30cd06. This is the CI merge commit with parents 47a04482488cd63c5b844091c154fcdc018dc718 and the PR head 15c2c650290d5e6bd9b702f8b7c7a197b615ce4b, not a stale package.
• Artifact-producing run: https://github.com/microsoft/aspire/actions/runs/34762863086
• Status: Verified. The PR head was checked again after testing and had not changed.

Changes Analyzed

• src/Aspire.Hosting.Kafka/KafkaBuilderExtensions.cs: registers a DI-owned keyed singleton per Kafka resource and resolves it from the health-check factory.
• tests/Aspire.Hosting.Kafka.Tests/AddKafkaTests.cs: public-API coverage for ownership, scoped reuse, distinct resources, and deferred connection availability.
• Categories: Kafka hosting integration and tests only. No CLI, dashboard, template, client integration, extension, or CI-infrastructure changes.

Test Scenario Executed

Fresh packaged AppHost lifecycle, including outage and recovery

Objective: Validate the shipped PR artifacts, real resource orchestration, dashboard state, health-check reuse, and disposal without private reflection.

Status: Passed.

Created a fresh file-based C# AppHost using the installed PR CLI and the explicit PR template hive:

"$cli" new aspire-empty --name KafkaPrSmoke --output "$temp/KafkaPrSmoke" \
  --source "$hive" --version 13.6.0-pr.20092.g15c2c650 \
  --language csharp --localhost-tld false --suppress-agent-init --non-interactive

Added the PR-version Kafka package and two AddKafka resources. Started the app with:

"$cli" start --apphost "$apphost" --launch-profile http --isolated --format Json --non-interactive

The temporary harness observed public health-check factory results and captured managed stacks using dotnet-stack. It did not replace the checks, inspect private fields, or change producer configuration. It set the public health-check registration timeout to three seconds to bound unavailable-broker probes. Normal Aspire health monitoring remained enabled.

All phases deliberately used the same fresh AppHost so reuse could be verified across failure and recovery:

Phase	Coverage	Expected and observed result
Before connection availability	Boundary	Resolving each check throws InvalidOperationException("Connection string is unavailable"). AppHost subsequently starts successfully.
Startup and dashboard	Happy path	Both real Kafka containers reach Running/Healthy. Dashboard displays both resources.
Repeated healthy probes	Happy path	40 rounds, checking both brokers each time, all return Healthy. Exactly one check instance per resource and two producer polling threads.
Stop broker A	Unhappy path	aspire resource kafka-a stop moves A to Exited; B remains Running/Healthy. Dashboard confirms the isolated failure state.
Repeated unavailable-broker probes	Unhappy path	Four rounds consistently return A=Unhealthy and B=Healthy. Still one instance per resource and two polling threads.
Restart and recovery	Recovery	aspire resource kafka-a start, then aspire wait kafka-a --status healthy --timeout 180, succeeds in about 5.7 seconds. Ten further rounds return Healthy for both resources without replacing the checks. Original endpoint allocations are retained.
AppHost disposal	Lifecycle	StopAsync and DisposeAsync remove both containers. While the console process is still alive, polling threads drop from two to zero.

All CLI follow-up commands supplied the explicit --apphost path.

Lifetime observations

Counts include normal background monitoring as well as explicit probe rounds.

Checkpoint	Factory executions	Distinct checks	Kafka producer polling threads
Ready	3	2	2
After 40 healthy rounds	87	2	2
After four outage rounds	98	2	2
After ten recovery rounds	122	2	2
After AppHost disposal	124	2	0

The captured thread stacks contain Confluent.Kafka.Producer<,>...<StartPollTask>b__0. These are thread-lifetime observations, not a reproduction of the original report's exact CPU percentage.

Evidence

Preserved under the session's files/pr20092-evidence/:

• apphost.cs: standalone packaged-PR harness.
• apphost-run.log: sanitized phase/probe output, including PRTEST PASS all scenarios.
• observations.json and checkpoint *.done files: exact counts.
• ready.stacks.txt, healthy-probes.stacks.txt, outage-probes.stacks.txt, recovery-probes.stacks.txt, disposed.stacks.txt: live managed stack captures.
• dashboard-healthy.png, dashboard-outage.png, dashboard-recovered.png: dashboard screenshots.
• dashboard-outage.txt, dashboard-recovered.txt: browser accessibility snapshots.

Installation, template creation, broker lifecycle, and aspire describe logs are retained alongside this folder as pr20092-*.log. Temporary login-bearing browser/startup logs were removed.

Summary

Validation	Status
CLI and Kafka package provenance	Passed
Fresh PR-template project and real broker startup	Passed
Missing connection-string boundary	Passed
Repeated healthy probes and independent resource instances	Passed
Unavailable broker isolation	Passed
Broker recovery using the same checks	Passed
Producer polling-thread cleanup on AppHost disposal	Passed

Overall Result

PR Kafka behavior verified against the packaged artifacts.

At the end of this test run, CI had failed VS Code browser-debugger E2E jobs on Linux and Windows, with the aggregate test/final checks consequently failed. Those jobs are outside the changed Kafka area and were not diagnosed by this focused run; their cause is not established here.

Cleanup

The AppHost and detached CLI processes exited; both owned broker containers were removed; the isolated headless browser was closed. The temporary project and PR installation were deleted. Evidence and the harness were retained in session artifacts. The repository working tree remained clean.

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused implementation follows the established Kafka client pattern and includes appropriate regression coverage.

Review tier: Balanced
Findings: None

[David Fowler (davidfowl)]

/backport to release/13.5

[github-actions]

Started backporting to release/13.5 (link to workflow run)

[github-actions]

Tests selector

2 / 99 PR test projects · 2 PR jobs · 0 advisory-only targets, from 2 changed files.

Selected PR test projects (2 / 99)

Aspire.Hosting.Kafka.Tests, Aspire.Playground.Tests

Selected PR jobs (2)

extension-e2e, typescript-api-compat

Advisory workflow impact (0)

none

---

How these were chosen — grouped by what changed

🔧 src/Aspire.Hosting.Kafka/KafkaBuilderExtensions.cs (changed source)
→ 1 directly: Aspire.Hosting.Kafka.Tests
→ 1 via the project graph: Aspire.Playground.Tests (2 hops)

🧪 tests/Aspire.Hosting.Kafka.Tests/AddKafkaTests.cs (changed test)
→ 1 directly: Aspire.Hosting.Kafka.Tests

Job reasons

Job	Triggered by
extension-e2e	• src/Aspire.Hosting.Kafka/KafkaBuilderExtensions.cs • affected project Aspire.Hosting.Kafka
typescript-api-compat	affected project Aspire.Hosting.Kafka

---

Selection computed for commit f1aa806.

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused lifecycle fix follows existing Kafka integration patterns and has appropriate regression coverage.

Review tier: Balanced
Findings: None

[github-actions]

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

[github-actions]

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

[James Newton-King (JamesNK)]

Reviewed the Kafka health-check lifetime change and regression coverage. 0 issues found.

[mahdiyar (mahdiyar021)]

I think this PR may also fixes this issue