Native (non-Docker) test mode with Windows support by honza-kasik · Pull Request #6 · modcluster/mod_cluster-testsuite

honza-kasik · 2026-03-26T13:32:07Z

Add support for running the mod_cluster test suite without Docker, using local WildFly and httpd processes (-Pnative)
Refactor container abstractions (WildFlyContainer → WildFlyWorker, BalancerContainer → Balancer) to be platform-independent, with Docker and native implementations
Fix numerous test issues discovered during Windows native runs (timing, path handling, process lifecycle)
Disable two tests blocked by upstream WildFly bugs
Add GitHub Actions CI with a matrix covering Linux Docker + Windows native

Architecture changes

WildFlyWorker (abstract) with DockerWildFlyWorker and NativeWildFlyWorker implementations
Balancer (abstract) with DockerUndertowBalancer, DockerHttpdBalancer, NativeUndertowBalancer, and NativeHttpdBalancer
CommandResult — platform-independent exec result replacing Testcontainers' Container.ExecResult
NativePortAllocator — manages port offsets for concurrent native workers
NativeProcessManager — process lifecycle with shutdown hooks and tree kill
NativeServerExtractor — extracts and caches WildFly distributions
TestMode — selects Docker vs native based on -Dtest.mode=native
All tests updated to use new abstractions; platform-specific logic isolated in implementations

Test fixes

Fix StickySessionForce to capture first response instead of polling
Fix exit.jsp to use System.exit instead of halt for reliable TCP teardown
Fix health-check-interval/broken-node-timeout units (seconds → milliseconds)
Fix port offset handling for multiple balancers in native mode
Disable HTTP keep-alive and HTTP/2 for reliable mod_cluster routing
Tolerate transient unknown responses in broken node verification
Clean up stale SSL configs on native httpd balancer start
Relax stop-context failure threshold from 10 to 15

Disabled tests

testStatefulEjbStickiness — JBEAP-33250: WildFly regression wildfly/wildfly@d3b318b sets JSESSIONID cookie path without leading /, breaking EJB-over-HTTP session stickiness
testNodeTimeout — JBEAP-9624 / JBEAP-26262: node-timeout not applied correctly on Undertow balancer

CI

GitHub Actions workflow running StickySessionTest on:
- Ubuntu / Undertow / JDK 21 (Docker)
- Ubuntu / Undertow / JDK 17 (Docker)
- Windows / Undertow / JDK 21 (Native)

jajik

A few things caught my eye (but frankly, these huge PRs are hard to review).

jajik · 2026-05-25T12:25:32Z

Some failures are caused by the old httpd url (https://dlcdn.apache.org/httpd/httpd-2.4.66.tar.gz -> https://dlcdn.apache.org/httpd/httpd-2.4.67.tar.gz).

honza-kasik · 2026-05-25T12:28:10Z

@jajik Thanks, I just found out. I switched to archive instead of using mirror. I don't want to dig into CI everyday :)

marekkopecky · 2026-05-26T11:34:13Z

It seems that #6 (comment) is last unresolved thread ....

jajik

Fine by me. Thank you.

jajik · 2026-05-26T13:32:55Z

(Of course, it would be nice to squash the commits meaningfully.)

Add NativeWildFlyWorker, NativeHttpdBalancer, NativeUndertowBalancer, NativePortAllocator, NativeProcessManager, and NativeServerExtractor for running tests against locally installed servers instead of Docker containers. Introduce TestMode and factory methods on Balancer.create()/WildFlyWorker.create() to dispatch on -Dtest.mode=native. Includes config caching for faster startup, process tree management, httpd binary and config discovery on Windows/Linux, idempotent balancer configuration, standalone.xml backup/restore, and updated architecture documentation.

Disable HTTP keep-alive and HTTP/2 in Undertow mod_cluster filter for reliable routing. Fix health-check-interval and broken-node-timeout units to milliseconds and reduce to 1s/3s. Use System.exit instead of halt in exit.jsp for clean TCP teardown. Also: fix port offset handling for multiple native balancers, fix StickySessionForce test capture, relax stop-context threshold to 15, use stricter request count on Linux with relaxed threshold on Windows, tag CPU load test as docker-only, disable testNodeTimeout (JBEAP-9624), and remove redundant post-failover verification request.

Prevent EJB client from bypassing the balancer via topology discovery by disabling it in EjbClientAppBuilder. Bind native workers to 127.0.0.1 instead of 0.0.0.0. Add TRACE logging for org.wildfly.httpclient, org.jboss.ejb.client, and org.jboss.remoting. Disable testStatefulEjbStickiness due to WildFly bug WFLY-21930 (malformed routing cookie path).

Add StickySessionTest across OS/JDK combinations (Ubuntu JDK 17+21 Docker, Windows JDK 21 native). Use single-line shell commands for Windows compatibility, extract shared CI variables, and rework matrix as cross-product of os/java/balancer with excludes. Also switch httpd source download to archive.apache.org for permanent URL retention.

Clean up stale SSL configs on native httpd balancer start to prevent duplicate Listen directives. Replace sed-based VirtualHost patching with a shipped mod_proxy_cluster_ssl.conf config file. Use Path.of for platform-safe SSL subpath joins. Add SslFailoverTest to CI matrix.

Extract shared isWindows() to TestMode, removing 5 private copies. Extract WFLYSRV0025 startup pattern constant to WildFlyWorker base class. Add missing Javadoc to Balancer, WildFlyWorker, CommandResult, and ContainerUtils. Update bug reference from JBEAP-33250 to WFLY-21930.

honza-kasik marked this pull request as draft March 26, 2026 13:32

honza-kasik force-pushed the native branch 7 times, most recently from 13194e5 to 200e363 Compare May 13, 2026 13:23

honza-kasik force-pushed the native branch from 9074515 to 348a3ad Compare May 25, 2026 07:22

honza-kasik changed the title ~~Phase 1+2 of supporting native runs implemented~~ Native (non-Docker) test mode with Windows support May 25, 2026

honza-kasik marked this pull request as ready for review May 25, 2026 07:50

honza-kasik requested a review from marekkopecky May 25, 2026 07:50

jajik suggested changes May 25, 2026

View reviewed changes