Performance: Maven 4 at parity with Maven 3 on large reactors (2:45 → 1:18)

[gnodet]

Summary

JFR profiling of Maven 4 RC5/RC6 on a 4,383-module generated reactor project revealed several performance regressions compared to Maven 3. A clean install -DskipTests that takes ~1:15 on Maven 3.9.16 was taking ~2:45 on Maven 4 RC6.

After a series of targeted optimizations across 4 repositories, Maven 4 now completes the same build in ~1:18 — at parity with Maven 3.

Benchmark Results

Test machine: Apple M4 Pro, JDK 21. Project: 4,383-module diamond-graph reactor, clean install -DskipTests -q.

Configuration	Wall time	vs RC6
Maven 3.9.16	1:14 – 1:18	—
Maven 4 RC6 (unpatched)	2:45	baseline
+ PathConflictResolver default	1:45	-36%
+ resolver optimizations	1:22	-50%
+ all optimizations	1:18	-53%

Hotspots & Fixes

Each optimization was identified via JFR CPU profiling. The fixes are grouped by repository and listed in order of impact.

1. Conflict Resolver: O(N²) → O(N) — apache/maven#12662

JFR: ClassicConflictResolver.gatherConflictItems — 46.7% CPU

The ClassicConflictResolver performs O(N×M) recursive DFS in gatherConflictItems(). Switching the default from classic to path (PathConflictResolver, available since resolver 2.0.11, same resolution results) reduces this to O(N). Single biggest win: 2:45 → 1:45.

2. TransitiveDependencyManager — apache/maven-resolver#2014

JFR: deriveChildManager + Key.equals + MMap.done — 28% CPU combined

TransitiveDependencyManager (Maven 4 default, deriveUntil=Integer.MAX_VALUE) creates a new child manager at every graph node. Four layered fixes:

• Instance self-reuse — when no new management data is collected (common for transitive POMs without <dependencyManagement>), return this instead of a new instance. Restores pool cache transparency.
• Key coordinate caching — cache groupId/artifactId/extension/classifier strings eagerly to avoid RelocatedArtifact virtual dispatch per equals().
• Cons-list parent pointer — replace O(depth)-copied ArrayList<AbstractDependencyManager> with a single parent reference. O(1) derive, cascading hashCode, identity-shortcircuit equals.
• Varargs elimination — replace Objects.hash(Object...) with manual 31*h + field.hashCode() chains throughout Key, constructor, Holder, and MMap to eliminate ~10 temporary Object[] allocations per derive call.
• MEMO_CACHE_SIZE 4 → 16 — self-reuse causes long-lived managers to serve children from many BFS wavefronts.
• PathConflictResolver O(1) cycle detection — each Path node maintains a Set<String> of all conflict IDs from root, making hasConflictIdOnPathToRoot a Set.contains() instead of O(depth) parent walk.

3. InstallMojo / DeployMojo: O(N²) → O(N) — apache/maven-install-plugin#427, apache/maven-deploy-plugin#684

JFR: PluginContainer.getPluginsAsMap — 5.6% CPU

Both InstallMojo.execute() and DeployMojo.execute() scan the entire reactor project list on every module invocation to check which projects use the plugin. In a 4,383-module build this is ~19.2M filter evaluations. Fixed by caching the filtered list in the first reactor project's plugin context on first invocation.

4. Reactor Sort: O(N² log N) → O(N log N) — apache/maven#12652

JFR: MavenProject.equals — 23.5% CPU

DefaultGraphBuilder uses result.sort(comparing(sortedProjects::indexOf)) where ArrayList.indexOf() is O(n), causing O(N² log N) total MavenProject.equals() calls (~230M for 4,383 modules). Replaced with a HashMap<MavenProject, Integer> index for O(1) lookups.

Also includes:

• DefaultModelObjectPool (~8% CPU) — cache getPooledTypes() set at construction, inline Objects.hash() varargs in PoolKey, add hashCode fast-rejection in equals()
• PhaseComparator (~2% CPU) — pre-build HashMap<String, Integer> for O(1) phase lookups

5. Model Building Pipeline — apache/maven#12653

JFR: DefaultModelObjectPool.PoolKey + Dependency builders — 37% + 18% CPU

• Add Builder getters and *ToBuilder merger variants to generated model classes
• Defer Dependency.build() in DefaultDependencyManagementInjector — accumulate as Builder objects, build once at end
• Optimize computeLocations() — replace Stream.concat().collect() with HashMap.putAll() + Map.copyOf()
• Precompute locationsHashCode at build time for fast inequality checks

6. PrintWriter Lock Contention — apache/maven#12654

JFR: 1,470ms blocked time on PrintWriter.println()

AsyncDrainWriter wraps the logging Consumer<String> with a lock-free ConcurrentLinkedQueue + non-blocking drain via ReentrantLock.tryLock(). Eliminates all contention during parallel model building (-T1C).

7. Location Tracking Wire-up — apache/maven#12655

JFR: 97 InputLocation.of() allocations per POM

ModelBuilderRequest.isLocationTracking() existed but was only checked in one place. Now wired through XmlReaderRequest → DefaultModelXmlFactory → MavenStaxReader so the parser actually skips all InputLocation.of() calls when tracking is disabled.

All PRs

Repository	PR	Status	Description
apache/maven	#12662	✅ Ready	Enable PathConflictResolver by default
apache/maven	#12652	✅ Ready	Optimize reactor sort, model pool, phase comparator
apache/maven	#12653	✅ Ready	Optimize model building pipeline
apache/maven	#12654	✅ Ready	AsyncDrainWriter — lock contention elimination
apache/maven	#12655	✅ Ready	Wire location tracking to XML parser
apache/maven-resolver	#2014	✅ Ready	TransitiveDependencyManager performance
apache/maven-install-plugin	#427	✅ Ready	Cache projectsUsingPlugin — O(N²) → O(N)
apache/maven-deploy-plugin	#684	✅ Ready	Cache projectsWithDeployExecution — O(N²) → O(N)

Reproducing

# Generate the test project
git clone https://github.com/maven-turbo-reactor/maven-multiproject-generator
cd maven-multiproject-generator && ./generate.sh

# Build with gnodet's bench branch that includes all resolver + maven patches
git clone -b bench/resolver-2.0.22 https://github.com/gnodet/maven
cd maven && mvn install -DskipTests -q

# Benchmark
cd generated && time path/to/patched-maven/bin/mvn clean install -DskipTests -q

See also: benchmark gist