Java GC — Unbounded Cache Full GC Spiral

Every Java application runs a second program inside the JVM — the Garbage Collector. It decides when memory gets freed, how long your threads pause, and whether your latency SLAs hold up under load. Most developers treat it like a black box and then wonder why their microservice spikes to 500ms every few seconds in production.

Before automatic memory management, C and C++ developers had to manually allocate and free every byte. Java solved this with a managed heap and a runtime that tracks object reachability — if nothing in your program can reach an object, its memory can be reclaimed. That single idea eliminated an entire class of bugs but introduced a new challenge: the collector itself consumes CPU and introduces pauses.

The core misconception: GC pauses are inevitable and unfixable. They are not. Modern collectors offer pause-time guarantees independent of heap size — but only if you understand the trade-offs and tune correctly for your workload.

What is Garbage Collection in Java?

Garbage Collection in Java is the JVM's automatic memory management mechanism. The GC periodically identifies objects that are no longer reachable from any GC root (thread stacks, static fields, JNI references) and reclaims their heap memory. This eliminates manual memory management but introduces pauses and CPU overhead that must be managed in production.

The JVM determines object reachability through a reachability analysis starting from GC roots. An object is considered dead — and eligible for collection — when no chain of references from any root can reach it. This is fundamentally different from reference counting (used in early Python/PHP) which cannot handle cyclic references. Java's tracing GC handles cycles naturally because it only cares about reachability, not reference count.

The key production insight: GC does not run when memory is low. GC runs when allocation pressure triggers it. This means a service with a large heap but low allocation rate may run GC infrequently, while a service with a small heap and high allocation rate may run GC constantly. Allocation rate, not heap size, is the primary driver of GC frequency.

package io.thecodeforge.gc;

import java.util.ArrayList;
import java.util.List;

/**
 * Demonstrates how Java GC determines object reachability.
 *
 * Key concept: An object is reachable if any GC root can access it
 * through a chain of references. When the chain breaks, the object
 * becomes eligible for collection.
 */
public class ReachabilityDemo {

    public static void main(String[] args) {
        // Object created on the heap — referenced by local variable 'order'
        // 'order' is a GC root (stack reference)
        Order order = new Order("ORD-001", 149.99);
        System.out.println("Order created: " + order.getId());

        // After this reassignment, the original Order object has no
        // reachable references. It becomes eligible for GC.
        order = new Order("ORD-002", 299.99);
        // The first Order("ORD-001") is now unreachable — GC will reclaim it

        // Demonstrating cyclic references — GC handles this correctly
        OrderNode nodeA = new OrderNode("A");
        OrderNode nodeB = new OrderNode("B");
        nodeA.next = nodeB;
        nodeB.next = nodeA; // cycle: A -> B -> A

        // Even though A and B reference each other, if we null out
        // our stack references, both become unreachable and are collected
        nodeA = null;
        nodeB = null;
        // The cycle A -> B -> A is still intact in memory, but no GC root
        // can reach either node. Both are eligible for collection.
    }

    static class Order {
        private final String id;
        private final double amount;

        Order(String id, double amount) {
            this.id = id;
            this.amount = amount;
        }

        String getId() { return id; }
    }

    static class OrderNode {
        final String name;
        OrderNode next;

        OrderNode(String name) {
            this.name = name;
        }
    }
}

The Generational Heap — Why Most Objects Die Young

The JVM heap is divided into generations based on the weak generational hypothesis: most objects die young, and objects that survive one collection are likely to survive many more. This observation drives the generational heap design that every modern JVM collector uses.

The young generation consists of eden space (where new objects are allocated) and two survivor spaces (S0 and S1). New objects are allocated in eden. When eden fills up, a minor GC (young collection) runs: live objects in eden are copied to one survivor space, and live objects in the other survivor space are also copied and aged. Objects that survive enough young collections (controlled by -XX:MaxTenuringThreshold) are promoted to the old generation.

The old generation holds long-lived objects. When the old generation fills up or a collection threshold is reached, a major GC runs. In G1, this is a mixed GC that collects both young and old regions. In extreme cases, a full GC (stop-the-world compaction of the entire heap) is triggered — this is the catastrophic failure mode you must avoid.

The critical production insight: the tenuring threshold determines how quickly objects move to old generation. Too low, and short-lived objects pollute old generation, increasing old gen GC frequency. Too high, and survivor spaces overflow, forcing premature promotion. Both paths degrade performance.

package io.thecodeforge.gc;

import java.util.ArrayList;
import java.util.List;

/**
 * Demonstrates how allocation patterns interact with the generational heap.
 *
 * Objects that survive young collections are promoted to old generation.
 * Understanding this promotion mechanism is critical for tuning.
 */
public class GenerationalBehaviorDemo {

    /**
     * Pattern 1: Short-lived objects — ideal for generational GC.
     * These objects die in eden and never reach old generation.
     * GC can reclaim them with a fast young collection.
     */
    public void processRequest() {
        // These objects are created, used, and become unreachable
        // within a single method call. They die in eden.
        String requestId = java.util.UUID.randomUUID().toString();
        byte[] payload = new byte[4096];
        List<String> validationErrors = new ArrayList<>();

        // After this method returns, all three objects become unreachable
        // because they are only referenced by local variables (stack roots).
    }

    /**
     * Pattern 2: Long-lived cached objects — promoted to old gen.
     * These objects survive young collections and get promoted.
     * They occupy old generation permanently (or until eviction).
     *
     * Production risk: If this cache grows unbounded, old generation
     * fills up and triggers full GC or OOM.
     */
    private final List<byte[]> longLivedCache = new ArrayList<>();

    public void cacheData(byte[] data) {
        // This reference keeps the byte array alive indefinitely.
        // After surviving MaxTenuringThreshold young collections,
        // it is promoted to old generation.
        longLivedCache.add(data);
    }

    /**
     * Pattern 3: Premature promotion — objects that should die young
     * but get promoted because survivor space is full.
     *
     * If allocation rate exceeds survivor space capacity, objects
     * are promoted directly to old generation even if they are short-lived.
     * This is called premature promotion and it pollutes old generation.
     *
     * Fix: Increase survivor space ratio (-XX:SurvivorRatio)
     *       or reduce allocation rate.
     */
    public void burstAllocation() {
        // If this loop runs fast enough to fill eden AND overflow
        // survivor space, these temporary objects get promoted to
        // old generation even though they die after each iteration.
        for (int i = 0; i < 100_000; i++) {
            byte[] temp = new byte[256];
            // temp is short-lived, but under pressure it may be
            // prematurely promoted to old generation
        }
    }

    /**
     * Production tuning flags for generational behavior:
     *
     * -XX:NewRatio=2              // old:young = 2:1 (default for most collectors)
     * -XX:SurvivorRatio=8         // eden:survivor = 8:1 (default)
     * -XX:MaxTenuringThreshold=15 // objects survive 15 young GCs before promotion
     * -XX:+AlwaysTenure            // promote immediately (dangerous — avoid)
     * -XX:+NeverTenure             // never promote (survivor overflow → old gen)
     *
     * Monitor promotion rate with:
     *   jstat -gcutil <pid> 1000
     *   Watch 'O' column (old gen utilization) for steady growth.
     *   Steady growth with low live data = premature promotion.
     */
}

GC Algorithms — Mark-Sweep, Copying, and Compaction

All GC algorithms are built on three fundamental operations: marking (identifying live objects), sweeping (reclaiming dead objects' memory), and compacting (defragmenting live objects to create contiguous free space). Different collectors combine these operations differently to optimize for pause time, throughput, or memory efficiency.

Mark-and-sweep identifies live objects (mark phase) then reclaims unmarked memory (sweep phase). The problem: it creates fragmentation. After many allocation-deallocation cycles, free memory is scattered in small chunks. Large object allocations may fail even when total free memory is sufficient — this is external fragmentation.

Copying collectors solve fragmentation by copying live objects to a fresh region and discarding the old region entirely. This is inherently compacting — live objects end up contiguous. The cost: copying live objects takes time proportional to the live data set, and you need double the memory (from-space and to-space). The generational heap reduces this cost by only copying in young generation.

Mark-and-compact identifies live objects then slides them to one end of the heap, creating one contiguous free region. This avoids the double-memory cost of copying but requires updating every reference to moved objects — a potentially expensive operation that must be done during a stop-the-world pause or with complex concurrent mechanisms.

package io.thecodeforge.gc;

import java.util.ArrayList;
import java.util.List;

/**
 * Demonstrates how different GC algorithm characteristics
 * affect production behavior.
 *
 * This is not a GC implementation — it illustrates the concepts
 * that drive real collector design decisions.
 */
public class GCAlgorithmDemo {

    /**
     * MARK-AND-SWEEP characteristic:
     * - Fast reclaim but creates fragmentation
     * - External fragmentation: total free > requested, but not contiguous
     *
     * Production impact: After hours of operation, allocation of large
     * objects fails even though 40% of heap is free — it's fragmented.
     * This triggers unnecessary full GC or OOM.
     */
    public void demonstrateFragmentation() {
        // Imagine this array is the heap, each index is a memory block
        // true = occupied, false = free
        boolean[] heap = new boolean[100];

        // Simulate allocation pattern: allocate and free alternating blocks
        for (int i = 0; i < 100; i++) {
            heap[i] = true; // allocate
        }
        for (int i = 0; i < 100; i += 2) {
            heap[i] = false; // free every other block
        }
        // Result: 50% free, but no contiguous block of size > 1
        // A request for 3 contiguous blocks fails despite 50 free blocks
        // This is external fragmentation — the problem compaction solves
    }

    /**
     * COPYING COLLECTOR characteristic:
     * - Copies live objects to to-space, discards from-space
     * - Inherently compacting — no fragmentation
     * - Cost: proportional to live data, not dead data
     * - Requires double the memory (from + to spaces)
     *
     * Production insight: Copying cost is why large live data sets
     * cause longer young collection pauses. If your service has
     * 2GB of live objects in young gen, copying takes measurable time.
     */
    public void demonstrateCopyingCost() {
        // Simulating live data that must be copied during young GC
        List<byte[]> liveObjects = new ArrayList<>();
        for (int i = 0; i < 10_000; i++) {
            liveObjects.add(new byte[1024]); // 1KB each = ~10MB live data
        }

        // During young GC, all 10MB must be copied to survivor space.
        // If only 1MB were live, the cost would be 10x lower.
        // This is why reducing live data in young gen reduces pause time.
        //
        // Real production fix: avoid holding references to temporary
        // objects across request boundaries. Let them die in eden.
    }
}

G1 GC — The Default Workhorse

G1 (Garbage-First) has been the default JVM collector since Java 9. It divides the heap into equal-sized regions (1MB to 32MB) and prioritizes collecting regions with the most garbage — hence 'garbage-first'. G1 maintains a remembered set per region tracking incoming references, enabling independent region collection without scanning the entire heap.

G1 operates in young-only and mixed collection cycles. Young GC collects survivor and eden regions. When the heap occupancy exceeds the Initiating Heap Occupancy Percent (IHOP), G1 triggers a concurrent marking cycle. After marking completes, subsequent mixed GCs collect both young and old regions identified as mostly garbage.

The critical production insight: G1's pause time is primarily driven by the number of regions it must collect in a single pause, not heap size. A 64GB heap with aggressive evacuation can pause longer than a 4GB heap with conservative settings. This is the opposite of what most engineers assume.

package io.thecodeforge.gc;

import java.util.concurrent.ConcurrentHashMap;
import java.util.Map;

/**
 * Demonstrates allocation patterns that stress G1 differently.
 *
 * Key insight: G1 humongous objects (>50% region size) bypass normal
 * allocation and can trigger to-space exhausted failures.
 */
public class G1TuningExample {

    // Cache with large value objects — common source of humongous allocations
    private final Map<String, byte[]> payloadCache = new ConcurrentHashMap<>();

    /**
     * BAD: Allocates objects that may exceed humongous threshold.
     * With default 1MB region size, objects > 512KB are humongous.
     * With 32MB regions, threshold is 16MB — much safer for large payloads.
     *
     * Tuning: -XX:G1HeapRegionSize=32M
     *         -XX:G1ReservePercent=15
     *         -XX:InitiatingHeapOccupancyPercent=35
     */
    public void cacheLargePayload(String key, int sizeBytes) {
        byte[] payload = new byte[sizeBytes];
        for (int i = 0; i < Math.min(sizeBytes, 1024); i++) {
            payload[i] = (byte) (i & 0xFF);
        }
        payloadCache.put(key, payload);
    }

    /**
     * BETTER: Chunk large payloads to stay below humongous threshold.
     * Each chunk is independently collectible as a regular object.
     */
    public void cacheChunkedPayload(String key, byte[] fullPayload) {
        int chunkSize = 256 * 1024; // 256KB chunks
        int numChunks = (fullPayload.length + chunkSize - 1) / chunkSize;

        for (int i = 0; i < numChunks; i++) {
            int offset = i * chunkSize;
            int length = Math.min(chunkSize, fullPayload.length - offset);
            byte[] chunk = new byte[length];
            System.arraycopy(fullPayload, offset, chunk, 0, length);
            payloadCache.put(key + ":chunk:" + i, chunk);
        }
    }

    /**
     * Production G1 flags for a 16GB heap with mixed allocation profile:
     *
     * -XX:+UseG1GC
     * -Xms16g -Xmx16g
     * -XX:G1HeapRegionSize=16m
     * -XX:MaxGCPauseMillis=200
     * -XX:G1ReservePercent=15
     * -XX:InitiatingHeapOccupancyPercent=35
     * -XX:G1MixedGCCountTarget=8
     * -XX:G1MixedGCLiveThresholdPercent=85
     * -Xlog:gc*,gc+humongous=debug:file=/var/log/gc.log:time,uptime,level,tags
     */
}

ZGC — Sub-Millisecond Pause Collector

ZGC (Z Garbage Collector) was introduced as experimental in JDK 11 and became production-ready in JDK 15. Its defining characteristic: pause times stay below 10ms regardless of heap size — tested up to 16TB heaps. ZGC achieves this through concurrent everything: marking, relocation, and reference processing all happen while application threads run.

ZGC uses load barriers with colored pointers. Every object reference carries metadata bits (marked0, marked1, remap, finalize) embedded in the pointer itself. The load barrier intercepts every object access to check if the reference needs remapping. This is the fundamental trade-off: ZGC replaces long GC pauses with per-access overhead on every object load.

As of JDK 21, ZGC supports generational mode (-XX:+ZGenerational) which dramatically improves throughput by focusing collection on young objects. Non-generational ZGC collects the entire heap every cycle, which limits throughput on allocation-heavy workloads.

package io.thecodeforge.gc;

import java.util.concurrent.atomic.AtomicLong;

/**
 * ZGC-specific considerations for production workloads.
 *
 * ZGC trades per-access overhead for near-zero pause times.
 * The load barrier adds ~4-8% overhead on pointer-heavy workloads.
 */
public class ZGCTuningExample {

    private final AtomicLong allocationCounter = new AtomicLong(0);

    /**
     * Production ZGC flags for a 32GB heap, latency-sensitive service:
     *
     * -XX:+UseZGC
     * -XX:+ZGenerational              // JDK 21+ — critical for throughput
     * -Xms32g -Xmx32g                // Always set Xms=Xmx for ZGC
     * -XX:SoftMaxHeapSize=28g         // ZGC-specific: target heap occupancy
     * -XX:ZCollectionInterval=5       // Suggest GC cycle every 5 seconds
     * -XX:ConcGCThreads=4             // Concurrent GC threads
     * -Xlog:gc*:file=/var/log/zgc.log:time,uptime,level,tags
     *
     * CRITICAL: ZGC uses ~20% native memory overhead beyond -Xmx.
     * Container memory limit must be heap * 1.25 minimum.
     */

    /**
     * ZGC SoftMaxHeapSize is unique — it tells ZGC to try to stay below
     * this threshold but can exceed it under allocation pressure.
     *
     * Use case: Set heap to 32GB, SoftMaxHeapSize to 28GB.
     * ZGC will trigger cycles aggressively to stay under 28GB.
     * Only allocates into the remaining 4GB under extreme pressure.
     */
    public void demonstrateSoftMaxHeapConcept() {
        // With SoftMaxHeapSize=28g and Xmx=32g:
        // - ZGC targets 28GB occupancy
        // - If allocation pressure pushes past 28GB, ZGC cycles more aggressively
        // - If it hits 32GB, allocation stalls (not OOM, but backpressure)
    }
}

Shenandoah — Red Hat's Low-Pause Contender

Shenandoah is Red Hat's concurrent compacting collector, available as production-ready since JDK 12. It achieves low pause times through concurrent evacuation — moving live objects while application threads run — using Brooks pointers (an indirection layer on every object).

Shenandoah differs from ZGC in a critical way: it uses Brooks pointers (every object has a forwarding pointer field) instead of colored pointers. This means Shenandoah does not require specific pointer bit layouts and works with compressed oops, reducing memory overhead compared to ZGC on heaps under 32GB.

Shenandoah operates in three concurrent phases: concurrent mark, concurrent evacuate, and concurrent update-refs. The initial mark and final mark phases are short stop-the-world pauses, typically under 10ms. Shenandoah's pacing mechanism backpressures allocation threads proportionally when the collector falls behind, creating smoother degradation than ZGC's hard allocation stalls.

package io.thecodeforge.gc;

import java.util.ArrayList;
import java.util.List;

/**
 * Shenandoah-specific production considerations.
 *
 * Shenandoah uses Brooks pointers — every object has an extra forwarding
 * pointer field. This adds 8 bytes per object on 64-bit systems.
 */
public class ShenandoahTuningExample {

    /**
     * Brooks pointer overhead calculation:
     *
     * Object with 2 fields (16 bytes header + 16 bytes data = 32 bytes)
     * + 8 bytes Brooks pointer = 40 bytes per object
     * Overhead: 25% increase per object
     *
     * For 10 million small objects: ~80MB additional memory
     * For 100 million small objects: ~800MB additional memory
     */
    public long estimateBrooksOverhead(int objectCount) {
        return (long) objectCount * 8;
    }

    /**
     * Production Shenandoah flags for a 16GB heap:
     *
     * -XX:+UseShenandoahGC
     * -Xms16g -Xmx16g
     * -XX:ShenandoahGCHeuristics=adaptive
     * -XX:ShenandoahAllocationThreshold=10
     * -XX:+UseCompressedOops               // works with Shenandoah (unlike ZGC)
     * -Xlog:gc*:file=/var/log/shenandoah.log:time,uptime,level,tags
     */

    /**
     * Shenandoah pacing is a unique feature that backpressures allocation
     * threads when the collector falls behind.
     *
     * Unlike ZGC which stalls allocation entirely, Shenandoah slows down
     * allocating threads proportionally. This creates smoother latency
     * degradation under load rather than sharp spikes.
     */
    public void demonstratePacingBehavior() {
        List<byte[]> allocations = new ArrayList<>();

        // Under heavy allocation, Shenandoah will pace this loop
        // by adding small delays to each allocation.
        // The delay is proportional to how far behind the collector is.
        for (int i = 0; i < 100_000; i++) {
            allocations.add(new byte[1024]);
        }
    }
}

JVM Flags That Actually Matter

Most JVM GC flags have sensible defaults. A small subset moves the needle in production. Understanding which flags to adjust — and when — prevents the common anti-pattern of blindly copying flags from blog posts without understanding their impact on your specific workload.

Flags fall into three categories: heap sizing, collector behavior, and logging. Heap sizing flags (-Xms, -Xmx, -XX:NewRatio) control memory layout. Collector behavior flags (-XX:MaxGCPauseMillis, -XX:InitiatingHeapOccupancyPercent) control collection strategy. Logging flags (-Xlog:gc) enable observability. The third category is the most important — you cannot tune what you cannot measure.

→ JVM Memory Issues in Production: Debugging Guide (OOM, GC, Leaks) — When flags alone are not enough and you need live incident triage

package io.thecodeforge.gc;

/**
 * Production JVM flag configurations organized by collector.
 * These are starting points — tune based on measured workload characteristics.
 */
public class ProductionJVMFlags {

    /**
     * UNIVERSAL FLAGS (apply to all collectors):
     *
     * -Xms<size> -Xmx<size>         // Set min=max to avoid resize overhead
     * -XX:+AlwaysPreTouch             // Pre-zero heap pages at startup
     * -XX:+DisableExplicitGC          // Ignore System.gc() calls
     * -XX:+HeapDumpOnOutOfMemoryError // Auto heap dump on OOM
     * -XX:HeapDumpPath=/var/log/      // Where to write heap dumps
     * -XX:+UseContainerSupport        // Respect cgroup limits (default JDK 10+)
     * -XX:MaxRAMPercentage=75.0       // Set heap as % of container memory
     * -XX:NativeMemoryTracking=detail // Track off-heap memory usage
     *
     * LOGGING FLAGS (always enable in production):
     * -Xlog:gc*:file=/var/log/gc.log:time,uptime,level

GC Algorithm Comparison: Serial, Parallel, G1, ZGC, Shenandoah

Choosing the right garbage collector depends on your workload's pause-time sensitivity, heap size, and throughput requirements. The table below summarizes the key characteristics of each major collector available in the JVM.

Key takeaway: For most web services, start with G1. Only move to ZGC or Shenandoah when your measured p99 latency exceeds 100ms after tuning G1. Serial and Parallel are legacy choices for resource-constrained or batch workloads.

System.gc() and finalize() — Patterns to Avoid

Two legacy Java mechanisms that should be avoided in production: System.gc() and finalize(). Both degrade GC performance and unpredictability.

System.gc() — An explicit request to run the garbage collector. It's a hint, not a command, but JVM often treats it as a full GC trigger (especially with -XX:+DisableExplicitGC disabled). Calling it frequently causes unnecessary full GC pauses, wrecking latency. Also, some frameworks like RMI, NIO, and JNDI call it internally. Always set -XX:+DisableExplicitGC in production to mitigate accidental calls.

finalize() — The finalize() method, defined in Object, runs before an object is reclaimed. It's unpredictable — the JVM may never call it before exit, and GC threads can finalize objects out of order. Additionally, finalize() can resurrect objects by assigning this to a reachable reference. The method also introduces latency as the JVM must finalize objects in a separate pass. Since Java 9, finalize() is deprecated. Use Cleaner (JDK 9+), PhantomReference with a cleanup thread, or AutoCloseable / try-with-resources instead.

package io.thecodeforge.gc;

import java.lang.ref.Cleaner;

/**
 * Demonstrates how to avoid System.gc() and finalize().
 * 
 * BAD PRACTICES:
 * 1. Calling System.gc() - triggers unnecessary full GC
 * 2. Overriding finalize() - unpredictable, deprecated.
 *
 * GOOD: Use Cleaner (JDK 9+) or PhantomReference with reference queue.
 */
public class AvoidSystemGCAndFinalize {

    // BAD - Avoid this
    @Override
    @Deprecated(since = "9")
    protected void finalize() throws Throwable {
        try {
            // Cleanup logic here - but this may never run!
            close();
        } finally {
            super.finalize();
        }
    }

    private void close() {
        System.out.println("Cleanup (if finalize runs)");
    }

    // GOOD - Use Cleaner (JDK 9+)
    private static final Cleaner CLEANER = Cleaner.create();

    // State that needs cleaning
    private final Cleaner.Cleanable cleanable;

    public AvoidSystemGCAndFinalize() {
        // Register a cleaning action
        this.cleanable = CLEANER.register(this, () -> {
            // This runs when the object becomes phantom-reachable
            System.out.println("Cleanup via Cleaner");
        });
    }

    public static void main(String[] args) {
        // NEVER do this:
        // System.gc(); // tells JVM to run GC - pauses, unpredictable

        // Better: let GC decide.
        // Disable explicit calls with -XX:+DisableExplicitGC
        
        // Use Cleaner or try-with-resources for cleanup.
    }
}

Advantages and Disadvantages of Garbage Collection

Garbage Collection is a mixed blessing. It eliminates manual memory management bugs but introduces new operational challenges. The table below summarizes the trade-offs.

Key takeaway: The disadvantages can be mitigated with proper collector selection and tuning. For most production services, the benefits far outweigh the costs, but ignore the downsides at your peril.

GC Tuning Flags Reference Table

This table lists the most important GC tuning flags along with their purpose and typical values. Use it as a quick reference when configuring JVM options for production.

Key takeaway: The most impactful flags are GC logging (for observability) and heap sizing. Tuning collector-specific flags without enabling logs is like fixing a car engine blindfolded – possible but wasteful.

Practice Problems: GC Diagnosis and Tuning

Test your understanding of GC concepts with these five practical problems. Each problem presents a real-world scenario; identify the issue and propose a fix or tuning change.

Problem 1: Unbounded Cache Scenario: A user service caches profile objects in a HashMap. Over a weekend spike, GC logs show rising old-gen occupancy followed by frequent full GCs. P99 latency jumps from 50ms to 5s. Question: What is the likely cause and the immediate fix? Answer: Unbounded cache retains all entries, filling old gen. Immediate fix: apply size and time-based eviction (e.g., Caffeine with maximumSize and expireAfterWrite).

Problem 2: Large Object Allocation Scenario: A service using G1 with default region size (1MB) allocates many 800KB byte arrays. GC logs show numerous humongous allocation warnings and to-space exhaustion. Question: What tuning change can reduce humongous objects? Answer: Increase G1HeapRegionSize (e.g., -XX:G1HeapRegionSize=4m) so 800KB objects are under the 50% humongous threshold. Alternatively, chunk large allocations.

Problem 3: Metaspace OOM Scenario: After deploying a new microservice, pods restart every few hours with OutOfMemoryError. Heap is not full; metaspace shows steady growth. Thread count is stable. Question: What is the likely root cause and how to diagnose? Answer: Class loader leak (e.g., from repeated dynamic class generation or redeployment). Use -XX:NativeMemoryTracking=detail and jcmd to monitor metaspace. Consider -XX:MaxMetaspaceSize to limit, but fix the leak.

Problem 4: Long Pauses on 64GB Heap Scenario: A data processing service uses Parallel GC on a 64GB heap. Full GC pauses exceed 60 seconds. Changing to G1 reduces pauses but they are still >2s. Question: What should be the next step? Answer: G1 pauses scale with live data. If live data is >30GB, G1 cannot meet sub-second pauses. Consider switching to ZGC or Shenandoah, which have pause times independent of heap size.

Problem 5: Allocation Rate Spike Scenario: During a flash sale, the order service's allocation rate spikes from 100 MB/s to 2 GB/s. GC is triggered every few hundred milliseconds, CPU at 80% GC threads. Question: What is the best approach to reduce GC pressure? Answer: Optimize application code to reduce allocation (use object pooling, reuse buffers, avoid String concatenation in loops). If spikes are unavoidable, adjust heap sizing and consider ZGC for concurrent collection. Also, increase NewSize to absorb young allocation spikes.

[2026-03-05T14:23:45.123+0000] GC(52) Pause Young (Normal) (G1 Evacuation Pause) 2048M->512M(8192M) 48.123ms
[2026-03-05T14:23:45.171+0000] GC(53) Pause Young (Normal) (G1 Evacuation Pause) 2560M->1024M(8192M) 51.789ms
[2026-03-05T14:23:45.223+0000] GC(54) Pause Full (Allocation Failure) 4096M->2048M(8192M) 12345.678ms  # <-- Problem 1: full GC due to old gen exhaustion
[2026-03-05T14:23:57.568+0000] GC(55) Pause Young (Normal) (G1 Evacuation Pause) 2048M->1024M(8192M) 52.345ms
[2026-03-05T14:24:01.234+0000] GC(56) Humongous allocation of size 819200 bytes (800KB) detected. Region size 1MB, threshold 512KB.  # <-- Problem 2: humongous allocation warning

Why Objects Become Unreachable (And Why That Matters)

Every production outage I've debugged that boiled down to a GC problem started with one thing: an object that should have died but didn't. Or worse, an object that died too late.

Unreachable means zero active references. Not "I think it's done." Not "nobody should need it." Zero references on the stack or from any GC root (static fields, JNI handles, active threads). The JVM doesn't care about your intentions. It traces live references from roots outward. Everything not reached during that trace is dead.

Here's the kicker: objects can become unreachable faster than you expect. A local reference inside a method block? Gone after the method returns. An object passed to a collection that gets cleared? Eligible immediately. But the reverse is also true — a single stray reference keeps an entire object graph alive. That's how "small" memory leaks bring down production services.

Understanding reachability isn't academic. It's the difference between writing code that the GC can efficiently reclaim and code that forces full GCs every hour.

// io.thecodeforge — java tutorial

public class ReachabilityTrap {
    private static List<byte[]> leakList = new ArrayList<>();

    public static void main(String[] args) {
        // This object is local — becomes unreachable after method exit
        processData();

        // This object is held by a static reference — stays alive forever
        while (true) {
            leakList.add(new byte[1024 * 1024]); // 1 MB each
            try { Thread.sleep(100); } catch (InterruptedException e) {}
        }
    }

    static void processData() {
        byte[] temp = new byte[10 * 1024 * 1024]; // 10 MB
        // temp is the ONLY reference to this 10 MB array
        // After this method returns, temp goes out of scope
        System.out.println("Data processed");
    }
}

The Two Types of GC Activity: Minor vs. Major

You can't tune GC properly if you don't understand that garbage collection runs on two distinct modes: minor and major. They're not the same thing, and confusing them gets you fired.

Minor GC happens in the Young Generation. It's fast. The JVM stops the world, copies all live objects from Eden to a survivor space, clears Eden, and resumes. Typical pause: 1-10 milliseconds. This is your friend. A healthy application should survive on minor GCs alone for 99% of its lifetime.

Major GC (or Full GC) hits the Old Generation. This is where the JVM does mark-sweep-compact across the entire heap. Pause times balloon: 100ms, 500ms, even seconds. A full GC every few hours? Fine. Every few minutes? You have a problem — either your survivor space sizing is wrong, or you're creating long-lived objects that should be short-lived.

The critical insight: you want to avoid promoting objects to Old Generation prematurely. Each object that survives a minor GC gets an age increment. When it exceeds tenuring threshold (default 15 for G1), it's promoted. If your survivor spaces are too small, objects get promoted early, fill Old Gen, and trigger frequent full GCs.

Monitor your promotion rate. If it's higher than expected, your objects are living too long.

// io.thecodeforge — java tutorial
// Simulate different object lifetimes to see GC impact

public class PromotionAnalysis {
    private static final int OBJECT_COUNT = 1_000_000;

    public static void main(String[] args) {
        // Short-lived objects: die in Young Gen
        for (int i = 0; i < OBJECT_COUNT; i++) {
            byte[] temp = new byte[100];
        }
        System.out.println("Short-lived done — minor GC only");

        // Long-lived objects: promoted to Old Gen
        List<byte[]> holders = new ArrayList<>();
        for (int i = 0; i < OBJECT_COUNT / 10; i++) {
            holders.add(new byte[100]);
        }
        // Holders survive — these get promoted
        System.out.println("Long-lived done — full GC coming");
    }
}

Requesting GC: System.gc() Is a Hint, Not a Command

I've seen junior devs sprinkle System.gc() like seasoning. "The app's memory is high, I'll tell GC to run." Stop. Please.

System.gc() is a suggestion. The JVM can ignore it entirely. Modern collectors like G1 and ZGC often do. But even when they run it, you're paying for a full GC — and you just threw away all of the collector's adaptive sizing data. The JVM has been monitoring allocation rates, promotion patterns, and pause times to optimize future GCs. Calling System.gc() resets those heuristics. Your app will run slower for minutes afterward.

There are three legitimate reasons to call System.gc(): 1. Right before a heap dump (to minimize garbage in the dump) 2. During testing, to verify GC behavior under controlled conditions 3. After a known burst of short-lived object creation that the collector hasn't processed yet

That's it. If you think you need it in production, you almost certainly have a different problem: a memory leak, oversized heap, or wrong collector choice. Fix the real problem, don't call System.gc().

And for heaven's sake, never call System.gc() in a loop, in a request handler, or inside a timer thread. I've seen all three. Each time it caused a production incident.

// io.thecodeforge — java tutorial
// Demonstrates why System.gc() hurts performance

public class DontDoThis {
    public static void main(String[] args) {
        long start = System.nanoTime();

        // Real work
        List<Integer> data = new ArrayList<>();
        for (int i = 0; i < 10_000_000; i++) {
            data.add(i);
        }

        long mid = System.nanoTime();
        System.out.println("Work took: " + (mid - start) / 1_000_000 + " ms");

        // Production antipattern: force GC
        for (int i = 0; i < 5; i++) {
            System.gc();
        }

        long end = System.nanoTime();
        System.out.println("After GC spam: " + (end - mid) / 1_000_000 + " ms wasted");
    }
}

Customizing GC Settings in Jelastic PaaS

Jelastic PaaS exposes heap and collector flags via topology manifests and environment variables, not raw JVM arguments. You set JAVA_OPTS or use the Cloud Scripting env block to override GC strategies per node. For example, switching from G1 to Shenandoah on a production layer requires adding JAVA_OPTS=-XX:+UseShenandoahGC to the manifest. Memory limits are tied to cloudlet quotas: the heap maximum defaults to 85% of the container's RAM, which you can cap with -Xmx inside the same variable. The trap is that Jelastic's auto-tuning may silently revert your flags on node restart if you edit them via the admin panel instead of the jelastic.env file. Always commit GC changes through version-controlled manifest sections — every cloudlet restart will read from there. This prevents production surprises when horizontal scaling spawns new nodes that inherit the wrong collector.

// io.theforge — java tutorial
// (C) Jelastic manifest snippet — not runnable directly

public class JelasticGCConfig {
    // Exposed via Jelastic env variable JAVA_OPTS
    // Switch default G1 to Shenandoah:
    //      -XX:+UseShenandoahGC
    //      -Xmx2048m
    //      -XX:MaxGCPauseMillis=100
    // This is injected before main() by the platform
    public static void main(String[] args) {
        // Verify active GC at runtime:
        for (GarbageCollectorMXBean gc : ManagementFactory
            .getGarbageCollectorMXBeans()) {
            System.out.println(gc.getName());
            // Expected: Shenandoah Pauses
        }
    }
}

GC Implementations: HotSpot's Historical Lineage

The JVM isn't one GC — it's seven major implementations baked into HotSpot. The Serial collector uses a single thread for both minor and full GCs, suitable for single-core or client machines. Parallel (Throughput) GC employs multiple threads but stops-the-world for both young and old collection — good for batch jobs. G1 splits the heap into regions and predicts pause targets, becoming the default in JDK 9. ZGC uses load barriers and colored pointers to achieve sub-millisecond pauses regardless of heap size; it starts scanning live objects before pausing. Shenandoah evolved differently — it uses a Brooks pointer forwarding technique to relocate objects concurrently, even during the compaction phase. Each implementation betrays a different trade-off: Serial sacrifices throughput for footprint, Parallel sacrifices latency for throughput, and ZGC/Shenandoah sacrifices throughput for latency. Choose based on your application's tolerance for pause time versus raw processing speed.

// io.theforge — java tutorial
import java.lang.management.ManagementFactory;

public class ListGCImplementations {
    public static void main(String[] args) {
        System.out.println("Active GCs:");
        ManagementFactory.getGarbageCollectorMXBeans()
            .forEach(gc -> System.out.println("  " + gc.getName()));
        // Run with: -XX:+UseZGC
        // Output — ZGC only shows one collector name
    }
}

Overview

Garbage collection in Java is an automatic memory management process that reclaims heap space occupied by objects no longer referenced by the application. It frees developers from manual memory deallocation, preventing two critical bugs: dangling pointers and memory leaks. However, GC is not free—it consumes CPU cycles and introduces pauses (stop-the-world events). The JVM’s heap is divided into young generation (Eden, Survivor spaces) and old generation. Most objects die young; a Minor GC in Eden is cheap. Objects that survive multiple cycles get promoted to the old generation, where a Major GC (full collection) is more expensive. Understanding when and why objects become unreachable is essential: losing all strong references, circular references between unreachable objects, or references from cleared weak/soft references. The choice of GC algorithm—Serial, Parallel, G1, ZGC, or Shenandoah—depends on latency vs. throughput trade-offs. Java’s GC has evolved from a simple mark-sweep to ultra-low-pause collectors that handle terabytes of heap without freezing applications.

// io.thecodeforge — java tutorial
// 25 lines max
public class GCOverviewDemo {
    public static void main(String[] args) {
        // Object becomes unreachable after scope ends
        for (int i = 0; i < 100_000; i++) {
            String s = new String("temp");
        } // s eligible for GC here
        
        // Explicitly nulling a reference (unnecessary usually)
        Object leak = new Object();
        leak = null; // eligible now
        
        // Circular reference still collectable
        class Node { Node next; }
        Node a = new Node();
        Node b = new Node();
        a.next = b; b.next = a;
        a = null; b = null; // both collectable
        System.gc(); // hint, not guarantee
    }
}

Conclusion

Java’s garbage collection is a powerful abstraction that eliminates manual memory management, but it requires thoughtful tuning to avoid latency spikes and throughput degradation. The key insight is that GC behavior is determined by object reachability: as long as references exist from live roots (stack, static fields, JNI handles), objects remain alive. Understanding why objects become unreachable—scope exit, null assignment, weak reference clearing—lets you predict GC load. The two types of GC activity, Minor and Major, have drastically different pause profiles; optimizing object allocation rates reduces Minor GC frequency, while avoiding accidental retention prevents expensive Major collections. Modern collectors like ZGC and Shenandoah achieve sub-millisecond pauses even on multi-terabyte heaps by performing most work concurrently. However, no collector is a silver bullet: low-latency collectors trade CPU overhead for responsiveness. The future of Java GC includes generational ZGC and continued improvements to G1. Effective GC tuning starts with monitoring (GC logs, JFR), identifying pause patterns, and then adjusting flags like heap size, survivor ratio, and concurrent threads. Always test changes under realistic production load, and favor default settings from Java 17+ unless metrics prove otherwise.

// io.thecodeforge — java tutorial
// 25 lines max
import java.util.ArrayList;
import java.util.List;

public class GCTuningCheck {
    public static void main(String[] args) {
        List<byte[]> holder = new ArrayList<>();
        // Simulate accidental retention (bad)
        for (int i = 0; i < 100; i++) {
            holder.add(new byte[1024 * 1024]); // 1MB each
        }
        // Without clearing holder, objects stay reachable
        // This forces Major GC if memory tight
        System.out.println("Allocated 100 MB; clearing reference");
        holder.clear(); // now eligible for GC
        System.gc(); // hint only
    }
}