Performance Smell Detection

Performance Smell Detection Skill

Identify potential code-level performance issues in Java code.

Philosophy

"Premature optimization is the root of all evil" - Donald Knuth

This skill helps you notice potential performance smells, not blindly "fix" them. Modern JVMs (Java 21/25) are highly optimized. Always:

1. Measure first - Use JMH, profilers, or production metrics
2. Focus on hot paths - 90% of time spent in 10% of code
3. Consider readability - Clear code often matters more than micro-optimizations

When to Use

• Reviewing performance-critical code paths
• Investigating measured performance issues
• Learning about Java performance patterns
• Code review with performance awareness

Scope

This skill: Code-level performance (streams, collections, objects) For database: Use jpa-patterns skill (N+1, lazy loading, pagination) For architecture: Use architecture-review skill

Quick Reference: Potential Smells

String Operations (Java 9+ / 21 / 25)

What Changed

Since Java 9 (JEP 280), string concatenation with + uses invokedynamic, not StringBuilder. The JVM optimizes simple concatenation well.

Java 25 adds String::hashCode constant folding for additional optimization in Map lookups with String keys.

Still Valid: StringBuilder in Loops

javaCopy code
// 🔴 Still problematic - new String each iteration
String result = "";
for (String s : items) {
    result += s;  // O(n²) - creates n strings
}

// ✅ StringBuilder for loops
StringBuilder sb = new StringBuilder();
for (String s : items) {
    sb.append(s);
}
String result = sb.toString();

// ✅ Or use String.join / Collectors.joining
String result = String.join("", items);

Now Fine: Simple Concatenation

javaCopy code
// ✅ Fine in Java 9+ - JVM optimizes this
String message = "User " + name + " logged in at " + timestamp;

// ✅ Also fine
return "Error: " + code + " - " + description;

Avoid in Hot Paths: String.format

javaCopy code
// 🟡 String.format has parsing overhead
log.debug(String.format("Processing %s with id %d", name, id));

// ✅ Parameterized logging (SLF4J)
log.debug("Processing {} with id {}", name, id);

Stream API (Nuanced View)

The Reality

Streams have overhead, but it's often acceptable:

• < 100 items: Streams can be 2-5x slower (but still microseconds)
• 1K-10K items: Difference narrows significantly
• > 10K items: Often within 50% of loops
• GraalVM: Can optimize streams to match loops

Recommendation: Prefer streams for readability. Optimize to loops only when profiling shows a bottleneck.

When Streams Are Problematic

javaCopy code
// 🔴 Stream created per iteration in hot loop
for (int i = 0; i < 1_000_000; i++) {
    boolean found = items.stream()
        .anyMatch(item -> item.getId() == i);
}

// ✅ Pre-compute lookup structure
Set<Integer> itemIds = items.stream()
    .map(Item::getId)
    .collect(Collectors.toSet());

for (int i = 0; i < 1_000_000; i++) {
    boolean found = itemIds.contains(i);
}

When Streams Are Fine

javaCopy code
// ✅ Single pass, readable, not in tight loop
List<String> names = users.stream()
    .filter(User::isActive)
    .map(User::getName)
    .sorted()
    .collect(Collectors.toList());

// ✅ Primitive streams avoid boxing
int sum = numbers.stream()
    .mapToInt(Integer::intValue)
    .sum();

Parallel Streams: Use Carefully

javaCopy code
// 🔴 Parallel on small collection - overhead > benefit
smallList.parallelStream().map(...);  // < 10K items

// 🔴 Parallel with shared mutable state
List<String> results = new ArrayList<>();
items.parallelStream()
    .forEach(results::add);  // Race condition!

// ✅ Parallel for CPU-intensive + large collections
List<Result> results = largeDataset.parallelStream()  // > 10K items
    .map(this::expensiveCpuComputation)
    .collect(Collectors.toList());

Boxing/Unboxing

Still a Real Issue

Boxing creates objects on heap, adds GC pressure. JVM caches small values (-128 to 127) but not larger ones.

Future: Project Valhalla will improve this significantly.

javaCopy code
// 🔴 Boxing in tight loop - creates millions of objects
Long sum = 0L;
for (int i = 0; i < 1_000_000; i++) {
    sum += i;  // Unbox, add, box
}

// ✅ Primitive
long sum = 0L;
for (int i = 0; i < 1_000_000; i++) {
    sum += i;
}

Use Primitive Streams

javaCopy code
// 🟡 Boxing overhead
int sum = list.stream()
    .reduce(0, Integer::sum);

// ✅ Primitive stream
int sum = list.stream()
    .mapToInt(Integer::intValue)
    .sum();

Regex

Always Pre-compile in Loops

This advice is not outdated - Pattern.compile is expensive.

javaCopy code
// 🔴 Compiles pattern every iteration
for (String input : inputs) {
    if (input.matches("\\d{3}-\\d{4}")) {  // Compiles regex!
        process(input);
    }
}

// ✅ Pre-compile
private static final Pattern PHONE = Pattern.compile("\\d{3}-\\d{4}");

for (String input : inputs) {
    if (PHONE.matcher(input).matches()) {
        process(input);
    }
}

Collections

Capacity Hint (Minor Optimization)

javaCopy code
// 🟢 Low severity - but free optimization if size known
List<User> users = new ArrayList<>(expectedSize);
Map<String, User> map = new HashMap<>(expectedSize * 4 / 3 + 1);

Right Collection for the Job

javaCopy code
// 🟡 O(n) lookup in loop
List<String> allowed = getAllowed();
for (Request r : requests) {
    if (allowed.contains(r.getId())) { }  // O(n) each time
}

// ✅ O(1) lookup
Set<String> allowed = new HashSet<>(getAllowed());
for (Request r : requests) {
    if (allowed.contains(r.getId())) { }  // O(1)
}

Unbounded Collections

javaCopy code
// 🔴 Memory risk - could grow unbounded
@GetMapping("/users")
public List<User> getAllUsers() {
    return userRepository.findAll();  // Millions of rows?
}

// ✅ Pagination
@GetMapping("/users")
public Page<User> getUsers(Pageable pageable) {
    return userRepository.findAll(pageable);
}

Modern Java (21/25) Patterns

Virtual Threads for I/O (Java 21+)

javaCopy code
// 🟡 Traditional thread pool for I/O - wastes OS threads
ExecutorService executor = Executors.newFixedThreadPool(100);
for (Request request : requests) {
    executor.submit(() -> callExternalApi(request));  // Blocks OS thread
}

// ✅ Virtual threads - millions of concurrent I/O operations
try (ExecutorService executor = Executors.newVirtualThreadPerTaskExecutor()) {
    for (Request request : requests) {
        executor.submit(() -> callExternalApi(request));
    }
}

Structured Concurrency (Java 21+ Preview)

javaCopy code
// ✅ Structured concurrency for parallel I/O
try (StructuredTaskScope.ShutdownOnFailure scope = new StructuredTaskScope.ShutdownOnFailure()) {
    Future<User> user = scope.fork(() -> fetchUser(id));
    Future<Orders> orders = scope.fork(() -> fetchOrders(id));

    scope.join();
    scope.throwIfFailed();

    return new UserProfile(user.resultNow(), orders.resultNow());
}

Performance Review Checklist

🔴 High Severity (Usually Worth Fixing)

• Regex Pattern.compile in loops
• Unbounded queries without pagination
• String concatenation in loops (StringBuilder still valid)
• Parallel streams with shared mutable state

🟡 Medium Severity (Measure First)

• Streams in tight loops (>100K iterations)
• Boxing in hot paths
• List.contains() in loops (use Set)
• Traditional threads for I/O (consider Virtual Threads)

🟢 Low Severity (Nice to Have)

• Collection initial capacity
• Minor stream optimizations
• toArray(new T[0]) vs toArray(new T[size])

When NOT to Optimize

• Not a hot path - Setup code, config, admin endpoints
• No measured problem - "Looks slow" is not a measurement
• Readability suffers - Clear code > micro-optimization
• Small collections - 100 items processed in microseconds anyway

Analysis Commands

bashCopy code
# Find regex in loops (potential compile overhead)
grep -rn "\.matches(\|\.split(" --include="*.java"

# Find potential boxing (Long/Integer as variables)
grep -rn "Long\s\|Integer\s\|Double\s" --include="*.java" | grep "= 0\|+="

# Find ArrayList without capacity
grep -rn "new ArrayList<>()" --include="*.java"

# Find findAll without pagination
grep -rn "findAll()" --include="*.java"