staff-engineering-skills-memory-leaks

Memory Leaks Trap

It works fine in testing. After a week in production, the OOM killer visits. Before writing any code that accumulates state, ask: what removes entries from this, and what limits its size?

How Memory Leaks Happen in GC Languages

In JavaScript and Go, leaks aren't about forgetting free(). They're about holding references you no longer need. The GC only reclaims unreachable memory; a reference held through a Map, closure, listener, or goroutine pins that memory forever.

The danger is the time dimension. A 200-byte-per-request leak is invisible in a 100-request test. At 1,000 req/s, it's 17GB per day.

The Five Leak Patterns

Detection: When You're Writing a Leak

Stop and fix if you see:

1. new Map()/new Set()/[] at module scope with .set()/.push() but no removal -- if nothing limits its size or evicts old entries, it grows until OOM.

2. .on("event", handler) or addEventListener without matching .off()/removeEventListener -- every call adds a listener; per-request or per-connection, they accumulate. Node warns with MaxListenersExceededWarning -- never suppress it.

3. setInterval without clearInterval -- the timer runs forever and pins any objects its callback references. Clear it when the owning resource is destroyed.

4. A closure passed to a long-lived callback (scheduler, event bus, queue) -- check what it captures. It may pin the entire request context or a large dataset even if it uses one field.

5. go func() reading a channel or waiting on a lock without a context/done channel -- if the channel never closes and the context never cancels, it blocks forever. Each leaked goroutine holds 8KB+ of stack plus referenced data.

6. defer inside a loop (Go) -- defer runs when the function returns, not the iteration. In a 100,000-item loop, all deferred calls accumulate until the function exits.

Patterns

Bounded cache with LRU eviction

Never use a plain Map as a cache. Always use LRU (or similar) with a max size and ttl so the memory ceiling is predictable.

import { LRUCache } from "lru-cache";

const userCache = new LRUCache<string, User>({ max: 10_000, ttl: 5 * 60 * 1000 });

export async function getUser(id: string): Promise<User> {
  const cached = userCache.get(id);
  if (cached) return cached;
  const user = await db.users.findUnique({ where: { id } });
  userCache.set(id, user);
  return user;
}

Event listener cleanup

Every .on() needs a corresponding .off() tied to the lifecycle of the resource that owns the listener. With modern APIs, an AbortController signal removes all listeners tied to it on abort():

export function handleWebSocket(ws: WebSocket, userId: string) {
  const ac = new AbortController();
  eventBus.on("notification", (event) => {
    if (event.userId === userId) ws.send(JSON.stringify(event));
  }, { signal: ac.signal });
  ws.on("close", () => ac.abort()); // removes the listener
}

Closures that don't capture more than they need

// BAD: closure captures largeDataset (500MB) even though it only uses summary
async function processReport(reportId: string) {
  const largeDataset = await fetchDataset(reportId);
  const summary = summarize(largeDataset);
  scheduler.onComplete(reportId, () => notifyUser(`Report ${reportId}: ${summary}`));
  // largeDataset is pinned by the closure's scope -- can't be GC'd
}

// GOOD: extract needed values before creating the closure
async function processReport(reportId: string) {
  const largeDataset = await fetchDataset(reportId);
  const message = `Report ${reportId}: ${summarize(largeDataset)}`;
  scheduler.onComplete(reportId, () => notifyUser(message));
  // largeDataset is now eligible for GC
}

WeakMap for metadata on transient objects

Use WeakMap to attach metadata to objects whose lifecycle you don't control. Unlike Map, it doesn't prevent GC of its keys, so entries vanish automatically with no manual cleanup.

const requestMetadata = new WeakMap<Request, { startTime: number; traceId: string }>();

export function trackRequest(req: Request) {
  requestMetadata.set(req, { startTime: Date.now(), traceId: crypto.randomUUID() });
}
// Entry removed automatically when the Request is GC'd after the response.

Timer cleanup

export function startHeartbeat(connection: Connection) {
  const timer = setInterval(() => connection.ping(), 30_000);
  connection.on("close", () => clearInterval(timer));
  process.on("SIGTERM", () => clearInterval(timer)); // also clear on shutdown
}

Go: goroutine lifecycle with context

Every goroutine needs an exit path -- channel closed, context cancelled, or both. Without one, it's a leak.

func watchForUpdates(ctx context.Context, ch <-chan Update, handler func(Update)) {
    go func() {
        for {
            select {
            case update, ok := <-ch:
                if !ok { return } // channel closed
                handler(update)
            case <-ctx.Done():
                return // context cancelled
            }
        }
    }()
}

Go: defer in loops -- extract to a function

// BAD: all file handles stay open until processFiles returns
func processFiles(paths []string) error {
    for _, path := range paths {
        f, _ := os.Open(path)
        defer f.Close() // accumulates!
    }
    return nil
}

// GOOD: defer runs per-iteration because it lives in its own function scope
func processFiles(paths []string) error {
    for _, path := range paths {
        if err := processFile(path); err != nil { return err }
    }
    return nil
}

func processFile(path string) error {
    f, err := os.Open(path)
    if err != nil { return err }
    defer f.Close()
    return nil
}

The Container Restart Mask

Containers mask leaks. A 20MB/hour leak in a 512MB pod OOM-restarts every ~24 hours -- teams see one restart a day and call it "container flakiness." It persists for months until a traffic increase makes it OOM every 6 hours and someone investigates.

Periodic restarts at a regular interval = suspect a memory leak. Plot memory over time -- a linear upward trend is the signature.

Anti-Patterns

// Unbounded Map: grows until OOM
const cache = new Map<string, any>();
function getItem(key) { if (!cache.has(key)) cache.set(key, fetchItem(key)); }

// Listener leak: new listener per connection, never removed
eventBus.on("event", (data) => ws.send(data)); // in a per-connection handler

// Timer leak: no clearInterval anywhere
setInterval(() => refreshData(), 60_000);

// Closure capture: pins 500MB dataset via scope
const data = await fetchLargeDataset();
const summary = summarize(data);
scheduler.later(() => log(`Done: ${summary}`)); // data is pinned

Related Traps

• Cache Invalidation -- unbounded caches are the most common source of memory leaks. Every cache needs a max size and eviction policy. See the Cache Invalidation skill.
• Backpressure -- unbounded buffers (queues, arrays) that grow because the consumer can't keep up are both a backpressure failure and a memory leak. Apply backpressure before the buffer exhausts memory.
• Thundering Herd -- when an OOM-killed process restarts with cold caches, the sudden spike in cache misses can cause a thundering herd on the database.