Создание и сравнение снимков кучи

A heap snapshot is a complete dump of every JavaScript object alive in V8 at the moment you capture it. For a Node.js backend, it is the single most precise tool for answering one question: what is still being retained, and why?

• A leaking process keeps allocating objects that never get garbage-collected because something is still holding a reference.
• One snapshot tells you what exists now; comparing two snapshots over time tells you what is growing — which is the actual leak signal.

In this lesson you will capture snapshots from a running Node service, load them into Chrome DevTools, diff them, and read the retainer chain back to the offending code.

To capture snapshots from a real service you first attach the V8 inspector. Start the process with --inspect so DevTools (or the inspector protocol) can connect.

• node --inspect server.js opens the inspector on 127.0.0.1:9229.
• Open chrome://inspect in Chrome, click inspect on your target, then go to the Memory tab.
• Never bind the inspector to a public interface in production — it grants full code execution.

The snippet below is the exact CLI invocation you would script in your start command.

You cannot always open DevTools against a production box. The built-in v8 module can write a .heapsnapshot file to disk on demand, which you later load into DevTools offline.

• v8.writeHeapSnapshot(filename) serializes the entire heap synchronously.
• Trigger it from a signal handler or an internal admin route so you can grab snapshots without restarting.

This program is fully standalone — it writes a snapshot and exits.

A common production pattern is to listen for an OS signal and dump a snapshot without touching the running service. You send kill -USR2 <pid> and the process writes a file.

• SIGUSR2 is conventionally free for app use (Nodemon uses it for restarts, so pick another if you run Nodemon).
• Always force a GC consideration in mind: snapshots include unreachable objects until the next GC, so capture after the heap settles.

A single snapshot is noisy — it contains everything, including legitimate long-lived caches. The classic leak-hunting recipe is the three-snapshot technique:

• Snapshot 1 — baseline, right after warm-up.
• Exercise the suspect code path many times (e.g. hit an endpoint 1,000 times).
• Snapshot 2 — after the workload.
• Run the workload again, then take Snapshot 3.

Objects that appear in Snapshot 2 and persist into Snapshot 3 are the real leak — transient request objects will have been collected by then.

DevTools automatically runs a full GC before each snapshot, so what you see is truly reachable memory. When capturing programmatically you should do the same to avoid counting garbage that is about to disappear.

• Run Node with --expose-gc to make global.gc() available.
• Call global.gc() twice before snapshotting — the second pass cleans up objects freed during the first.

Load a .heapsnapshot into DevTools Memory tab. The default Summary view groups objects by constructor.

• Objects Count — how many live instances of that constructor exist.
• Shallow Size — memory held by the object itself, excluding what it references.
• Retained Size — memory that would be freed if this object were deleted, including everything only it keeps alive. This is the number that matters for leaks.

Sort by Retained Size descending to find the objects that dominate the heap.

The real power is the Comparison view. After loading two snapshots, switch the dropdown from Summary to Comparison and pick the baseline as the comparison base.

• #New — objects allocated since the baseline.
• #Deleted — objects collected since the baseline.
• #Delta — net change. A constructor with a large positive delta that keeps growing across diffs is your leak.
• Size Delta — net retained bytes added.

Focus on positive deltas whose count rises in lockstep with your workload iterations.

To practice the workflow, you need a deterministic leak. A module-level array (or Map) that you keep pushing into — and never clear — is the canonical example. Each request appends data that can never be collected.

• The leaked array is reachable from the module scope, so V8 must keep every entry.
• In DevTools this shows up as a growing (array) or your closure constructor with a rising delta.

This standalone program leaks on purpose and prints heap growth.

Once a suspicious constructor is found, click it to expand instances, then inspect the bottom Retainers pane. Retainers answer the only question that matters: who is holding this object alive?

• The chain reads from the object up to a GC root (the global object, a module closure, a still-pending Promise, an active timer, etc.).
• A yellow-highlighted node is reachable directly from a JS variable; red marks a detached DOM-style node (rare in Node).
• Follow the chain until you recognize your variable — that is the line of code to fix.

Closures over large scopes, unbounded caches, and EventEmitter listeners that are never removed are the usual culprits.

After patching the leak, prove it with the same three-snapshot ritual. Re-run the identical workload and diff baseline against the post-workload snapshot.

• The previously growing constructor should now show a delta near zero — allocations and collections balance out.
• Use a WeakMap or a bounded LRU cache so entries become eligible for GC when no longer referenced.
• Automate a guardrail: assert that heapUsed after GC stays under a threshold across N iterations in a soak test.

You take three heap snapshots using the standard leak-hunting technique. Which objects are the strongest evidence of a real memory leak?

You now have a complete heap-snapshot leak-hunting workflow for Node.js backends:

• Capture snapshots via --inspect + DevTools, v8.writeHeapSnapshot(), or a SIGUSR2 handler in production.
• Force GC (--expose-gc, double global.gc()) so snapshots reflect only reachable memory.
• Use the three-snapshot technique to separate real leaks from transient allocations.
• In the Comparison view, hunt rising #Delta counts; sort by Retained Size.
• Follow the retainer chain to the GC root to find the exact variable at fault — usually an unbounded cache, a lingering closure, or an unremoved listener.
• Confirm the fix with another diff showing a flat delta; prefer WeakMap or bounded caches.