Deco Site Memory Debugging

Deco Site Memory Debugging

Two runtime targets, two different diagnostic stories — pick the section that matches how the site is deployed.

Superseded content. This skill previously assumed Deno-on-Kubernetes (Deno.memoryUsage(), CDP over kubectl port-forward into a pod's port 9229, Deno-specific V8 inspector quirks). That deployment model does not exist for current blocks sites — neither Cloudflare Workers nor Node/RSC exposes Deno.memoryUsage() or lives in a kubectl-managed pod. cdp-connection.md and memory-analysis.md are now written for Node/RSC (@decocms/nextjs) specifically; if you're debugging a legacy Deno/Fresh/Knative site outside this package split, pull the Deno-flow version of this skill from git history instead of following the current files.

When to Use This Skill

• A Cloudflare Workers site is hitting exceededMemory / exceededCpu outcomes (128 MB / CPU-time isolate limits)
• A Node/RSC site's process memory is high or growing over time
• Need to identify what's consuming memory inside a Node process
• Investigating Response/Request body leaks
• Need to determine if memory is a real leak vs lazy GC (Node target only — Workers isolates don't live long enough for "lazy GC accumulation" to be the diagnosis)

Quick Start

Cloudflare Workers (@decocms/tanstack) — Part 1, no live debugger:
  1. QUERY TAIL CAPTURE   → ClickHouse otel_logs, _source='tail-worker', _outcome IN (exceededMemory, exceededCpu)
  2. STREAM LIVE          → wrangler tail <worker-name> --format pretty
  3. REPRO LOCALLY        → wrangler dev, attach via chrome://inspect (Chrome-shaped CDP)
  4. FIX AT THE CODE LEVEL → no --max-old-space-size knob exists on Workers; reduce actual allocations

Node/RSC (@decocms/nextjs) — Part 2, real CDP flow:
  1. CONNECT INSPECTOR    → node --inspect / kill -USR1 <pid>
  2. GET WS URL           → curl http://127.0.0.1:9229/json/list
  3. CONNECT CDP          → WebSocket to the debuggerUrl
  4. FORCE GC             → HeapProfiler.collectGarbage
  5. ANALYZE MEMORY       → process.memoryUsage() + queryObjects + heap snapshot
  6. DIAGNOSE             → Is it a leak or lazy GC?
  7. RECOMMEND            → Tune NODE_OPTIONS=--max-old-space-size or fix leak

Files in This Skill

Part 1: Cloudflare Workers (@decocms/tanstack)

Why there's no live-heap CDP flow in production

Workers isolates are ephemeral and horizontally distributed — there is no single addressable "pod" to kubectl port-forward into, and no long-running process to attach a debugger to mid-incident. Memory is hard-capped at 128 MB per isolate; when an isolate exceeds it, Cloudflare kills the isolate before any in-Worker code — including a would-be CDP responder — can run. This is a fundamentally different failure mode from a long-running Deno process slowly accumulating garbage.

What you actually get

1. exceededMemory / exceededCpu outcomes, captured 100% by the tail worker (deco-otel-tail) and landing in ClickHouse otel_logs with Attributes['_source'] = 'tail-worker'. See docs/observability.md, "Error capture — three-channel model".

   SELECT Timestamp, ServiceName,
          Attributes['_outcome'] AS outcome,
          Body
   FROM otel_logs
   WHERE ServiceName = '<site-worker-name>'
     AND Attributes['_source'] = 'tail-worker'
     AND Attributes['_outcome'] IN ('exceededMemory', 'exceededCpu')
     AND Timestamp > now() - INTERVAL 24 HOUR
   ORDER BY Timestamp DESC
   LIMIT 50;
   

   Requires the site to have adopted the tail worker per docs/tail-worker-recipe.md (check tail_consumers in the site's wrangler.jsonc first). If it hasn't, this query returns nothing — either onboard the recipe or fall back to the Cloudflare dashboard's own per-Worker Metrics panel, which shows exceeded-limit invocation counts without the tail worker.

2. wrangler tail <worker-name> --format pretty for live streaming during an active incident. Doesn't show history, but useful for reproducing the failure live against real traffic shape.

3. Cloudflare dashboard → Workers & Pages → <worker> → Metrics shows CPU-time percentiles and exceeded-limit counts natively.

Diagnosing WHY, not just THAT

You can't attach a debugger to the isolate that actually failed, so root-causing exceededMemory means code-level investigation rather than live inspection — but the smells to look for are the same ones the old Deno methodology targeted:

• Buffered instead of streamed bodies — await res.arrayBuffer() / .text() on a large upstream response instead of streaming or bounding it. Same "Response body leak" class of bug as before, different consequence (128 MB hard kill instead of slow RSS growth).
• Unbounded per-isolate caches — isolates are reused across several requests, so an in-memory Map/array that's never evicted grows request-over-request within one isolate's lifetime. Check any hand-rolled cache in site code for a missing eviction policy.
• Large synchronous JSON.parse of full multi-MB commerce API responses (VTEX catalog/search payloads in particular) instead of extracting only the needed fields.
• Local repro: run the same request pattern against wrangler dev, which runs a real workerd isolate under Miniflare and enforces the same 128 MB limit. wrangler dev exposes a real inspector — open chrome://inspect (or pass --inspect) and attach. This is Chrome-shaped CDP, the opposite of the old Deno quirk: Runtime.queryObjects returns its array under result.result, not result.objects.
• No --max-old-space-size knob. The 128 MB ceiling is a platform limit, not a tunable V8 flag on Workers — there is no "lower the GC threshold" fix here. The only fix is reducing what's actually allocated.

Part 2: Node/RSC (@decocms/nextjs)

If the site's Node process runs as a normal long-lived server (not serverless/edge), the old Deno flow's underlying methodology still holds — force GC, compare before/after, take a heap snapshot if memory is really retained — but every concrete command changes. See cdp-connection.md for the connection flow and memory-analysis.md for the analysis procedures, both rewritten for Node's inspector and process.memoryUsage() instead of Deno's.

This repo does not prescribe a hosting/orchestration layer for @decocms/nextjs sites — if the Node process runs inside a container or orchestrator, forward its inspector port (9229 by default) the way that infra provides today (docker exec + published port, an SSH tunnel, or your PaaS's own port-forwarding CLI). There is no kubectl-specific step here because blocks itself makes no Kubernetes assumption for the Node/RSC target.

Key Concept: GC is Lazy (Node/RSC target)

V8's garbage collector is lazy by design. It won't collect garbage until memory pressure forces it to. A Node process showing 1.8 GB RSS might drop to 700 MB after a forced GC — meaning there was no leak, just uncollected garbage. (This is the one piece of the old methodology that has no Workers-side analog — see Part 1's "no --max-old-space-size knob" note.)

Always force GC before concluding there's a leak:

HeapProfiler.collectGarbage  (via CDP)

Then check process.memoryUsage() again. The difference between before-GC and after-GC tells you how much was reclaimable garbage vs actual retained memory.

Recommendation: Reduce Max Heap Size

If post-GC memory is reasonable but pre-GC memory is causing OOM kills or high process memory:

Decrease the V8 max old space size so GC runs more frequently:

NODE_OPTIONS=--max-old-space-size=512

This forces V8 to GC more aggressively instead of letting garbage accumulate. For most Deco sites this doesn't affect performance because the actual live heap is much smaller than the default limit. The GC runs are fast (milliseconds) and the trade-off is worth it to keep RSS predictable.

Example:

NODE_OPTIONS=--max-old-space-size=512 node server.js

Memory Breakdown Model (Node/RSC target)

RSS is composed of multiple layers. You must understand what each layer represents:

RSS = V8 Heap + V8 External (incl. ArrayBuffers) + Native (untracked)

V8 Heap     → JavaScript objects, closures, strings, compiled code
V8 External → ArrayBuffers, Buffers, TypedArrays
Native      → Node's own C++ runtime, libuv, JIT code cache, mmap'd files, thread stacks

Use process.memoryUsage() to get the breakdown:

• rss: Total resident set size
• heapUsed: V8 JavaScript heap
• heapTotal: V8 heap capacity
• external: V8 external allocations
• arrayBuffers: the subset of external that's ArrayBuffer/Buffer — Node reports this as its own field, unlike Deno which folds it into external only
• rss - heapUsed - external = native/untracked memory

The native gap is normal. For a large Next.js/RSC app with thousands of modules, several hundred MB of native memory is expected (JIT compiled code, libuv, module cache). This is NOT a leak by itself.

Common Memory Consumers

Diagnostic Decision Tree (Node/RSC target)

Process memory high?
├── Force GC → Memory drops significantly?
│   ├── YES → Not a leak. Recommend reducing NODE_OPTIONS=--max-old-space-size
│   └── NO → Real retained memory. Continue investigation:
│       ├── Check Response objects (queryObjects Response.prototype)
│       │   └── bodyUsed=false count high? → Response body leak
│       ├── Check ArrayBuffers
│       │   └── Many large OTEL/JSON buffers? → Export/cache leak
│       ├── Check heap snapshot top consumers
│       │   └── Large HTML/RSC strings? → SSR cache unbounded
│       └── Large native gap (RSS - heap - external)?
│           └── Normal for large Node apps (JIT + libuv); only a concern if it grows unboundedly over time

For the Cloudflare Workers target, there is no equivalent decision tree to walk through against a live process — see Part 1's "Diagnosing WHY, not just THAT" instead, which substitutes code-review of the same failure classes for live inspection.

Related Skills

• deco-site-scaling-tuning — now scoped to legacy Deno/Kubernetes sites only; no current scaling-tuning equivalent exists for Cloudflare Workers or Node/RSC (see that skill's note)
• deco-incident-debugging — for general incident response and triage
• deco-performance-audit — for deep performance analysis