.agents/skills/profile-worker-memory/SKILL.md
Never ask anyone to upload a heap snapshot. It contains connection tokens, API keys, and customer step data verbatim — a customer will refuse, and they are right to. This procedure keeps the snapshot on their box and moves only an aggregate histogram off it. That is the difference between an investigation that stalls and one that closes.
Every command below runs over SSH against the worker host.
A worker under pm2 dies and restarts without the container ever leaving Up, so docker ps and docker stats both look calm. The real tells:
docker inspect <container> --format '{{.State.OOMKilled}}|{{.RestartCount}}'
docker exec <container> pm2 list # the ↺ column is the crash count
dmesg -T | grep -aiE "Memory cgroup out of memory|Killed process"
OOMKilled=true on a running container means the cgroup OOM-killed a child while PID 1 survived. In dmesg, the process name is truncated to 15 characters: node /usr/src/a is the worker (…/worker/dist/src/bootstrap.js), not the engine — the engine runs as /usr/local/bin/node. Getting this backwards sends you profiling the wrong process.
docker exec <c> pm2 jlist # pm2_env.axm_monitor["Used Heap Size"]
docker exec <c> cat /proc/<pid>/smaps_rollup
Compare Used Heap Size against Rss. Worker baseline is roughly heap + ~100 MB flat. If RSS grows and heap grows with it, it is a JS heap leak — continue. If RSS grows while heap stays flat, it is native (isolated-vm, buffers, addons) and a heap snapshot will show you nothing.
Sample across all containers on the host before picking a target; snapshot the one with the largest heap.
Raise the cgroup first — serializing the heap allocates, and a 1 GiB container will OOM mid-snapshot. Restore it afterwards and verify.
docker update --memory 6g --memory-swap 6g <container>
# ... snapshot ...
docker update --memory 1g --memory-swap 1g <container>
docker inspect <container> --format '{{.HostConfig.Memory}}' # confirm restored
Two mechanics cost an hour each if you learn them the hard way:
sandbox-<nanoid>. Matching on /usr/local/bin/node finds the code-step child (~45 MB, its own module registry, require.cache of size 1) and tells you nothing about the pieces. Match sandbox-* in /proc/<pid>/cmdline. ps is not in the image; walk /proc directly.WebSocket cannot talk to the V8 inspector — the handshake is accepted, then the socket dies with a bare error. Use the ws the image already ships: find /usr/src/app/node_modules -type d -name ws -path "*node_modules/ws", and pass { perMessageDeflate: false, maxPayload: 0 }.Send SIGUSR1 to open the inspector, then drive CDP over 127.0.0.1:9229:
process.kill(pid, 'SIGUSR1') // opens inspector
// GET /json/list -> webSocketDebuggerUrl, then over the socket:
// HeapProfiler.enable
// HeapProfiler.collectGarbage <- so you profile survivors, not garbage
// HeapProfiler.takeHeapSnapshot
// stream HeapProfiler.addHeapSnapshotChunk events to a file
collectGarbage first is load-bearing: without it the histogram is full of collectable junk and the real retainers are buried.
A full snapshot of a ~950 MB heap kills the process mid-serialization (the cgroup can be raised; --max-old-space-size cannot — see the Gotchas on [[workers]]). If the question is "which modules are resident" rather than "who retains this", one Runtime.evaluate answers it for a few KB and cannot OOM anything:
// returnByValue: true, includeCommandLineAPI: true
const cache = process.mainModule.constructor._cache
const keys = Object.keys(cache)
// → totalModules, count matching @activepieces/shared, pieces-framework,
// distinct @activepieces/piece-* packages, and process.memoryUsage()
Run it against the sandbox-<id> pid while a flow holds the sandbox open. To get that window, end the probe flow with a CODE step that sleeps — a delay piece step over 10 s creates a waitpoint and pauses the run, releasing the sandbox, so the process you wanted is gone before you arrive.
Never parse the snapshot inside the constrained container. docker cp it out and run the parser in a throwaway fat container using the same image:
docker run --rm --memory 10g -v /tmp:/data --entrypoint node <image> \
--max-old-space-size=8192 /data/parse.js /data/heap.heapsnapshot
The format is flat typed arrays described by snapshot.meta: read node_fields for the stride, then walk nodes summing self_size grouped by strings[name]. That histogram alone usually names the leak.
A histogram tells you what is retained; only the retainer path tells you who holds it, which is the line of code you fix. Build firstEdge as a prefix sum over each node's edge_count, BFS from node 0 recording a parent per node, then walk parents back from the offending node. Print the edge names — the property names on the path are the variable names in the source.
The 2026-08 worker leak resolved to exactly this, and the last two edges were the whole diagnosis:
Object
property:/usr/src/app/cache/v13/bundles/<flowVersionId> -> Object
property:<flowVersionId> -> string "{\"flowVersion\":…" [90 MB]
A module-level object keyed by cache path, holding whole flow-bundle manifests, never evicted (cache-state.ts). See the Gotchas on [[workers]].
The snapshot never leaves the box — but its output carries the same data in miniature. Constructor histograms and retainer paths are full of real flowVersionIds, and a retained string's preview can expose flow names, step config, and timestamps. Replace them with placeholders (<flow-version-a>) before they go into a PR description, an issue, a Slack message, or a doc. This repo's PRs are public. Getting this right at the snapshot step and then pasting the raw histogram into a public PR undoes the whole point.
Remove the snapshot (it is large and it is sensitive), the helper scripts, and confirm every cgroup limit you touched is back to its original value.