anti
825d7d72c9
Fleet resident set ~2.57GB across 18 workers; ~1.5GB is the 86MB import floor paid 18x. Pinned root cause: topology/__init__ eager re-export of generate drags the full SQLModel ORM (26 tables, ~38MB) into every worker.
113 lines
5.3 KiB
Markdown
113 lines
5.3 KiB
Markdown
# DECNET RAM / Process-Footprint Improvements
|
||
|
||
Status: analysis complete, implementation not started.
|
||
Measured 2026-06-17 on the dev box, 18 live `decnet` workers, CPython 3.11 (`.311`).
|
||
|
||
## Headline
|
||
|
||
Fleet resident set ≈ **2.57 GB across 18 processes**. The bulk is not workload —
|
||
it is the same import floor paid 18 times over.
|
||
|
||
## Part A — the universal import tax (measured)
|
||
|
||
Every worker pays **~86 MB at startup before doing any work**:
|
||
|
||
```
|
||
interpreter 12 MB
|
||
+ import decnet.cli 74 MB ← SQLModel/SQLAlchemy/Pydantic (~32MB)
|
||
+ EVERY decnet.web.db.models.* table
|
||
+ decnet.config + decnet.models
|
||
= floor 86 MB paid 18× ≈ 1.5 GB of the 2.57 GB total
|
||
```
|
||
|
||
Measured cold, fresh interpreter each time (RSS):
|
||
|
||
| Layer | Resident | Who pays |
|
||
|---|---:|---|
|
||
| CPython interpreter | ~12 MB | everyone (shared COW) |
|
||
| `import decnet.cli` | +74 MB | **every worker** |
|
||
| └ SQLModel/SQLAlchemy/Pydantic | ~32 MB | the ORM chain |
|
||
| └ all `decnet.web.db.models.*` tables | ~20 MB | eagerly imported |
|
||
| `scapy.all` | +76 MB | only `collect`, `probe`, `sniffer` |
|
||
|
||
Confirmed NOT in the universal path:
|
||
- **scapy** — `scapy loaded after import decnet.cli? False`. Only the sniff/probe workers pay it.
|
||
- **pandas / numpy / sklearn** — no module-scope imports anywhere; already lazy-imported
|
||
inside the functions that use them. Codebase got this right; leave it.
|
||
|
||
### Root cause
|
||
|
||
`decnet/cli/__init__.py:22-48` eagerly does `from . import (agent, api, ... ttp)` —
|
||
all 26 command modules imported at process start. Each pulls `decnet.config` +
|
||
`decnet.models` + the `decnet.web.db.models.*` chain at module top. So `decnet canary`
|
||
(which never touches TTP/swarm/webhook tables) still parses every table's SQLModel
|
||
metaclass.
|
||
|
||
importtime top offenders (pure model-table import cost, self time):
|
||
```
|
||
decnet.web.db.models.topology 21ms
|
||
decnet.web.db.models.attackers 15ms
|
||
decnet.models 13ms
|
||
decnet.web.db.models.logs 11ms
|
||
... canary, ttp, swarm, auth, webhooks, orchestrator ...
|
||
```
|
||
|
||
## Part B — architecture map (for consolidation)
|
||
|
||
All 18 workers are **already asyncio coroutines subscribing to one shared UNIX-socket
|
||
bus** (`decnet/bus/`), with a `system.{worker}.control` shutdown topic already wired and a
|
||
`system.{worker}.health` heartbeat every 10s. They are already independent tasks — nothing
|
||
needs re-architecting, only re-hosting.
|
||
|
||
| Tier | Workers | Verdict |
|
||
|---|---|---|
|
||
| Broker | `bus` | Stays alone — it's the hub. |
|
||
| Already multiprocess by design | `api`/uvicorn, `web` (ThreadingTCPServer) | Leave them. |
|
||
| scapy + blocking sniff threads | `collect`, `probe`, `sniffer` | Keep out of main loop (76 MB scapy + GIL-thrashing threads). **Merge these 3** → pay scapy once. |
|
||
| Heavy resident state / CPU | `profiler` (353 MB), `ttp` (308 MB) | Keep separate — big live heaps, real CPU work; co-locating serializes them under GIL. |
|
||
| **The idle herd** ⭐ | `webhook`, `canary`, `listener`, `forwarder`, `mutate`, `orchestrator`, `reconciler`, `enrich`, + lighter clusterers | **The prize.** ~10 mostly-idle event-driven tasks each paying the 86 MB floor to `await` a bus event. Collapse into ONE supervisor. |
|
||
|
||
Loop-type evidence (from architecture map):
|
||
|
||
| Worker | Loop entry | Loop kind |
|
||
|---|---|---|
|
||
| bus | `cli/bus.py:10` → `bus/worker.py:44` | asyncio serve_forever + heartbeat |
|
||
| profiler | `cli/profiler.py:10` → `:33` | asyncio, 30s wakeup, batch 500 logs |
|
||
| ttp | `cli/ttp.py:46` → `:80` | asyncio queue pump on `attacker.observation.*` |
|
||
| clusterer | `cli/workers.py:260` → `:304` | bus-woken on `attacker.observed` |
|
||
| campaign-clusterer | `cli/workers.py:308` → `:362` | bus-woken on `identity.>` |
|
||
| web | `cli/web.py:27` → `:148` | ThreadingTCPServer.serve_forever (blocking) |
|
||
| api | `cli/api.py:18` → `:37` | subprocess.Popen uvicorn |
|
||
|
||
## Recommendation — ordered, stop when RAM is fine
|
||
|
||
### Step 1 — Lazy command registration (do first; safe, high-leverage)
|
||
Smallest diff, zero new failure modes, helps with or without consolidation. Typer only
|
||
needs a module imported to *run* a command, not to *register* it. Defer the
|
||
`from . import (...)` so `decnet canary` loads canary's deps only, not all 26 tables.
|
||
Reversible. Expected: idle workers drop well below the 86 MB floor.
|
||
|
||
### Step 2 — Consolidate the idle herd (only if RAM still bites after step 1)
|
||
`decnet supervise` runs the idle event-driven workers as tasks in ONE process — pay the
|
||
floor 1× instead of ~10×. Plumbing already exists; the supervisor is ~10 lines:
|
||
|
||
```python
|
||
async with asyncio.TaskGroup() as tg:
|
||
for w in IDLE_WORKERS:
|
||
tg.create_task(w.run(bus)) # each already a bus-subscribed coroutine
|
||
```
|
||
|
||
**Cost to weigh:** consolidation trades RAM for **shared fate** — one crash takes down
|
||
~10 workers, one OOM kills the herd, and you lose per-worker systemd restart policy and
|
||
`MemoryMax=` caps. That's why step 1 comes first: free safety, and may make step 2
|
||
unnecessary.
|
||
|
||
### Step 3 — Merge the 3 scapy workers
|
||
Share the 76 MB scapy import once instead of 3×.
|
||
|
||
### Projected trajectory
|
||
- 2.57 GB → **~1.3 GB** from lazy imports alone (nearly free)
|
||
- → **~0.9 GB** if also consolidating the herd + merging scapy (costs isolation)
|
||
|
||
The first 1.3 GB is nearly free; the last 400 MB costs you process isolation.
|