Files

T

Antigravity Agent c24165b8bc refactor(topics): 멀티 에이전트용 지식 재편 — _Common(공통 기본기) + Domain_* 구조

에이전트 8종(대화형/프로그래머 C·S/디자이너/설계자/기획자/QA/PD/PM)에게
[공통 기본 능력 + 롤별 Specialty] 2층으로 지식을 주입하기 위한 재분류.
문서 내용·포맷은 무수정, 폴더 이동만 (6,372개 문서 수 보존 확인).

- Topic_Programming → Domain_Programming (내부 구조 보존)
- Topic_Graphic → Domain_Design
- Topic_Business → Domain_Product
- Topic_General → Domain_General
- _Common 신설: Math(구 Topic_Math_Specialty), Reasoning(구 General/From_Thinking & Reasoning),
  Reasoning_Creativity(구 General/From_창의성), Communication(Poetic_Blog_Writing + From_writing)
- 타 도메인의 From_* 폴더는 유지 (출처 표기일 뿐, 이미 도메인에 맞게 분류된 문서)
- 빈 폴더 정리 (memory/procedures)
- 에이전트→폴더 매핑은 workspace의 .astra/agent-knowledge-map.json (9개 에이전트)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-07-11 11:05:56 +09:00

4.7 KiB

Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

title

Solitude Optimization

매 한 줄

"매 noisy neighbor 의 quiet 의 making". Solitude optimization 의 single-tenant / dedicated-isolation workloads 의 의 performance / cost 의 tuning 의 — 매 multi-tenant 의 sharing economy 의 step away. 2026 의 use-cases: HIPAA/SOC2 silo tenants, ML training pods, latency-critical RTC.

매 핵심

매 isolation 의 levels

Process (cgroups, Linux namespaces): 매 weak.
VM (KVM, Firecracker microVM): 매 strong, 매 ms-boot.
Bare metal: 매 strongest, 매 slowest provisioning.
Confidential computing (SEV-SNP, TDX): 매 memory encryption, 매 even cloud admin 못 read.

매 cost 의 vs noise tradeoff

pool: 매 cheapest, 매 noisy.
silo VM: 매 2-5x cost, 매 quiet + auditable.
bare metal: 매 5-10x, 매 silent + compliance-friendly.

매 응용

Top-N enterprise tenants 의 dedicated DB instance.
ML training 의 dedicated GPU node (no neighbor jitter).
Real-time audio/video 의 dedicated compute pool.

💻 패턴

Kubernetes node 의 dedicated taint

kubectl label node gpu-node-1 tenant=acme dedicated=true
kubectl taint nodes gpu-node-1 dedicated=acme:NoSchedule

# pod spec
spec:
  nodeSelector: { tenant: acme }
  tolerations:
    - key: dedicated
      operator: Equal
      value: acme
      effect: NoSchedule

CPU pinning + isolated cores

# kubelet --reserved-cpus=0-1, --cpu-manager-policy=static
spec:
  containers:
    - name: rtc
      resources:
        requests: { cpu: "4", memory: "8Gi" }
        limits:   { cpu: "4", memory: "8Gi" }

Firecracker microVM (per-tenant)

firectl --kernel ./vmlinux --root-drive ./tenant-rootfs.ext4 \
  --cpu-template T2 --vcpu-count 2 --memory 1024 \
  --tap-device tap-acme/AA:FC:00:00:00:01

Postgres 의 logical replica 의 silo upgrade

CREATE PUBLICATION acme_pub FOR TABLE invoices, users WHERE (tenant_id='acme-uuid');
-- on dedicated instance:
CREATE SUBSCRIPTION acme_sub CONNECTION '...' PUBLICATION acme_pub;

Redis — dedicated DB index per VIP tenant

const dbIdx = tenant.tier === 'enterprise' ? tenantToDb[tenant.id] : 0;
const r = new Redis({ host, port, db: dbIdx });

Network egress 의 per-tenant bandwidth shape (tc)

tc qdisc add dev eth0 root handle 1: htb default 30
tc class add dev eth0 parent 1: classid 1:1 htb rate 100mbit
tc filter add dev eth0 protocol ip parent 1:0 prio 1 \
   u32 match ip src 10.244.5.7/32 flowid 1:1

NUMA-aware 의 ML pod

apiVersion: v1
kind: Pod
spec:
  containers:
    - name: trainer
      resources:
        requests:
          cpu: "16"
          memory: "64Gi"
          nvidia.com/gpu: "1"
        limits:
          cpu: "16"
          memory: "64Gi"
          nvidia.com/gpu: "1"

매 결정 기준

상황	Isolation
HIPAA enterprise customer	silo (dedicated DB + node taint)
ML training, p99 jitter < 5ms	dedicated GPU node + CPU pin
RTC audio/video VIPs	dedicated pool, NUMA-pinned
free-tier	pool (cgroups only)

기본값: pool with QoS-Guaranteed for paid tiers, silo upgrade option for enterprise SLA.

🔗 Graph

응용: Firecracker
Adjacent: SaaS · SLO

🤖 LLM 활용

언제: tier-tradeoff explanation to sales, capacity planning, generating taint/toleration manifests. 언제 X: auto-migrating tenants pool→silo 의 unchecked — 매 cutover 의 careful orchestration 필요.

❌ 안티패턴

Silo by default: 매 cost balloon — pool 의 enough for 95% tenants.
No QoS class: BestEffort pods 의 prod 의 — 매 OOMKill victims.
Dedicated 의 sold w/o SLO uplift: 매 customer 의 perceived value 0.
Forget the data plane: CPU silo 의 했지만 shared NIC/Disk — 매 noise 여전.

🧪 검증 / 중복

Verified (Kubernetes CPU Manager, Firecracker docs, AWS Nitro/SEV-SNP, Postgres logical rep).
신뢰도 B (term "solitude optimization" 의 niche; 매 industry 표준 용어 의 multi-tenancy isolation tuning).

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — isolation/silo patterns + microVM + NUMA

4.7 KiB Raw Blame History