데이터 티어링 회의록

---

1. 회의 배경

1.1 현재 상황

단일 Tier (Hot Only)

고객사 설명:

"They don't have tiering right now, they just have one cluster."

현재 구성:

• Pool 0: 54 nodes (NVMe)
• 용량: 3 PB (사용 중)
• Tier: Hot only (단일)

향후 계획

고객사 계획:

"going forward, they will try to use whatever they have in the 
back... like [SATA] SSDs or whatever, to have a warm tier... 
sometime in this year, like, probably... second half."

타임라인:

• 시기: 2026년 하반기 (예상)
• 대상: Legacy SATA SSD
• 목적: Warm tier 구축

1.2 요구사항

고객사 질문:

"they will have this one massive NVMe cluster[s], hot tier... 
then, like, what are the things that they need to consider for 
warm tier? Like, how should they be connected? And what are the 
[lifecycle] management?"

핵심 질문:
1. Warm Tier 구성: 어떤 하드웨어?
2. 연결 방법: Hot ↔ Warm 통합?
3. Lifecycle Management: 자동 이동?
4. 데이터 선택: 무엇을 Warm으로?

---

2. Tiering 개념

2.1 Hot-Warm-Cold 정의

Hot Tier:

• Storage: NVMe SSD
• 성능: 매우 빠름 (μs latency)
• 용도: 자주 액세스 (Real-time analytics)
• 비용: 매우 비쌈
• 예시: 최근 1개월 데이터

Warm Tier:

• Storage: SATA SSD
• 성능: 빠름 (ms latency)
• 용도: 주기적 액세스 (Weekly reports)
• 비용: 중간
• 예시: 1~12개월 데이터

Cold Tier:

• Storage: HDD 또는 Tape
• 성능: 느림 (10+ ms latency)
• 용도: 드문 액세스 (Archive, Compliance)
• 비용: 저렴
• 예시: 12개월+ 데이터

2.2 현재 vs 목표

현재 (Hot Only):

All Data (3 PB)
  ↓
NVMe Pool 0 (54 nodes)
  ↓
고비용, 고성능

목표 (Hot + Warm):

Recent Data (500 GB)
  ↓
Hot: NVMe Pool 0 (54 nodes)

Older Data (2.5 PB)
  ↓
Warm: SATA SSD Pool (Legacy hardware)

비용 절감 + 성능 유지

---

3. Tiering 전략

3.1 데이터 분류

접근 빈도

벤더 질문:

"Is it data that you still be running through analytics, or is 
it just going to be... less frequently accessed?"

분류 기준:

Hot (자주 액세스):

• Real-time Analytics: 매일 쿼리
• Dashboard: 실시간 업데이트
• ML Training: 현재 모델 학습
• Active ETL: 진행 중 파이프라인

Warm (가끔 액세스):

• Historical Reports: 주간/월간
• Compliance: 규제 보관
• Archive: 참조용
• Completed Projects: 종료된 분석

예시:

-- Hot: 최근 1개월 매출 데이터
SELECT * FROM sales 
WHERE date >= '2026-01-01';

-- Warm: 작년 매출 (월별 리포트)
SELECT SUM(amount) 
FROM sales 
WHERE YEAR(date) = 2025 
GROUP BY MONTH(date);

성능 요구사항

벤더 경고:

"these are enterprise SSDs... you [will] notice the difference 
in access speed between [NVMe] and SSD. So... the data as you 
let it go to warm, should be data that you don't want to run 
analytics on top of anymore"

NVMe vs SATA SSD:

항목	NVMe	SATA SSD
Latency	10-20 μs	50-100 μs
IOPS	500K+	100K
Throughput	3-7 GB/s	500 MB/s
Use Case	Real-time	Batch

주의:

• ✅ Warm: Batch 분석 OK
• ❌ Warm: Real-time 분석 불가

3.2 시간 기반 정책

예시 정책:

정책 1: 30일 기준

Day 0-30: Hot (NVMe)
Day 31+: Warm (SATA SSD)

정책 2: 90일 기준

Day 0-90: Hot (NVMe)
Day 91-365: Warm (SATA SSD)
Day 366+: Cold (Delete or Archive)

정책 3: 액세스 기반

Last Access < 7 days: Hot
Last Access 7-90 days: Warm
Last Access > 90 days: Cold

---

4. 부록: 주요 인용문

Tiering 계획

"going forward, they will try to use whatever they have in the 
back... like [SATA] SSDs or whatever, to have a warm tier... 
sometime in this year, like, probably... second half."

요구사항

"they will have this one massive NVMe cluster[s], hot tier... 
then, like, what are the things that they need to consider for 
warm tier? Like, how should they be connected? And what are the 
[lifecycle] management?"

성능 고려

"these are enterprise SSDs... you [will] notice the difference 
in access speed between [NVMe] and SSD. So... the data as you 
let it go to warm, should be data that you don't want to run 
analytics on top of anymore"

ILM 상태

"ILM and replication is not being used as of now... but they 
have... [a] goal of... activating [it]."