벤더 설명:
"health Check and everything is more generic, like once a month,
like once a quarter thing, whereas MC Browse can be used when
there's a failure, or they're doing upgrade, or, like those
edge cases, like, it's an easy monitoring to pool."
회의 내용:
"So they ran couple of the scripts that we suggested.
And of course, like, they're most interested in performance for us."
발견 과정:
"We will critical aside from that, yeah, transparent, huge pages
being the OS level... So the top one there, I see that almost
all the time. Within our own cluster, we have the red flag
every time... the transparent, huge pages that is also a
performance issue."
질문: THP는 OS 레벨인가, 컨테이너 레벨인가?
벤더 답변:
"That's what I thought, because we're setting the transparent
huge pages, is that the OS level? So either it's not set
correctly at the OS level or the container, because that
setting for transparent huge pages could be defined at any time"
컨테이너 내부 확인
"We can always go into the container and try to see what
the computer sees... I don't know which profile they're
reading these value out"
OS 레벨 확인
기술 논의:
"For some reason that one note, it's getting caught like
it's not sticking... if you restarted just the container
on that node, would it pick that up? Typically, there's no..."
벤더 의견:
"I would say it's probably just it can be ignored...
if the rest of it's [normal]"
회의 내용:
"어떻게 진단이 되는지 내부적으로 확인 예정"
회의 내용:
"(SK) IOMMU 설정은 인프라에서 적용해도 무방하다는 의견을
받았음 (다만 노드 재기동 필요)"
회의 중 관찰:
"Go routines could mean that this particular node, there's
a process stock allocating Go routines... you see, I mean,
we were not able to find a CPU widget on your graph, but
you see memory is not spiking... the Go routine, it spiked,
and then it stayed."
"whatever happened at nine, yeah, something clearly at
nine triggered, yeah, and then it stopped and kept stable.
So it's definitely, could be a variety of things, but
obviously something at 915, some schedule, process,
some product started"
벤더 설명:
"in Go, Go routines are super cheap, yeah, very, very cheap
to cheaper than threads. So in the Go run time, every time
you start go routine, it's just kind of like an allocation
in memory. Just take some bytes"
긍정적 지표 (문제 없음):
1. 메모리 사용량: 급증하지 않음
2. CPU 사용량: 스파이크 없음
3. 안정적 유지: 추가 증가 없이 일정 수준 유지
회의 중 분석:
"Memory didn't grow. So it's not like an ongoing issue.
So memory is not skyrocketing. In go you can spawn 10 million
Go routines, and they're quite cheap."
부정적 지표 (주의 필요):
1. 지속적 높은 수준: 원래대로 돌아가지 않음
2. 원인 불명: 무엇이 트리거했는지 불확실
3. 잠재적 누수: 시간 경과 시 문제 발생 가능성
벤더 평가:
"So it's not like it's not like hot, right when it will
span it out of control, but whatever that server decided
to do, the spike under Go routines, and then it just
[stayed down]. So I don't think it's an ongoing problem per se."
실행 방법:
mc admin profile [alias] --type goroutine
벤더 설명:
"so if you see, look at the interesting circle, it just
profiles 10 seconds, right? So the default is 15 seconds,
because profile kind of like taxes the system. It slows
down... now that file, if you upload it, subnet engineers
can actually tell you what is the reason"
주의사항:
"profile kind of like taxes the system. It slows down"
→ 프로파일링 자체가 시스템에 부하를 줌
기술적 설명:
"if each one of them was doing something, it would be
concerning. So that's why CPU memory correlation, it shows
like there was no [spike] and even that metric you could
see at nine, there was like a spike"
MC Admin Profile 실행
결과 파일 벤더 전달
트렌드 관찰
성능 지표 확인
고객사 보고:
"identical error took place in four days, and normally
we would put it to subnet, but now that we're have
experts here, maybe [we can investigate]"
"it's not syncing with your policies. It's not syncing
IDP correction using LDAP or OIDC"
설명:
중요성:
"the only thing I can think of is there a Clock drift?
Clock drift between nodes, between the LDAP, the LDAP,
is there a clock drift between the LDAP server and
the cluster... authentication? They have to be in the
same time, right?"
기술적 배경:
벤더 확인:
"NTP issue, right? Yep."
설명:
벤더 질문:
"It could just be a timeout from that pool. You know,
what's your what's your sync rate?"
질문:
"or did your LDAP change its TLS, its TLS?"
설명:
과거 경험:
"if it's A bind issue that we spent hours debugging"
설명:
고객사 현황:
"when you're testing LDAP connections, are you always
doing it through the browser? [Yes,] normally, web browser"
한계점:
벤더 제공 예정:
"we have a script we can give you that'll let you test
that from command line as well. That grabs the token
and the secret, so you can actually test that as well
without having to go in."
스크립트 기능:
장점:
"[for] applications are going to be coming in using
the shared Token Service... Typical program."
NTP 동기화 상태
네트워크 연결성
TLS/SSL 설정
LDAP 서버 로그
동기화 설정
커맨드 라인 스크립트 적용
Clock Drift 확인
네트워크 진단
모니터링 강화
설정 최적화
업그레이드 배경:
"because there are so many things that's upgraded and
the UI is like, UX is completely different... I think
the reason that MC Bros is upgraded is because there
were some requests or customer pieces that we saw,
you know, that enterprise will be able to help."
개선 영역:
1. 모든 노드 및 디스크 모니터링
"remember like there was a new feature in MC AI store,
MC that you can literally monitor, like all the nodes
and disks"
글자 하나하나 모니터링 가능
성능 편차 확인 용이
엔터프라이즈 고객 요청 반영:
"엔터프라이즈 고객의 요청으로 MC의 진단 관련 UI/UX가
대폭 개선되었으며, 이를 통해 성능 편차를 확인하거나
모니터링하는 것이 훨씬 쉬워졌음"
단계:
1. 업그레이드 전 상태 스냅샷
2. 업그레이드 진행
3. MC Browse로 실시간 모니터링
4. 이상 징후 즉시 탐지
5. 업그레이드 후 검증
단계:
1. 기존 Pool 상태 확인
2. 새 노드 추가
3. 리밸런싱 모니터링
4. 성능 균형 확인
단계:
1. 장애 노드 식별
2. 상세 로그 확인
3. 드라이브 상태 점검
4. 복구 작업 모니터링
벤더 계획:
"give them a couple of scenarios, guidances, like,
they upgrade, what do they see? What do they do for
like, stuff like that, or even fully expansion...
we can, like, just drop down those things and share
it with them. Doesn't have to be done this week"
제공 예정 자료:
1. 시나리오별 사용 가이드
2. 화면 설명 및 해석 방법
3. 일반적인 패턴 및 이상 징후
4. 고객 사례 기반 Best Practice
회의 중 언급:
"개선된 버전은 조만간 Release될 예정임"
기술적 해석:
"the average is consistent with the 40 gigabit...
it's strange that it's closer to that single server
also is very low."
맥락:
샘플링 특성:
"right now this, the metrics will always report the
last 30 seconds only, so it's going to be a tougher
metric to choose the system"
참고 데이터:
기술 설명:
"400 text servers tend to be either access denied or
object not found to be some other reason, like services
and available [or] Max tower Out something else"
중요성:
"error actually is internal errors. It's something
happening between nodes that one didn't want to take
a request for some reason"
모니터링 필요성:
기술적 배경:
"현재 메모리 사용량이 낮은 이유는 네트워크 대역폭 대비
부하가 높지 않기 때문이며, 접속 수가 급증할 경우 연결 당
최대 64MB까지 버퍼를 사용하면서 메모리 수치가 상승할 수 있음"
| 항목 | 헬스체크 | 트러블슈팅 (MC Browse) |
|---|---|---|
| 목적 | 정기 점검 | 실시간 문제 해결 |
| 빈도 | 월 1회, 분기 1회 | 필요 시 즉시 |
| 시점 | 예방적 | 사후 대응적 |
| 범위 | 전체 시스템 | 특정 이슈 집중 |
| 상세도 | 일반적 | 매우 상세 |
| 도구 | 진단 스크립트 | MC Browse, MC Admin Profile |
MC Browse 시나리오 가이드 작성
LDAP 테스트 스크립트 제공
THP 진단 메커니즘 확인
Node 38 Go Routine 이슈
LDAP 동기화 진단
THP 설정 이슈
진단 도구 패키지 제공
모니터링 대시보드 템플릿
정기 헬스체크 체계 수립
MC Browse 활용 체계
모니터링 체계 고도화
문제 대응 프로세스 정립
| 메트릭 | 경고(Warning) | 심각(Critical) |
|---|---|---|
| Latency | > 50ms | > 100ms |
| 500 에러율 | > 0.1% | > 1% |
| Go Routine 증가 | > 200% | > 500% |
| 메모리 사용률 | > 70% | > 85% |
| CPU 사용률 | > 70% | > 85% |
문서 버전: 1.0
최종 수정일: 2026년 2월 6일
다음 리뷰: 2026년 2월 13일 (주간 진행상황 회의)
# Go Routine 프로파일
mc admin profile [alias] --type goroutine
# CPU 프로파일
mc admin profile [alias] --type cpu
# 메모리 프로파일
mc admin profile [alias] --type memory
# 전체 프로파일
mc admin profile [alias]
# MC Browse 실행
mc browse [alias]
# 특정 버킷 확인
mc browse [alias]/[bucket-name]
# 진단 스크립트 실행 (벤더 제공 예정)
./health-check.sh
# 결과 확인
cat health-check-results.txt