Command Center

Forge mode

할 일을 늘리는 페이지가 아니다. 네 개 문서를 하나의 실행 순서로 압축한다: campaign이 오늘 사고를 정하고, lab에서 직접 터뜨리고, move로 내 말 압축하고, sparring에서 영어로 맞는다.

Start sparring

Only rule

오늘은 한 stage만 한다. 한 lab만 터뜨린다. 한 move만 내 말로 채운다. 그리고 `/sparring`에서 5번 맞는다.

전체 목록은 지도일 뿐이다. 체크박스 전체를 끝내는 게임이 아니라, 사고 하나를 몸에 새기는 게임이다.

Campaign

오늘 어떤 사고를 겪을지 정한다. 순서를 고민하지 않는다.

docs/assets/CAMPAIGN.md

Lab

순진하게 만들고, 공격하고, 망가진 숫자/로그를 직접 본다.

docs/assets/EXPERIENCE_LABS.md

Move

겪은 사고를 System 15로, 알고리즘 문제는 Algorithm 10으로 따로 압축한다.

docs/assets/15_MOVES_MY_WORDS.md

Sparring

화면을 보면서 영어로 재설명하고 AI한테 공격받는다.

/sparring

Campaign spine

순서를 고민하지 말고 위에서 아래로. 각 stage는 사고 하나다.

11 stages

Stage 0 - point shop baseline

Trigger: users, points, products, orders 4 tables and small Java/Spring APIs.

Symptom: It works because nothing has attacked it yet.

Lab

Build the naive version by hand.

Move

수1 진실은 한 곳에 / 수2 묶는다

Sparring

Explain the source of truth and transaction boundary.

Metric

Basic happy-path latency + row counts.

Stage 1 - 1M users, slow reads

Trigger: Marketing succeeds; orders grow from small to millions.

Symptom: Order lookup becomes seconds, not milliseconds.

Lab

EXPLAIN ANALYZE, compound index, N+1, index write cost.

Move

수6 덜 한다 / 수9 미리 한다

Sparring

Explain index order and what p95/p99 prove.

Metric

p95/p99 query latency, rows scanned, buffers, insert cost.

Stage 2 - read spike, cache pain

Trigger: Sale page drives thousands of reads per second.

Symptom: DB CPU spikes; cache fixes reads but creates staleness and stampede.

Lab

Cache-aside, stale value, stampede, singleflight.

Move

수7 복사한다 / 수13 한도를 건다

Sparring

Explain cache freshness, TTL, and stampede control.

Metric

cache hit rate, DB QPS, p95, stale response count.

Stage 3 - flash sale concurrency

Trigger: 100 items, 5000 buyers at once.

Symptom: Negative stock, double point deduction, deadlocks.

Lab

Lost update, CAS, lock order, isolation tour.

Move

수2 묶는다 / 수3 멱등하게 반복한다

Sparring

Explain the invariant and how concurrent requests preserve it.

Metric

incorrect balance count, conflict rate, lock wait time.

Stage 4 - real money, payment webhooks

Trigger: External PG/webhook enters the system.

Symptom: Duplicate webhook double-charges or double-credits points.

Lab

Idempotency key, state machine, reconciliation.

Move

수3 멱등 / 수5 대조

Sparring

Explain retry safety and where the idempotency key lives.

Metric

duplicate attempts, dedup hits, reconciliation mismatch count.

Stage 5 - async split

Trigger: Order flow includes notification, points, stats, and takes seconds.

Symptom: DB commits but server dies before sending downstream work.

Lab

Half-state, outbox, Kafka duplicate/order/lag/poison/DLQ.

Move

수4 먼저 기록하고 나중 행동 / 수10 줄 세운다

Sparring

Explain outbox and why queue means latency and eventual consistency.

Metric

queue lag, DLQ count, consumer error rate, end-to-end delay.

Stage 6 - external dependency failure

Trigger: Shipping API hangs for 30 seconds.

Symptom: Threads pile up and unrelated endpoints slow down.

Lab

Timeout, circuit breaker, bulkhead, fallback.

Move

수13 한도 / 수14 보이게 실패

Sparring

Explain timeout budget, circuit breaker state, and fallback trade-off.

Metric

p99 latency, timeout count, open circuit count, thread pool usage.

Stage 7 - observability era

Trigger: Users say it is slow, but nobody knows where.

Symptom: No dashboard, no p95/p99, no trace, no root cause.

Lab

Prometheus/Grafana, memory leak, 3am diagnosis, postmortem.

Move

수14 보이게 실패 / 수5 대조

Sparring

Explain what you check first from metrics, logs, and traces.

Metric

RPS, error rate, p95/p99, heap, GC, pool saturation.

Stage 8 - search outgrows LIKE

Trigger: Product search is added on millions of rows.

Symptom: LIKE '%x%' dies; search index becomes a copy with lag.

Lab

LIKE vs Postgres FTS vs optional Elasticsearch.

Move

수7 복사한다 / 수9 미리 한다

Sparring

Explain when FTS is enough and when ES earns its cost.

Metric

search p95, index lag, reindex time, result quality checks.

Stage 9 - data explosion

Trigger: orders reaches 100M rows.

Symptom: No single index makes all queries cheap; cross-partition work hurts.

Lab

Partitioning, batch inserts, CQRS-lite.

Move

수8 쪼갠다 / 수7 복사한다

Sparring

Explain partition boundary and the cost of cross-boundary queries.

Metric

partition scan count, write throughput, aggregation latency.

Stage 10 - deployment fear

Trigger: The system is large enough that every change feels risky.

Symptom: A 200 response lies; old code may still be serving.

Lab

CI gate, Docker, rollback, running image digest verification.

Move

수14 보이게 실패 / 수5 대조

Sparring

Explain deploy verification and rollback with metrics.

Metric

running image digest, error rate, p99, health checks, rollback time.

Move check

Lab 끝나면 System 15로 압축한다. 알고리즘 드릴은 Algorithm 10으로 따로 압축한다. 못 채우면 아직 네 것이 아니다.

진실은 한 곳에

Same fact cannot be owned by two places independently.

Where is the source of truth?

What duplicate fact can drift?

What constraint is the last defense?

묶는다

Things that must be true together must commit or fail together.

What is the transaction boundary?

What happens if the process dies halfway?

What invariant must hold?

멱등하게 만들고 반복한다

Make retry safe before relying on retry.

What is the stable key?

What result is returned on replay?

What duplicate side effect is prevented?

먼저 기록하고 나중 행동한다

Durably record intent before doing side effects.

What record survives a crash?

Who picks it up later?

How do you avoid double-send?

주기적으로 대조한다

When the first four leak, reconciliation catches drift.

Which two sources should agree?

What mismatch types exist?

Which ones auto-repair?

측정해서 증명한다

A fix is not real until a metric proves the symptom moved.

Why p99, not average?

What alert would fire?

What number proves the fix?

Source files

World model

docs/assets/SYSTEM_DESIGN_ROADMAP.md

System 15 + Algorithm 10. This is the judging frame.

My words

docs/assets/15_MOVES_MY_WORDS.md

You fill the blanks. AI only cross-examines.

Experience labs

docs/assets/EXPERIENCE_LABS.md

Build naive, attack, witness failure, fix, attack again, log.

Campaign

docs/assets/CAMPAIGN.md

The point-shop story mode. It decides the order.