2nd/10_Wiki/Topics/Domain_Programming/AI_and_ML/카산드라(Cassandra).md

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

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
wiki-2026-0508-카산드라-cassandra	카산드라(Cassandra)	10_Wiki/Topics	verified	self	Apache Cassandra Cassandra C*	none	A	0.9	applied	database nosql distributed wide-column ap-system		2026-05-10	pending	language framework Java/CQL Apache Cassandra 5.0

카산드라 (Cassandra)

매 한 줄

"매 write-optimized · 매 masterless · 매 AP 의 매 wide-column store". 매 Apache Cassandra 5.0 (2026) 은 매 Dynamo-style replication + 매 BigTable-style data model 의 매 합 — 매 single-region 1M+ writes/sec 의 매 linear scale, 매 multi-DC active-active, 매 tunable consistency. 매 partition key 설계 가 매 흥망 — 매 잘못된 model 은 매 hotspot · 매 large partition 의 매 재앙.

매 핵심

매 architecture

• 매 Masterless: 매 모든 node 가 매 동등 — 매 single point of failure 부재.
• 매 Consistent hashing: 매 token ring + 매 vnode (default 16) — 매 even distribution.
• 매 Replication: 매 RF=3 의 매 typical, 매 NetworkTopologyStrategy 로 매 multi-DC.
• 매 Gossip: 매 peer-to-peer cluster state.
• 매 LSM tree storage: 매 memtable → SSTable, 매 compaction (STCS / LCS / TWCS).

매 consistency

• 매 Tunable: ANY, ONE, QUORUM, LOCAL_QUORUM, EACH_QUORUM, ALL.
• 매 Strong: R+W > N (e.g., RF=3, R=QUORUM, W=QUORUM).
• 매 Eventual: ONE/ANY — 매 fast 하지만 매 stale read 가능.
• 매 LWT (Paxos): 매 conditional write — 매 비싸지만 매 linearizable.

매 응용

1. 매 time-series (IoT, metrics, logs).
2. 매 messaging / 매 feed (Discord 의 매 trillion+ msgs).
3. 매 session / 매 cart store.
4. 매 GenAI 의 매 vector + Cassandra 5 의 매 SAI vector index.

💻 패턴

Pattern 1: 매 Schema Design (query-first)

-- 매 BAD: 매 hotspot — 매 single partition
CREATE TABLE messages (
  channel_id uuid PRIMARY KEY,
  msg_id timeuuid,
  body text
);

-- 매 GOOD: 매 bucketed time partition
CREATE TABLE messages (
  channel_id uuid,
  bucket text,        -- 매 'YYYY-MM-DD'
  msg_id timeuuid,
  body text,
  PRIMARY KEY ((channel_id, bucket), msg_id)
) WITH CLUSTERING ORDER BY (msg_id DESC);

Pattern 2: 매 Vector Search (Cassandra 5 SAI)

CREATE TABLE products (
  id uuid PRIMARY KEY,
  name text,
  embedding vector<float, 1536>
);

CREATE CUSTOM INDEX ON products(embedding)
  USING 'StorageAttachedIndex'
  WITH OPTIONS = { 'similarity_function' : 'cosine' };

-- 매 ANN search
SELECT id, name FROM products
ORDER BY embedding ANN OF [0.1, 0.2, ...]
LIMIT 10;

Pattern 3: 매 Driver Async (Java)

CqlSession session = CqlSession.builder().build();
PreparedStatement ps = session.prepare(
  "INSERT INTO messages (channel_id, bucket, msg_id, body) VALUES (?, ?, ?, ?)"
);

CompletionStage<AsyncResultSet> f = session.executeAsync(
  ps.bind(channelId, bucket, msgId, body)
    .setConsistencyLevel(ConsistencyLevel.LOCAL_QUORUM)
);

Pattern 4: 매 Compaction Strategy 선택

-- 매 time-series → TWCS
ALTER TABLE metrics WITH compaction = {
  'class': 'TimeWindowCompactionStrategy',
  'compaction_window_size': '1',
  'compaction_window_unit': 'DAYS'
};

-- 매 read-heavy → LCS
ALTER TABLE users WITH compaction = {
  'class': 'LeveledCompactionStrategy',
  'sstable_size_in_mb': '160'
};

-- 매 write-heavy general → STCS (default)

Pattern 5: 매 Multi-DC Replication

CREATE KEYSPACE app
  WITH replication = {
    'class': 'NetworkTopologyStrategy',
    'us-east': 3,
    'eu-west': 3,
    'ap-northeast': 2
  } AND durable_writes = true;

Pattern 6: 매 LWT (conditional)

-- 매 unique constraint
INSERT INTO users (email, id) VALUES ('a@b.com', uuid())
  IF NOT EXISTS;

-- 매 비쌈 — 매 4 round trip Paxos. 매 hot path 회피.

Pattern 7: 매 Anti-pattern 진단

-- 매 nodetool tablestats 로 매 large partition 확인
-- nodetool tablestats keyspace.table | grep "Compacted partition maximum"

-- 매 100MB+ partition = 매 redesign signal

매 결정 기준

상황	Approach
매 write 1M+/sec	Cassandra (자연 fit)
매 strong consistency 필수	LWT or 매 다른 DB (CockroachDB, Spanner)
매 ad-hoc query / JOIN	Postgres / Trino — 매 Cassandra 부적합
매 time-series	Cassandra + TWCS or ScyllaDB
매 vector + scale	Cassandra 5 SAI or Milvus/Qdrant
매 small data (< 1TB)	Postgres — 매 Cassandra overkill

기본값: 매 query-first schema, 매 LOCAL_QUORUM, 매 RF=3, 매 partition < 100MB.

🔗 Graph

• 부모: NoSQL
• Adjacent: CAP Theorem & PACELC

🤖 LLM 활용

언제: 매 large-scale write workload 의 매 design, 매 multi-DC active-active 요건, 매 time-series storage, 매 schema review. 언제 X: 매 transactional / OLTP / JOIN 매 heavy — 매 RDBMS 가 매 적합. 매 small data — 매 over-engineering.

❌ 안티패턴

• 매 Large partition (>100MB): 매 OOM, 매 compaction failure, 매 read latency 폭발.
• 매 Hotspot key: 매 single-channel 모든 msg → 매 partition 폭발.
• 매 ALLOW FILTERING: 매 full scan — 매 production X.
• 매 Secondary index 의 매 high cardinality: 매 매번 매 fanout — 매 SAI 사용.
• 매 LWT 의 매 hot path: 매 4× latency.
• 매 SQL mindset (JOIN, GROUP BY): 매 denormalize 의 매 의무.

🧪 검증 / 중복

• Verified (Apache Cassandra 5.0 docs, DataStax docs, Discord engineering blog 2026).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Cassandra 5.0 (SAI vector) full 정리