CAP is about behavior during a network partition: a replicated service cannot guarantee both linearizable responses and a successful response from every non-failing node. It does not mean a database permanently “chooses two of three.”
Name the client-visible guarantee
| Model | Client-visible promise |
|---|---|
| Linearizability | Each operation appears at one instant between call and return |
| Sequential consistency | One global order respects each client’s program order |
| Causal consistency | Causal order is preserved; concurrent writes may differ |
| Eventual consistency | Replicas converge after updates stop, without a freshness bound |
Consistency is not a slider labeled strong/weak. Each model permits specific observations.
Quorums expose a tunable boundary
With replication factor (N), write acknowledgements (W), and read responses (R), the familiar condition (R + W > N) creates overlap. Correctness still depends on version selection, failure assumptions, sloppy quorums, and repair; the inequality alone is not a proof.
Cassandra’s write path
Cassandra partitions keys around a ring and replicates them across failure domains. A replica appends to a commit log, updates a memtable, and later flushes a sorted SSTable. Bloom filters avoid searching SSTables that definitely do not contain a key; compaction merges files and discards obsolete versions.
Hinted handoff, read repair, and anti-entropy help replicas converge. Tunable consistency chooses how many replica responses a request waits for.
HBase chooses a different contract
HBase stores ordered rows in regions backed by HDFS. Region servers maintain a write-ahead log and MemStore, then flush HFiles. Range scans follow row-key order. Coordination and single-region ownership favor stronger consistency, while a partition may make an affected region unavailable.
In plain terms: Cassandra tends to keep more doors open during trouble and reconcile later; HBase is more willing to close a door than serve conflicting views.
Design from access and failure behavior
Choose partition keys, row keys, consistency level, and replication placement together. Ask what a client may read after its own write, what happens when a region or rack fails, how conflicts are represented, and how long convergence can take.
Review card
- CAP applies specifically when communication is partitioned.
- Consistency models are sets of allowed observations.
- Quorum overlap helps only with compatible version and repair rules.
- Cassandra uses append, memtable, SSTable, and background repair.
- HBase uses ordered regions and favors consistent access to an available owner.
- Storage layout and consistency contract must match the workload together.
CAP 描述 network partition 發生時的行為：replicated service 不能同時保證 linearizable response，以及每個 non-failing node 都成功回覆。它不是 database 永久「三選二」。
先說清楚 client 看得到什麼
| Model | Client-visible promise |
|---|---|
| Linearizability | Operation 像在 call／return 間某一瞬間發生 |
| Sequential consistency | 單一 global order 尊重各 client program order |
| Causal consistency | 保留 causal order；concurrent write 可不同 |
| Eventual consistency | Update 停止後最終 converge，沒有 freshness bound |
Consistency 不是 strong／weak slider；每個 model 允許的 observation 不同。
Quorum 顯示可調邊界
Replication factor (N)、write acknowledgements (W)、read responses (R) 下，(R + W > N) 會產生 overlap。Correctness 仍取決於 version selection、failure assumption、sloppy quorum 與 repair；不等式本身不是完整 proof。
Cassandra write path
Cassandra 將 key partition 到 ring，並跨 failure domain replication。Replica append commit log、update memtable，之後 flush sorted SSTable。Bloom filter 避免搜尋確定不含 key 的 SSTable；compaction merge file 並清 obsolete version。
Hinted handoff、read repair、anti-entropy 幫 replica converge；tunable consistency 決定 request 等多少 replica response。
HBase 選不同 contract
HBase 將 ordered row 放在 HDFS-backed region。Region server 維護 WAL／MemStore，再 flush HFile；range scan 使用 row-key order。Coordination 與 single-region ownership 偏向強 consistency，partition 時受影響 region 可能 unavailable。
白話來說： Cassandra 傾向故障時多開幾扇門、之後 reconcile；HBase 寧可關門，也不提供互相衝突的 view。
從 access 與 failure behavior 設計
Partition key、row key、consistency level、replica placement 要一起選。要問 client 在 own write 後能讀到什麼、region／rack failure 時怎麼辦、conflict 如何表示，以及 convergence 最久多久。
複習卡
- CAP 特別描述 communication partition。
- Consistency model 是允許 observation 的集合。
- Quorum overlap 仍需相容 version／repair rule。
- Cassandra 使用 append、memtable、SSTable 與 background repair。
- HBase 使用 ordered region，偏向只從 available owner 提供 consistent access。
- Storage layout 與 consistency contract 必須一起符合 workload。