Rationale

Capstone Step 4: per-store CAP/PACELC + license rationale

exercises co-03, co-04, co-28

Every store this capstone uses is chosen because its own access pattern demands it -- not because "NoSQL" is a single interchangeable category. This page states, for each of the three stores kv.py, doc.py, and wide.py build on, the access pattern that justifies the choice, the CAP/PACELC position that follows from how the store is actually configured here, and the exact, checked license -- formalizing the same content Example 80 already drafted and verified in code.

Valkey (session/cache store, kv.py)

Access pattern: TTL-bound session tokens, single-key CRUD, disposable by design. kv.py's own create_session/get_session/touch_session/delete_session never read across sessions -- every operation targets exactly one session:<id> key, and a session that is never explicitly deleted still self-expires via its own TTL (co-20, co-24, co-21).

CAP/PACELC: AP/PA/EL. A single-node cache favors availability and low latency over strict consistency: if this session store were ever replicated, a stale read of "is this session still valid" for a few milliseconds is a far cheaper mistake than refusing a login because a replica is momentarily unreachable. There is no cross-key transaction here to protect -- each session round-trips independently, which is exactly the shape AP/PA/EL is built for (co-03, co-04).

License: BSD-3-Clause -- Valkey, the Linux Foundation fork of Redis (the actual store kv.py connects to in this capstone, per Example 25's own citation). A permissive, OSI-approved license with no source-availability or field-of-use restriction.

MongoDB (document store, doc.py)

Access pattern: two NAMED reads -- "fetch a product with its current stock level" and "fetch every product in a given category, cheapest first" -- both served by indexes chosen specifically to answer them (co-08, co-17). doc.py's seed_products creates exactly the sku index and the (category, price) compound index those two patterns need, nothing more.

CAP/PACELC: CP/PC/EC. doc.py's own update_stock writes a real inventory count -- a value a reader must never see go backward or read stale after a genuine sale, the kind of correctness a product catalog's stock level specifically needs. A majority write concern (the production-grade choice this pattern implies) favors consistency over availability under a partition: better to briefly refuse a write than let two replicas disagree about how many units are left (co-03, co-04).

License: SSPLv1 (Server Side Public License), NOT OSI-approved open source (per Example 26's own citation). Source-available, but SSPLv1's own terms extend copyleft obligations to any service offered "as a service" around the software -- a materially different commitment from a permissive or traditional copyleft license, and one that must be understood before adopting MongoDB behind a managed-service offering.

Cassandra (wide-column store, wide.py)

Access pattern: order history per customer, an append-heavy, partition-scoped feed. wide.py's own fetch_order_history reads exactly one partition (WHERE customer_id = ?), already clustering-ordered newest-first -- and wide.py's own two-customer test proves that partition isolation holds: cust-2's orders never appear when reading cust-1's history (co-22, co-25).

CAP/PACELC: AP/PA/EL at consistency level ONE, or CP/PC/EC at QUORUM -- tunable per query (co-07). wide.py's own single-node local cluster does not exercise this tuning directly (replication_factor 1 has no quorum to speak of), but the SAME partition-scoped read shape this capstone uses is exactly what a production Cassandra deployment would read at ONE for availability, or at QUORUM if a customer's own order history must never show a stale write immediately after checkout.

License: Apache License 2.0 -- Cassandra is an Apache Software Foundation top-level project (per Example 27's own citation). A permissive, OSI-approved license, matching the license tier wide.py's own driver (cassandra-driver, also Apache-2.0) carries.

Summary table

Store	Access pattern (justifies the choice)	CAP/PACELC	License
Valkey	TTL-bound session CRUD, single-key, disposable	AP/PA/EL	BSD-3-Clause
MongoDB	Two named, index-served reads over a product catalog	CP/PC/EC (majority write concern)	SSPLv1 (NOT OSI-approved)
Cassandra	Per-customer order history, partition-scoped, append-heavy	AP/PA/EL at ONE, CP/PC/EC at QUORUM	Apache License 2.0 (ASF top-level)

Every row states three things together, matching Example 80's own discipline: the access pattern that justifies the choice, the CAP/PACELC position, and the exact, checked license -- a store recommendation missing any one of the three is an incomplete rationale.

---

← Back to Overview