Skip to content

Implementation References and Freshness

Review date: 14 November 2026 · Fundamentals are durable; product capabilities and limits are not.

This guide teaches mechanisms rather than a list of product slogans. Selected implementation claims were checked against the primary sources below during this refresh. This is not a claim that every provider limit or release was exhaustively audited. Check the exact engine version, service edition, region and request settings before using a capability in production.

Important Updates to Older Mental Models

Older shortcut More accurate current model Primary reference
SQL scales vertically; NoSQL scales horizontally Query language, data model and deployment architecture are separate choices Spanner reads, writes and splits
A composite B-tree cannot help without its first column Leading equalities are usually best, but PostgreSQL 18 can choose skip scans PostgreSQL multicolumn indexes
BigQuery has no indexes It supports specialized search and vector indexes; scan pruning is still central Search indexes, vector indexes
All Cassandra secondary indexes have the same limitations Evaluate SAI separately from legacy 2i, with measured selectivity and fan-out Cassandra SAI
DynamoDB global tables are always eventually consistent MREC and MRSC are distinct modes with different constraints DynamoDB read consistency
A quorum formula proves linearizability Intersection is one ingredient; versioning, concurrent writes and protocol rules matter Cassandra guarantees
MongoDB chunks have a universal 64 MB size Current documentation describes a configurable 128 MB default range size MongoDB data partitioning
Spanner only offers one isolation level Serializable is the default; repeatable read is also documented Spanner transaction behavior

Read the Contract at the Correct Scope

Before copying a feature comparison, specify:

  1. Operation: one item, batch, transaction, scan or change stream?
  2. Scope: one partition, one region or multiple regions?
  3. Settings: isolation, read concern, write concern, timestamp bound or consistency mode?
  4. Failure model: process crash, disk loss, minority partition or regional outage?
  5. Evidence: documented guarantee, benchmark observation or inferred implementation?

A historical paper is valuable for the algorithm it describes, but does not prove a managed service still uses that exact storage layout. Avoid unsupported claims about proprietary B-tree/LSM hybrids, Bloom filters or internal quorum counts.

Foundational Reading

Diagrams and Learning Questions

The site's new SVG illustrations are exported from adjacent editable .drawio files. They use explicit labels and high-contrast fills; color is supplementary, not the only source of meaning. The SVGs render without a diagram service at page-view time. Existing Mermaid diagrams remain for lightweight sequences and comparisons.

Questions throughout the chapters use the MkDocs Quiz plugin with shuffled choices, explicit submission and explanatory feedback. They are original learning and interview-style exercises, not claimed to be recent real-company interview questions. Use them to explain assumptions, trace a failure and justify a trade-off. Future question-specific updates can link back to these foundations without turning the guide into memorized answers.

Dependency Policy

The requirements pin the tested direct dependencies, including the two configured MkDocs plugins. The update selects stable non-yanked releases at least seven days old where available; it deliberately does not jump to a just-published version solely to say "latest". Direct pins are not a full transitive lock file. See the repository's verification instructions for the tested install/build workflow.