[Four Speeds of Memory: HOOX Multi-Tier Storage]
D1, KV, R2, Durable Objects, Vectorize—assign each job by consistency, latency, and cost.
Sources: docs/devops/architecture/storage.mdx, DESIGN.md data models, paper data-layer sections, README primitives.
Thesis
HOOX does not pretend one database can be a mutex, a config plane, a ledger, and a PDF archive. It assigns each job to the Cloudflare® primitive whose consistency, latency, and cost match the job. That assignment is the data architecture.
1. Tier map
| Primitive | Consistency | Feel | Ideal use |
|---|---|---|---|
| Durable Objects | Strong (single-threaded) | <2 / <5 ms class | Idempotency locks, WS session managers |
| Workers KV | Eventual | Sub-ms reads; slow global write | Kill switch, queue mode, routing, rate windows, watermarks |
| D1 (SQLite) | Relational edge SQL | ~5–12 ms class | Trades, positions, balances, system_logs |
| R2 | Strong per object | Tens of ms | Verbose exchange JSON, PDFs, heavy logs |
| Vectorize | ANN index | Query-bound | RAG over trade history |
| Analytics Engine | Append time-series | Non-blocking write | Latency & API metrics |
Adjacent: Queues, Browser Rendering, Secrets (never KV/D1 for API keys).
2. D1 through one door
All business SQL → d1-worker. Centralization enables:
- Query guards (dangerous statements, injection shapes, table allowlists)
- Schema evolution in one place
- Binding auth in front of the ledger
Core tables: trade_signals, trades, positions, balances, system_logs—with trace/request ids for reconstruction.
3. KV as control plane
CONFIG_KV is where operators and agents meet: kill switch, queue mode, routing overrides, rate windows, trailing watermarks. Reads are cheap; writes are eventually consistent. Policy belongs here; double-spend locks do not.
4. R2 as cold / bulky path
Exchange dumps and PDF reports would punish D1 quotas. R2 takes bulk; Telegram gets links; the ledger stays lean. Enterprise thickens this into jurisdiction-aware immutable audit (essay 13).
5. Analytics Engine as free black box
trackAnalytics after response via waitUntil. Metrics must not lengthen the 22 ms path. Dashboard SQL against hoox-analytics reconstructs latency without a third-party APM bill at retail scale.
6. Decision procedure
For every new field:
- Must concurrent requests serialize? → DO
- Runtime operator flag? → KV
- Money / relational agent state? → D1 via d1-worker
- Huge or archival? → R2
If the answer is “all of the above,” you are designing a workflow, not a column.
Previous: agent-worker · Next: Five-layer security