Files
JiboExperiments/OpenJibo/docs/persistence-architecture.md

137 lines
3.6 KiB
Markdown
Raw Normal View History

# Persistence Architecture
## Goal
Keep OpenJibo's stateful behavior portable now and Azure-ready later.
The current in-memory stores are fine as the default implementation, but the app should depend on stable persistence contracts rather than directly on in-memory collections or file formats.
## Design Principles
- Application code talks to small, intent-specific interfaces.
- Persistence keys are always scoped by tenant and person where relevant.
- In-memory, local JSON, and hosted Azure stores are adapters, not behavior sources.
- Long-lived data should be versioned so we can add optimistic concurrency later.
- Ephemeral turn/session state should stay separate from durable user and device state.
## Current Seams
These are the contracts we should preserve:
- `IPersonalMemoryStore`
- personal facts: names, birthdays, preferences, affinities, important dates, household lists
- scope: account + loop + device + optional person
- `ICloudStateStore`
- account, robot, loops, people, sessions, updates, media, backups, holidays, keys
- scope: system-level state with loop/device/person records inside it
- `IJiboExperienceContentRepository`
- catalog/content layer only
## Recommended Storage Split
### 1. Identity and topology store
Responsible for:
- account profile
- robot/device registration
- loop membership
- person records
- greeting/proactive presence metadata when it becomes durable
This is the seam most likely to become Azure SQL or Cosmos later.
### 2. Personal memory store
Responsible for:
- names
- birthdays
- preferences
- affinities
- important dates
- household lists
This can remain in memory now and later move to a durable store keyed by account/loop/device/person.
### 3. Session and short-lived orchestration state
Responsible for:
- websocket/session tokens
- temporary skill state
- active report/list/greeting interaction state
This can stay in-process for now, but should be clearly separated from durable memory.
### 4. Media and backup store
Responsible for:
- uploaded media metadata
- backup manifests
- binary references
This is a good candidate for Azure Blob Storage plus a metadata table later.
## Record Shape Guidance
For durable records, prefer a small shared envelope:
- `AccountId`
- `LoopId`
- `DeviceId`
- `PersonId` when relevant
- `RecordType`
- `RecordKey`
- `Value`
- `CreatedUtc`
- `UpdatedUtc`
- `Revision` or `ETag`
That gives us:
- easy partitioning later
- clear tenant boundaries
- room for concurrency checks
- a path to Azure Table, Cosmos, or SQL without changing behavior code
## Adapter Plan
### Phase 1
- keep `InMemoryPersonalMemoryStore`
- keep `InMemoryCloudStateStore`
- make sure all callers use the interfaces only
- add tests against behavior, not implementation details
### Phase 2
- introduce durable adapters behind the same interfaces
- likely split:
- SQL or Cosmos for identity/topology
- Blob or table-backed store for media/backup metadata
- table/SQL-backed memory store for personal facts
### Phase 3
- add replication/sync primitives if we need multi-server state convergence
- prefer explicit change records or versioned snapshots over hidden shared state
## Non-Goals For Now
- no Azure SDK types in application logic
- no event-sourcing rewrite
- no giant generic repository
- no distributed transaction work before single-node semantics are stable
## Immediate Next Step
Before building durable adapters, tighten the store contracts around:
- tenant/person scoping
- record versioning
- explicit load/save operations for durable state
That lets us swap the backing store later without changing the personality, report, greeting, or list behaviors already built on top.