Files
Aether/crates/aether-data
fawney19 9fc9c334c9 Merge remote-tracking branch 'origin/pr/487'
# Conflicts:
#	crates/aether-data/src/lifecycle/bootstrap/postgres.rs
#	crates/aether-data/src/lifecycle/migrate/tests.rs
#	crates/aether-data/src/repository/oauth_providers/postgres.rs
#	crates/aether-data/src/repository/oauth_providers/sqlite.rs
#	frontend/src/views/admin/OAuthSettings.vue
2026-05-19 02:27:39 +08:00
..
2026-05-06 02:22:32 +08:00
2026-05-05 18:27:36 +08:00

aether-data

aether-data is the runtime data-access crate. It owns concrete SQL/database drivers, concrete repository implementations, migration/backfill/export workflows, and the composition layer that wires those pieces into the rest of the application.

It does not own the cross-crate DTO contracts. Shared repository records and errors that are consumed by scheduler, billing, admin, usage runtime, and video task crates live in ../aether-data-contracts.

Directory Map

Path Responsibility
src/database.rs Logical SQL driver selection and shared pool configuration.
src/config.rs Data-layer config for SQL drivers and repository wiring.
src/maintenance.rs Maintenance DTOs and aggregation summaries used by backend dispatch and runtime maintenance entrypoints.
src/driver/{postgres,mysql,sqlite} Low-level SQL driver primitives such as pools, transactions, and leases. These modules should not contain domain repository logic.
src/repository Domain repository traits/types re-exported from contracts plus concrete in-memory/Postgres/MySQL/SQLite implementations.
src/backend Composition root. Builds concrete driver backends and exposes app-facing read/write/worker/lock/lease handles.
src/backend/{maintenance,stats,wallet,system}.rs Backend-owned maintenance, aggregation, wallet ledger, and system config workflows that are not normal request-path repositories.
src/lifecycle/migrate.rs and src/lifecycle/migrate/* Runtime migration entry points and migration-specific tests/helpers.
src/lifecycle/backfill.rs Backfill entry points and backfill discovery.
src/lifecycle/export.rs Cross-database export/import workflows.
migrations/{postgres,mysql,sqlite} Executable sqlx migrations embedded at compile time.
schema Schema maintenance workspace for logical definitions, driver fragments, and generated output.
schema/logical Human-maintained logical table definitions used by aether-data-schema.
schema/drivers/{postgres,mysql,sqlite} Human-maintained driver fragments that compose back into executable SQL while generation is being promoted.
schema/bootstrap/postgres Human-maintained source fragments for the Postgres bootstrap snapshot. build.rs composes them into the runtime embedded artifact during crate builds.
schema/generated Machine-written SQL generated from logical schema for audit and drift detection only.
schema/overrides Rare driver-specific SQL escape hatch. Keep README-only until a real override is needed.
backfills/{postgres,mysql,sqlite} Executable backfill SQL grouped by driver.

Layering Rules

The crate is easiest to read as five layers:

  1. Contracts: DTOs, input structs, repository traits, and DataLayerError. Prefer aether-data-contracts for anything that another crate needs to compile against.
  2. Driver primitives: driver/postgres, driver/mysql, and driver/sqlite connect to infrastructure and expose pools/runners.
  3. Repository implementations: repository/<domain>/{sql,mysql,sqlite,memory} translate contract types to driver-specific SQL.
  4. Backend composition: backend chooses one SQL driver from config and wires repository implementations into app-facing handles. Backend-owned runtime maintenance workflows live in focused backend modules rather than in the driver pool files.
  5. Maintenance workflows: lifecycle/migrate, lifecycle/backfill, lifecycle/export, and schema manage database lifecycle outside normal request handling.

Do not add domain queries to low-level pool modules. Do not add driver selection logic inside individual repository implementations. Keep cross-crate contracts out of aether-data unless they are implementation-only.

Repository Layout

Most domain repositories use this shape:

src/repository/<domain>/
  mod.rs      # exports trait/type names and concrete implementations
  types.rs    # implementation-local DTOs when they are not already in contracts
  postgres.rs # Postgres implementation
  mysql.rs    # MySQL implementation
  sqlite.rs   # SQLite implementation
  memory.rs   # tests/dev in-memory implementation

Use explicit driver filenames for repository implementations. Do not introduce new generic sql.rs modules for driver-specific code.

SQL Driver Policy

The project supports three SQL drivers at the repository/backend boundary: Postgres, MySQL, and SQLite. That does not mean every raw SQL file is shared. The portable contract is the Rust shape and behavior; the physical SQL stays driver-specific where syntax, indexes, JSON support, timestamps, locking, or upsert semantics differ.

Use logical types in design docs and reviews:

Logical type Postgres MySQL SQLite
json json or jsonb json or text JSON text JSON
bool boolean boolean / tinyint(1) integer
time_unix bigint or legacy timestamp bigint integer
money_decimal numeric / legacy double double real

jsonb is acceptable only in Postgres SQL. MySQL and SQLite migrations must not contain jsonb; this is guarded by migration tests. Prefer serde_json::Value or typed Rust structs at the repository boundary so callers do not depend on the physical storage type.

Schema Maintenance

Executable migrations stay under migrations/{postgres,mysql,sqlite} because sqlx::migrate! embeds those paths and existing deployments record those file versions.

The large baseline SQL files are maintained through schema fragments. New table-structure work should start in schema/logical/*.toml and be generated into driver-specific SQL before it is composed into executable migrations:

bash crates/aether-data/schema/compose_schema.sh generate
bash crates/aether-data/schema/compose_schema.sh compose
bash crates/aether-data/schema/compose_schema.sh check

schema/generated/** is machine-written by aether-data-schema; it is checked in to make generator drift reviewable, not because runtime reads it. Edit schema/logical/*.toml instead. Use schema/overrides/** only as an exception bucket for driver-specific SQL that cannot live cleanly in logical schema or the normal driver fragments.

compose_schema.sh check also verifies that required baseline/portable table-creation SQL is represented in schema/logical. This is the guardrail that keeps table structure from drifting back into three manually maintained definitions.

For executable fragments that have not been promoted to generated output yet, edit fragments under schema/drivers/{postgres,mysql,sqlite} directly, run compose, then run check. Do not edit baseline executable SQL and fragments independently.

When adding a table:

  1. Add or update schema/logical/*.toml first for the table structure.
  2. Run schema/compose_schema.sh generate and inspect the generated driver SQL.
  3. Add or update the executable driver-specific migration/fragments only for deployment compatibility or generator gaps.
  4. Add or update repository contracts in aether-data-contracts if other crates need the new shape.
  5. Add driver repository implementations only for the drivers that are actually supported for that domain.
  6. Wire new repositories through src/backend/read.rs or src/backend/write.rs only after the implementation exists for each selected driver.
  7. Update docs/architecture/data-schema-inventory.md for new tables or logical type changes.

Known Cleanup Targets

These are intentionally staged to keep the multi-database refactor reviewable:

  1. Group repository domains once file-level names are stable. Likely groups: identity, auth config, provider catalog, runtime tasks, wallet/billing, usage, stats, and proxy nodes.
  2. Continue shrinking Postgres stats SQL modules where useful by moving shared SQL fragments and row-mapping helpers behind focused backend/stats/* modules.
  3. Consider a later crate split only after module boundaries are stable. The likely split is schema/migration tooling versus runtime repository backends, not an ORM rewrite.