feat(sources)!: split source contracts into PollSource/StreamSource and add mode-aware source config

- Introduce explicit source interfaces: sources.PollSource and sources.StreamSource, with shared sources.Input (Name() only).
- Remove mandatory Kind() from the base source contract to support sources that emit multiple kinds.
- Add config.SourceMode (poll, stream, or omitted/auto) and SourceConfig.Kinds (plural expected kinds), while keeping legacy SourceConfig.Kind for compatibility.
- Enforce mode semantics in config validation (poll requires every, stream forbids every) and detect mode/driver mismatches in sources.Registry.
- Update docs and tests for the new source model and config behavior.
This commit is contained in:
2026-03-15 19:19:19 -05:00
parent fafba0f01b
commit 6c5f95ad26
12 changed files with 591 additions and 542 deletions

283
README.md
View File

@@ -1,255 +1,114 @@
# feedkit
**feedkit** provides the domain-agnostic core plumbing for *feed-processing daemons*.
`feedkit` provides domain-agnostic plumbing for feed-processing daemons.
A feed daemon is a long-running process that:
- polls one or more upstream providers (HTTP APIs, RSS feeds, etc.)
- normalizes upstream data into a consistent internal representation
- applies lightweight policy (dedupe, rate-limit, filtering)
- emits events to one or more sinks (stdout, files, databases, brokers)
feedkit is designed to be reused by many concrete daemons (e.g. `weatherfeeder`,
`newsfeeder`, `rssfeeder`) without embedding *any* domain-specific logic.
---
A daemon built on feedkit typically:
- ingests upstream input (polling APIs or consuming streams)
- emits domain-agnostic `event.Event` values
- applies optional processing (normalization, dedupe, policy)
- routes events to sinks (stdout, NATS, files, databases, etc.)
## Philosophy
feedkit is **not a framework**.
It does **not**:
- define domain schemas
- enforce allowed event kinds
- hide control flow behind inversion-of-control magic
- own your application lifecycle
Instead, it provides **small, composable primitives** that concrete daemons wire
together explicitly. The goal is clarity, predictability, and long-term
maintainability.
---
feedkit is not a framework. It provides small composable packages and leaves
lifecycle, domain schemas, and domain-specific validation in your daemon.
## Conceptual pipeline
Collect Normalize → Filter / Policy Route → Persist / Emit
Collect -> Normalize (optional) -> Policy -> Route -> Emit
In feedkit terms:
| Stage | Package(s) |
|---|---|
| Collect | `sources`, `scheduler` |
| Normalize | `normalize` (optional in `pipeline`) |
| Policy | `pipeline` |
| Route | `dispatch` |
| Emit | `sinks` |
| Configure | `config` |
| Stage | Package(s) |
|------------|--------------------------------------|
| Collect | `sources`, `scheduler` |
| Normalize | *(today: domain code; planned: pipeline processor)* |
| Policy | `pipeline` |
| Route | `dispatch` |
| Emit | `sinks` |
| Configure | `config` |
## Core packages
---
### `config`
## Public API overview
Loads YAML config with strict decoding and domain-agnostic validation.
### `config` — Configuration loading & validation
**Status:** 🟢 Stable
`SourceConfig` supports both source modes:
- `mode: poll` requires `every`
- `mode: stream` forbids `every`
- omitted `mode` means auto (inferred from the registered driver type)
- Loads YAML configuration
- Strict decoding (`KnownFields(true)`)
- Domain-agnostic validation only (shape, required fields, references)
- Flexible `Params map[string]any` with typed helpers
It also supports optional expected source kinds:
- `kinds: ["observation", "alert"]` (preferred)
- `kind: "observation"` (legacy fallback)
Key types:
- `config.Config`
- `config.SourceConfig`
- `config.SinkConfig`
- `config.Load(path)`
### `event`
---
Defines the domain-agnostic event envelope (`event.Event`) used across the system.
### `event` — Domain-agnostic event envelope
**Status:** 🟢 Stable
### `sources`
Defines the canonical event structure that moves through feedkit.
Includes:
- Stable ID
- Kind (stringly-typed, domain-defined)
- Source name
- Timestamps (`EmittedAt`, optional `EffectiveAt`)
- Optional `Schema` for payload versioning
- Opaque `Payload`
Key types:
- `event.Event`
- `event.Kind`
- `event.ParseKind`
- `event.Event.Validate`
feedkit infrastructure never inspects `Payload`.
---
### `sources` — Polling abstraction
**Status:** 🟢 Stable (interface); 🔵 evolving patterns
Defines the contract implemented by domain-specific polling jobs.
Defines source interfaces and driver registry:
```go
type Source interface {
type Input interface {
Name() string
Kind() event.Kind
}
type PollSource interface {
Input
Poll(ctx context.Context) ([]event.Event, error)
}
```
Includes a registry (sources.Registry) so daemons can register drivers
(e.g. openweather_observation, rss_feed) without switch statements.
Note: Today, most sources both fetch and normalize. A dedicated
normalization hook is planned (see below).
### `scheduler` — Time-based polling
**Status:** 🟢 Stable
Runs one goroutine per source on a configured interval with jitter.
Features:
- Per-source interval
- Deterministic jitter (avoids thundering herd)
- Immediate poll at startup
- Context-aware shutdown
Key types:
- `scheduler.Scheduler`
- `scheduler.Job`
### `pipeline` — Event processing chain
**Status:** 🟡 Partial (API stable, processors evolving)
Allows events to be transformed, dropped, or rejected between collection
and dispatch.
```go
type Processor interface {
Process(ctx context.Context, in event.Event) (*event.Event, error)
type StreamSource interface {
Input
Run(ctx context.Context, out chan<- event.Event) error
}
```
Current state:
- `pipeline.Pipeline` is fully implemented
Notes:
- a poll can emit `0..N` events
- stream sources emit events continuously
- a single source may emit multiple event kinds
- driver implementations live in downstream daemons and are registered via `sources.Registry`
Placeholder files exist for:
- `dedupe` (planned)
- `ratelimit` (planned)
### `scheduler`
This is the intended home for:
- normalization
- deduplication
- rate limiting
- lightweight policy enforcement
Runs one goroutine per source job:
- poll sources: cadence driven (`every` + jitter)
- stream sources: continuous run loop
### `dispatch` — Routing & fan-out
**Status:** 🟢 Stable
### `pipeline`
Routes events to sinks based on kind and isolates slow sinks.
Optional processing chain between collection and dispatch.
Processors can transform, drop, or reject events.
Features:
- Compiled routing rules
- Per-sink buffered queues
- Bounded enqueue timeouts
- Per-consume timeouts
- Sink panic isolation
- Context-aware shutdown
### `normalize`
Key types:
- `dispatch.Dispatcher`
- `dispatch.Route`
- `dispatch.Fanout`
Optional normalization package (already implemented). Typical use: sources emit raw
payload events, then normalize to canonical schemas in a pipeline stage.
### `sinks` — Output adapters
***Status:*** 🟡 Mixed
### `dispatch`
Defines where events go after processing.
Compiles routes and fans out events to sinks with per-sink queue/worker isolation.
```go
type Sink interface {
Name() string
Consume(ctx context.Context, e event.Event) error
}
```
### `sinks`
Registry-based construction allows daemons to opt into any sink drivers.
Defines sink interface and sink registry. Built-ins include `stdout` and `nats`, with
additional sink implementations at varying maturity.
Sink Status
stdout 🟢 Implemented
nats 🟢 Implemented
file 🔴 Stub
postgres 🔴 Stub
## Typical wiring
All sinks are required to respect context cancellation.
### Normalization (planned)
**Status:** 🔵 Planned (API design in progress)
Currently, most domain implementations normalize upstream data inside
`sources.Source.Poll`, which leads to:
- very large source files
- mixed responsibilities (HTTP + mapping)
- duplicated helper code
The intended evolution is:
- Sources emit raw events (e.g. `json.RawMessage`)
- A dedicated normalization processor runs in the pipeline
- Normalizers are selected by `Event.Schema`, `Kind`, or `Source`
This keeps:
- `feedkit` domain-agnostic
- `sources` small and focused
- normalization logic centralized and testable
### Runner helper (planned)
**Status:** 🔵 Planned (optional convenience)
Most daemons wire together the same steps:
- load config
- build sources
- build sinks
- compile routes
- start scheduler
- start dispatcher
A small, opt-in `Runner` helper may be added to reduce boilerplate while keeping the system explicit and debuggable.
This is not intended to become a framework.
## Stability summary
Area Status
Event model 🟢 Stable
Config API 🟢 Stable
Scheduler 🟢 Stable
Dispatcher 🟢 Stable
Source interface 🟢 Stable
Pipeline core 🟡 Partial
Normalization 🔵 Planned
Dedupe/Ratelimit 🔵 Planned
Non-stdout sinks 🔴 Stub
Legend:
🟢 Stable — API considered solid
🟡 Partial — usable, but incomplete
🔵 Planned — design direction agreed, not yet implemented
🔴 Stub — placeholder only
1. Load config.
2. Register/build sources from `cfg.Sources`.
3. Register/build sinks from `cfg.Sinks`.
4. Compile routes.
5. Start scheduler (`sources -> bus`).
6. Start dispatcher (`bus -> pipeline -> sinks`).
## Non-goals
`feedkit` intentionally does not:
feedkit intentionally does not:
- define domain payload schemas
- enforce domain-specific validation
- manage persistence semantics beyond sink adapters
- own observability, metrics, or tracing (left to daemons)
Those concerns belong in concrete implementations.
## See also
- NAMING.md — repository and daemon naming conventions
- event/doc.go — detailed event semantics
- **Concrete example:** weatherfeeder (reference implementation)
---
- enforce domain-specific event kinds
- own application lifecycle
- prescribe observability stack choices