Backend

The seelf backend is written in the Golang language for its simplicity and small footprint.

Architecture

Packages overview

• cmd/: contains application commands such as the serve one
• internal/: contains internal package representing the core features of this application organized by bounded contexts and app, domain and infra folders (see The Domain)
• pkg/: contains reusable stuff not tied to seelf which can be reused if needed

The Domain

I'm a big fan of Domain Driven Design and as such, this codebase is heavily influenced by that with minor tweaks to make it more Go friendly.

The internal/ follows a classic DDD structure with:

• app: commands and queries to orchestrate the domain logic
• domain: core stuff, entities and values objects, as pure as possible to be easily testable
• infra: implementation of domain specific interfaces for the current context

In Go, it's common to see entities as structs with every field exposed. In this project, I have decided to try something else to prevent unwanted mutations from happening and making things more explicit.

Domain entities encapsulates a lot of rules and should always be in a valid state. Entities only expose needed fields and methods to mutate them and keep their invariants true. When they mutate, they raise events (stored directly inside them) representing what have been changed on an entity. Events enable the system to react to entities mutation (See Persistence).

Value objects regroup multiple properties that operate together into one struct acting as a whole and keeping their own invariants true.

The rule of thumbs in this project regarding struct creation is to always pass by a constructor function if any. I could have enforced the valid creation by hiding struct behind an interface but that's a lot of additional complexity.

Immutability and pointer vs value receivers

Entities are mutable and always use pointer receivers. That's because they may hold a lot of data and that's the most reasonable way to think about them.

In DDD, value objects should be immutable. But in Go, everything is pass by value by default hence a copy is done. This is why value objects in seelf are mutable (use pointer receiver for mutation methods). Since an entity always exposes its value objects as values, no harm could be done.

To make calls chainable, getters on value objects use a value receiver instead (mixing pointer and value receivers against Go recommendations... but I favor usability here).

Persistence

Only entities could be persisted. Since they raise events representing what have changed after a mutation, those events make their way to entity stores (defined as interfaces in domain packages and implemented in infra packages) where they are translated to SQL queries.

This make it easy to execute surgical SQL updates as needed.

Every entities should be read from the persistent store as a whole (= it should be populated with all their fields set). In this project, every entity expose a method which takes a storage.Scanner and returns an entity of the given type. This method, since it needs access to unexposed fields, is defined next to the public constructor of an entity in the domain sub-package.

Some value objects implements the Scanner, Valuer, Marshaler and Unmarshaler interfaces when they must be persisted in a single column. I may eventually found another cleaner way to do this but this is sufficient for now.

Some types are represented as discriminated union to express dynamic types. For example, SourceData (archive, git or raw file) could be anything supported by a Source and should be persisted and rehydrated as such. To enable those kind of use cases, every supported discriminated union type expose a storage.DiscriminatedMapper on the specific needed type and each types supported should register on it by defining a function to call to rehydrate this type specifically from a raw string payload and a discriminator value.

Retrieving related data is easy thanks to something inspired by graphql dataloaders. When querying the database, you can provide an optional array of Dataloader[T] which will execute additional requests based on key extracted from the parent result set. This approach enables efficient querying of the database by avoiding N+1 queries.

Commands and Queries

The domain is never accessed directly by client applications (here, a REST API). That's why there's app packages in every domain represented by an internal/ sub-package.

app packages expose commands and queries which can be processed by a bus.Dispatcher everywhere. Handlers are registered at the application startup by each infra/mod.go Setup function. Commands and queries means different things:

• command/: mutate the system based on given inputs, only take and returns primitive types. Commands translate application usecases to domain actions.
• query/: read stuff from the database. Types manipulated by queries are defined in the package itself since the representation may differ from the mutating stuff. Here, every field is public since it will be serialized afterwards and represents only UI needs. Query handlers are for the most part directly implemented by infra/sqlite packages in each domain.

Optionality

To make things more explicit, optional values are not represented using a pointer but a specific monad.Maybe[T] type instead. This type implements some common interfaces such as Scanner, Valuer, Marshaler and Unmarshaler to enable persistence and JSON serialization.

Useful commands

These commands must be executed from the root folder.

• make serve-back: Serve the web server
• make build: Build the seelf executable
• make ts: Print the current timestamp (unix only), needed when writing migrations
• make outdated: Print package versions and the latest one for updating
• make test: Run all test suites (front & back), if you only wish to launch the backend ones, just run go test ./... --cover instead