Architecture

Package structure

Packages have strict rules of importing:

ahriman.application package must not be used outside of this package.
ahriman.core and ahriman.models packages don’t have any import restriction. Actually we would like to totally restrict importing of core package from models, but it is impossible at the moment.
ahriman.web package is allowed to be imported from ahriman.application (web handler only, only ahriman.web.web methods). It also must not be imported globally, only local import is allowed.

Full dependency diagram:

`ahriman.application` package

This package contains application (aka executable) related classes and everything for it. It also contains package called ahriman.application.handlers in which all available subcommands are described as separated classes derived from the base ahriman.application.handlers.Handler class.

ahriman.application.application.Application (god class) is used for any interaction from parsers with repository. It is divided into multiple traits by functions (package related and repository related) in the same package.

ahriman.application.application.workers package contains specific wrappers for local and remote build processes.

ahriman.application.ahriman contains only command line parses and executes specified Handler on success, ahriman.application.lock.Lock is additional class which provides file-based lock and also performs some common checks.

`ahriman.core` package

This package contains everything required for the most of application actions and it is separated into several packages:

ahriman.core.alpm package controls pacman related functions. It provides wrappers for pyalpm library and safe calls for repository tools (repo-add and repo-remove). Also this package contains ahriman.core.alpm.remote package which provides wrapper for remote sources (e.g. AUR RPC and official repositories RPC).
ahriman.core.auth package provides classes for authorization methods used by web mostly. Base class is ahriman.core.auth.Auth which must be instantiated by load method.
ahriman.core.build_tools is a package which provides wrapper for devtools commands.
ahriman.core.configuration contains extension for standard configparser library and some validation related classes.
ahriman.core.database is everything for database, including data and schema migrations.
ahriman.core.distributed package with triggers and helpers for distributed build system.
ahriman.core.formatters package provides Printer sub-classes for printing data (e.g. package properties) to stdout which are used by some handlers.
ahriman.core.gitremote is a package with remote PKGBUILD triggers. Should not be called directly.
ahriman.core.http package provides HTTP clients which can be used later by other classes.
ahriman.core.log is a log utils package. It includes logger loader class, custom HTTP based logger and some wrappers.
ahriman.core.report is a package with reporting triggers. Should not be called directly.
ahriman.core.repository contains several traits and base repository (ahriman.core.repository.Repository class) implementation.
ahriman.core.sign package provides sign feature (only gpg calls are available).
ahriman.core.status contains helpers and watcher class which are required for web application. Reporter must be initialized by using ahriman.core.status.client.Client.load method.
ahriman.core.support provides plugins for support packages (mirrorlist and keyring) generation.
ahriman.core.triggers package contains base trigger classes. Classes from this package must be imported in order to implement user extensions. In fact, ahriman.core.report, ahriman.core.upload and other built-in triggers use this package.
ahriman.core.upload package provides sync feature, should not be called directly.

This package also provides some generic functions and classes which may be used by other packages:

ahriman.core.exceptions provides custom exceptions.
ahriman.core.spawn.Spawn is a tool which can spawn another ahriman process. This feature is used by web application.
ahriman.core.tree is a dependency tree implementation.

`ahriman.models` package

It provides models for any other part of application. Unlike ahriman.core package classes from here provide only conversion methods (e.g. create class from another or convert to). It is mostly presented by case classes and enumerations.

`ahriman.web` package

Web application. It is important that this package is isolated from any other to allow it to be optional feature (i.e. dependencies which are required by the package are optional).

ahriman.web.middlewares provides middlewares for request handlers.
ahriman.web.schemas provides schemas (actually copy paste from dataclasses) used by swagger documentation.
ahriman.web.views contains web views derived from aiohttp view class.
ahriman.web.apispec provides generators for swagger documentation.
ahriman.web.cors contains helpers for cross origin resource sharing middlewares.
ahriman.web.routes creates routes for web application.
ahriman.web.web provides main web application functions (e.g. start, initialization).

Application run

Parse command line arguments, find subcommand and related handler which is set by the parser.
Call Handler.execute method.
Define list of architectures to run. In case if there is more than one architecture specified run several subprocesses or continue in current process otherwise. Class attribute ALLOW_MULTI_ARCHITECTURE_RUN controls whether the application can be run in multiple processes or not - this feature is required for some handlers (e.g. Web, which should be able to spawn child process in daemon mode; it is impossible to do from daemonic processes).
In each child process call lock functions.
After success checks pass control to Handler.run method defined by specific handler class.
Return result (success or failure) of each subprocess and exit from application.
Some handlers may override their status and throw ExitCode exception. This exception is just silently suppressed and changes application exit code to 1.

In the most cases handlers spawn god class ahriman.application.application.Application class and call required methods.

The application is designed to run from systemd services and provides parametrized by repository identifier timer and service file for that.

Subcommand design

All subcommands are divided into several groups depending on the role they are doing:

aur (aur-search) group is for AUR operations.
help (e.g. help) are system commands.
package subcommands (e.g. package-add) allow to perform single package actions.
patch subcommands (e.g. patch-list) are the special case of package subcommands introduced in order to control patches for packages.
repo subcommands (e.g. repo-check) usually perform actions on whole repository.
service subcommands (e.g. service-setup) perform actions which are related to whole service managing: create repository, show configuration.
user subcommands (user-add) are intended for user management.
web subcommands are related to web service management.

For historical reasons and in order to keep backward compatibility some subcommands have aliases to their shorter forms or even other groups, but the application doesn’t guarantee that they will remain unchanged.

Filesystem tree

The application supports two types of trees, one is for the legacy configuration (when there were no explicit repository name configuration available) and another one is the new-style tree. This document describes only new-style tree in order to avoid deprecated structures.

Having default root as /var/lib/ahriman (differs from container though), the directory structure is the following:

/var/lib/ahriman/
├── ahriman.db
├── cache
├── chroot
│   └── aur-clone
├── packages
│   └── aur-clone
│       └── x86_64
├── pacman
│   └── aur-clone
│       └── x86_64
│           ├── local
│           │   └── ALPM_DB_VERSION
│           └── sync
│               ├── core.db
│               ├── extra.db
│               └── multilib.db
│
└── repository
    └── aur-clone
        └── x86_64
            ├── aur-clone.db -> aur-clone.db.tar.gz
            ├── aur-clone.db.tar.gz
            ├── aur-clone.files -> aur-clone.files.tar.gz
            └── aur-clone.files.tar.gz

There are multiple subdirectories, some of them are commons for any repository, but some of them are not.

cache is a directory with locally stored PKGBUILD’s and VCS packages. It is common for all repositories and architectures.
chroot/{repository} is a chroot directory for devtools. It is specific for each repository, but shared for different architectures inside (the devtools handles architectures automatically).
packages/{repository}/{architecture} is a directory with prebuilt packages. When a package is built, first it will be uploaded to this directory and later will be handled by update process. It is architecture and repository specific.
pacman/{repository}/{architecture} is the repository and architecture specific caches for pacman’s databases.
repository/{repository}/{architecture} is a repository packages directory.

Normally you should avoid direct interaction with the application tree. For tree migration process refer to the migration notes.

Database

The service uses SQLite database in order to store some internal info.

Database instance

All methods related to the specific part of database (basically operations per table) are split into different traits located inside ahriman.core.database.operations package. The base trait ahriman.core.database.operations.Operations also provides generic methods for database access (e.g. row converters and transactional support).

The ahriman.core.database.SQLite class itself derives from all of these traits and implements methods for initialization, including migrations.

Schema and data migrations

The schema migrations are applied according to current pragma user_info values, located at ahriman.core.database.migrations package and named as m000_migration_name.py (the preceding m is required in order to import migration content for tests). Additional class ahriman.core.database.migrations.Migrations reads all migrations automatically and applies them in alphabetical order.

These migrations can also contain data migrations. Though the recommended way is to migrate data directly from SQL queries, sometimes it is required to have external data (like packages list) in order to set correct data. To do so, special method migrate_data is used.

Type conversions

By default, it parses rows into python dictionary. In addition, the following pseudo-types are supported:

dict[str, Any], list[Any] - for storing JSON data structures in database (technically there is no restriction on types for dictionary keys and values, but it is recommended to use only string keys). The type is stored as json data type and json.loads and json.dumps methods are used in order to read and write from/to database respectively.

Basic flows

By default package build operations are performed with PACKAGER which is specified in makepkg.conf, however, it is possible to override this variable from command line; in this case service performs lookup in the following way:

If packager is not set, it reads environment variables (e.g. SUDO_USER and USER), otherwise it uses value from command line.
It checks users for the specified username and tries to extract packager variable from it.
If packager value has been found, it will be passed as PACKAGER system variable (additional sudo configuration might be required).

Add new packages or rebuild existing

Idea is to add package to a build queue from which it will be handled automatically during the next update run. Different variants are supported:

If supplied argument is file, then application moves the file to the directory with built packages. Same rule applies for directory, but in this case it copies every package-like file from the specified directory.
If supplied argument is directory and there is PKGBUILD file there, it will be treated as local package. In this case it will queue this package to build and copy source files (PKGBUILD and .SRCINFO) to caches.
If supplied argument is not file then application tries to lookup for the specified name in AUR and clones it into the directory with manual updates. This scenario can also handle package dependencies which are missing in repositories.

This logic can be overwritten by specifying the source parameter, which is partially useful if you would like to add package from AUR, but there is local directory cloned from AUR. Also official repositories calls are hidden behind explicit source definition.

Rebuild packages

Same as add function for every package in repository. Optional filters by reverse dependency or build status can be supplied.

Remove packages

This flow removes package from filesystem, updates repository database and also runs synchronization and reporting methods.

Update packages

This feature is divided into to the following stages: check AUR for updates and run rebuild for required packages. Whereas check does not do anything except for check itself, update flow is the following:

Process every built package first. Those packages are usually added manually.
Run sync and report methods.
Generate dependency tree for packages to be built.
For each level of tree it does:
1. Download package data from AUR.
2. Bump pkgrel if there is duplicate version in the local repository (see explanation below).
3. Build every package in clean chroot.
4. Sign packages if required.
5. Add packages to database and sign database if required.
6. Process triggers.

After any step any package data is being removed.

In case if there are configured workers, the build process itself will be delegated to the remote instances. Packages will be partitioned to the chunks according to the amount of configured workers.

Distributed builds

This feature consists of two parts:

Upload built packages to the node.
Delegate packages building to separated nodes.

The upload process is performed via special API endpoint, which is disabled by default, and is performed in several steps:

Upload package to temporary file.
Copy content from temporary file to the built package directory with dot (.) prefix.
Rename copied file, removing preceding dot.

After success upload, the update process must be called as usual in order to copy built packages to the main repository tree.

On the other side, the delegation uses upload feature, but in addition it also calls external services in order to trigger build process. The packages are separated to chunks based on the amount of the configured workers and their dependencies.

pkgrel bump rules

The application is able to automatically bump package release (pkgrel) during build process if there is duplicate version in repository. The version will be incremented as following:

Get version of the remote package.
Get version of the local package if available.
If local version is not set, proceed with remote one.
If local version is set and epoch or package version (pkgver) are different, proceed with remote version.
If local version is set and remote version is newer than local one, proceed with remote.
Extract pkgrel value.
If it has major.minor notation (e.g. 1.1), then increment last part by 1, e.g. 1.1 -> 1.2, 1.0.1 -> 1.0.2.
If pkgrel is a number (e.g. 1), then append 1 to the end of the string, e.g. 1 -> 1.1.

Core functions reference

Configuration

ahriman.core.configuration.Configuration class provides some additional methods (e.g. getpath and getlist) and also combines multiple files into single configuration dictionary using repository identifier overrides. It is the recommended way to deal with settings.

Enumerations

All enumerations are derived from enum.StrEnum. Integer enumerations in general are not allowed, because most of operations require conversions from string variable. Derivation from string based enumeration is required to make json conversions implicitly (e.g. during calling json.dumps methods).

In addition, some enumerations provide from_option class methods in order to allow some flexibility while reading configuration options.

Utils

For every external command run (which is actually not recommended if possible) custom wrapper for subprocess is used. Additional functions ahriman.core.auth.helpers provide safe calls for aiohttp_security methods and are required to make this dependency optional.

Context variables

Package provides implicit global variables which can be accessed from ahriman.core package as context variable, wrapped by contextvars.ContextVar class. The value of the variable is defaulting to private _Context class which is defined in the same module. The default values - such as database and sign - are being set on the service initialization.

The _Context class itself mimics default collection interface (as is Mapping) and can be modified by _Context.set method. The stored variables can be achieved by _Context.get method, which is unlike default Mapping interface also performs type and presence checks.

In order to provide statically typed interface, the ahriman.models.context_key.ContextKey class is used for both _Content.get and _Content.set methods; the context instance itself, however, does not store information about types.

Submodules

Some packages provide different behaviour depending on configuration settings. In these cases inheritance is used and recommended way to deal with them is to call class method load from base classes.

Authorization

The package provides several authorization methods: disabled, based on configuration and OAuth2.

Disabled (default) authorization provider just allows everything for everyone and does not have any specific configuration (it uses some default configuration parameters though). It also provides generic interface for derived classes.

Mapping (aka configuration) provider uses hashed passwords with optional salt from the database in order to authenticate users. This provider also enables user permission checking (read/write) (authorization). Thus, it defines the following methods:

check_credentials - user password validation (authentication).
verify_access - user permission validation (authorization).

Passwords must be stored in database as hash(password + salt), where password is user defined password (taken from user input), salt is random string (any length) defined globally in configuration and hash is secure hash function. Thus, the following configuration

"username","password","access"
"username","$6$rounds=656000$mWBiecMPrHAL1VgX$oU4Y5HH8HzlvMaxwkNEJjK13ozElyU1wAHBoO/WW5dAaE4YEfnB0X3FxbynKMl4FBdC3Ovap0jINz4LPkNADg0","read"

means that there is user username with read access and password password hashed by sha512 with salt salt.

OAuth provider uses library definitions (aioauth-client) in order authenticate users. It still requires user permission to be set in database, thus it inherits mapping provider without any changes. Whereas we could override check_credentials (authentication method) by something custom, OAuth flow is a bit more complex than just forward request, thus we have to implement the flow in login form.

OAuth’s implementation also allows authenticating users via username + password (in the same way as mapping does) though it is not recommended for end-users and password must be left blank. In particular this feature can be used by service reporting (aka robots).

In addition, web service checks the source socket used. In case if it belongs to socket.AF_UNIX family, it will skip any further checks considering the request to be performed in safe environment (e.g. on the same physical machine). This feature, in particular is being used by the reporter instances in case if socket address is set in configuration.

In order to configure users there are special subcommands.

Triggers

Triggers are extensions which can be used in order to perform any actions on application start, after the update process and, finally, before the application exit.

The main idea is to load classes by their full path (e.g. ahriman.core.upload.UploadTrigger) by using importlib: get the last part of the import and treat it as class name, join remain part by . and interpret as module path, import module and extract attribute from it.

The loaded triggers will be called with ahriman.models.result.Result and list[Packages] arguments, which describes the process result and current repository packages respectively. Any exception raised will be suppressed and will generate an exception message in logs.

In addition triggers can implement on_start and on_stop actions which will be called on the application start and right before the application exit respectively. The on_start action is usually being called from handlers directly in order to make sure that no trigger will be run when it is not required (e.g. on user management). As soon as on_start action is called, the additional flag will be set; ahriman.core.triggers.TriggerLoader class implements __del__ method in which, if the flag is set, the on_stop actions will be called.

For more details how to deal with the triggers, refer to documentation and modules descriptions.

Remote synchronization

There are several supported synchronization providers, currently they are rsync, s3, github.

rsync provider does not have any specific logic except for running external rsync application with configured arguments. The service does not handle SSH configuration, thus it has to be configured before running application manually.

s3 provider uses boto3 package and implements sync feature. The files are stored in architecture specific directory (e.g. if bucket is repository, packages will be stored in repository/aur-clone/x86_64 for the aur-clone repository and x86_64 architecture), bucket must be created before any action and API key must have permissions to write to the bucket. No external configuration required. In order to upload only changed files the service compares calculated hashes with the Amazon ETags, the implementation used is described here.

github provider authenticates through basic auth, API key with repository write permissions is required. There will be created a release with the name of the architecture in case if it does not exist; files will be uploaded to the release assets. It also stores array of files and their MD5 checksums in release body in order to upload only changed ones. According to the GitHub API in case if there is already uploaded asset with the same name (e.g. database files), asset will be removed first.

Additional features

Some features require optional dependencies to be installed:

Version control executables (e.g. git, svn) for VCS packages.
gnupg application for package and repository sign feature.
rsync application for rsync based repository sync.
boto3 python package for S3 sync.
Jinja2 python package for HTML report generation (it is also used by web application).

Web application

Web application requires the following python packages to be installed:

Core part requires aiohttp (application itself), aiohttp_jinja2 and Jinja2 (HTML generation from templates).
Additional web features also require aiohttp-apispec (autogenerated documentation), aiohttp_cors (CORS support, required by documentation).
In addition, authorization feature requires aiohttp_security, aiohttp_session and cryptography.
In addition to base authorization dependencies, OAuth2 also requires aioauth-client library.
In addition if you would like to disable authorization for local access (recommended way in order to run the application itself with reporting support), the requests-unixsocket library is required.

Middlewares

Service provides some custom middlewares, e.g. logging every exception (except for user ones) and user authorization.

HEAD and OPTIONS requests

HEAD request is automatically generated by ahriman.web.views.base.BaseView class. It just calls GET method, removes any data from body and returns the result. In case if no GET method available for this view, the aiohttp.web.HTTPMethodNotAllowed exception will be raised.

On the other side, OPTIONS method is implemented in the ahriman.web.middlewares.exception_handler.exception_handler middleware. In case if aiohttp.web.HTTPMethodNotAllowed exception is raised and original method was OPTIONS, the middleware handles it, converts to valid request and returns response to user.

Web views

All web views are defined in separated package and derived from ahriman.web.views.base.Base class which provides typed interfaces for web application.

REST API supports only JSON data.

Different APIs are separated into different packages:

ahriman.web.views.api not a real API, but some views which provide OpenAPI support.
ahriman.web.views.*.service provides views for application controls.
ahriman.web.views.*.status package provides REST API for application reporting.
ahriman.web.views.*.user package provides login and logout methods which can be called without authorization.

The views are also divided by supporting API versions (e.g. v1, v2).

Templating

Package provides base jinja templates which can be overridden by settings. Vanilla templates actively use bootstrap library.

Requests and scopes

Service provides optional authorization which can be turned on in settings. In order to control user access there are two levels of authorization - read-only (only GET-like requests) and write (anything), settings for which are provided by each web view directly.

If this feature is configured any request will be prohibited without authentication. In addition, configuration flag auth.allow_read_only can be used in order to allow read-only operations - reading index page and packages - without authorization.

For authenticated users it uses encrypted session cookies to store tokens; encryption key is generated each time at the start of the application. It also stores expiration time of the session inside.

External calls

Web application provides external calls to control main service. It spawns child process with specific arguments and waits for its termination. This feature must be used either with authorization or in safe (i.e. when status page is not available world-wide) environment.

For most actions it also extracts user from authentication (if provided) and passes it to the underlying process.