Pirate Praveen 029fca15f4 New upstream version 13.2.1

2020-08-09 17:44:08 +05:30

17 KiB

Raw Blame History

Geo self-service framework (alpha)

NOTE: Note: This document might be subjected to change. It's a proposal we're working on and once the implementation is complete this documentation will be updated. Follow progress in the epic.

NOTE: Note: The Geo self-service framework is currently in alpha. If you need to replicate a new data type, reach out to the Geo team to discuss the options. You can contact them in #g_geo on Slack or mention @geo-team in the issue or merge request.

Geo provides an API to make it possible to easily replicate data types across Geo nodes. This API is presented as a Ruby Domain-Specific Language (DSL) and aims to make it possible to replicate data with minimal effort of the engineer who created a data type.

Nomenclature

Before digging into the API, developers need to know some Geo-specific naming conventions.

Model: A model is an Active Model, which is how it is known in the entire Rails codebase. It usually is tied to a database table. From Geo perspective, a model can have one or more resources.
Resource: A resource is a piece of data that belongs to a model and is produced by a GitLab feature. It is persisted using a storage mechanism. By default, a resource is not a replicable.
Data type: Data type is how a resource is stored. Each resource should fit in one of the data types Geo supports: :- Git repository :- Blob :- Database; For more detail, see Data types.
Geo Replicable: A Replicable is a resource Geo wants to sync across Geo nodes. There is a limited set of supported data types of replicables. The effort required to implement replication of a resource that belongs to one of the known data types is minimal.
Geo Replicator: A Geo Replicator is the object that knows how to replicate a replicable. It's responsible for: :- Firing events (producer) :- Consuming events (consumer); It's tied to the Geo Replicable data type. All replicators have a common interface that can be used to process (that is, produce and consume) events. It takes care of the communication between the primary node (where events are produced) and the secondary node (where events are consumed). The engineer who wants to incorporate Geo in their feature will use the API of replicators to make this happen.
Geo Domain-Specific Language: The syntactic sugar that allows engineers to easily specify which resources should be replicated and how.

Geo Domain-Specific Language

The replicator

First of all, you need to write a replicator. The replicators live in ee/app/replicators/geo. For each resource that needs to be replicated, there should be a separate replicator specified, even if multiple resources are tied to the same model.

For example, the following replicator replicates a package file:

module Geo
  class PackageFileReplicator < Gitlab::Geo::Replicator
    # Include one of the strategies your resource needs
    include ::Geo::BlobReplicatorStrategy

    # Specify the CarrierWave uploader needed by the used strategy
    def carrierwave_uploader
      model_record.file
    end

    # Specify the model this replicator belongs to
    def self.model
      ::Packages::PackageFile
    end
  end
end

The class name should be unique. It also is tightly coupled to the table name for the registry, so for this example the registry table will be package_file_registry.

For the different data types Geo supports there are different strategies to include. Pick one that fits your needs.

Linking to a model

To tie this replicator to the model, you need to add the following to the model code:

class Packages::PackageFile < ApplicationRecord
  include ::Gitlab::Geo::ReplicableModel

  with_replicator Geo::PackageFileReplicator
end

API

When this is set in place, it's easy to access the replicator through the model:

package_file = Packages::PackageFile.find(4) # just a random id as example
replicator = package_file.replicator

Or get the model back from the replicator:

replicator.model_record
=> <Packages::PackageFile id:4>

The replicator can be used to generate events, for example in ActiveRecord hooks:

  after_create_commit -> { replicator.publish_created_event }

Library

The framework behind all this is located in ee/lib/gitlab/geo/.

Existing Replicator Strategies

Before writing a new kind of Replicator Strategy, check below to see if your resource can already be handled by one of the existing strategies. Consult with the Geo team if you are unsure.

Blob Replicator Strategy

Models that use CarrierWave's Uploader::Base can be easily supported by Geo with the Geo::BlobReplicatorStrategy module.

First, each file should have its own primary ID and model. Geo strongly recommends treating every single file as a first-class citizen, because in our experience this greatly simplifies tracking replication and verification state.

For example, to add support for files referenced by a Widget model with a widgets table, you would perform the following steps:

Replication

Include Gitlab::Geo::ReplicableModel in the Widget class, and specify the Replicator class with_replicator Geo::WidgetReplicator.

At this point the Widget class should look like this:

# frozen_string_literal: true

class Widget < ApplicationRecord
  include ::Gitlab::Geo::ReplicableModel

  with_replicator Geo::WidgetReplicator

  mount_uploader :file, WidgetUploader

  def self.replicables_for_geo_node
    # Should be implemented. The idea of the method is to restrict
    # the set of synced items depending on synchronization settings
  end
  ...
end

Create ee/app/replicators/geo/widget_replicator.rb. Implement the #carrierwave_uploader method which should return a CarrierWave::Uploader. And implement the class method .model to return the Widget class.

# frozen_string_literal: true

module Geo
  class WidgetReplicator < Gitlab::Geo::Replicator
    include ::Geo::BlobReplicatorStrategy

    def self.model
      ::Widget
    end

    def carrierwave_uploader
      model_record.file
    end
  end
end

Create ee/spec/replicators/geo/widget_replicator_spec.rb and perform the setup necessary to define the model_record variable for the shared examples.

# frozen_string_literal: true

require 'spec_helper'

RSpec.describe Geo::WidgetReplicator do
  let(:model_record) { build(:widget) }

  it_behaves_like 'a blob replicator'
end

Create the widget_registry table so Geo secondaries can track the sync and verification state of each Widget's file:

# frozen_string_literal: true

class CreateWidgetRegistry < ActiveRecord::Migration[6.0]
  DOWNTIME = false

  disable_ddl_transaction!

  def up
    unless table_exists?(:widget_registry)
      ActiveRecord::Base.transaction do
        create_table :widget_registry, id: :bigserial, force: :cascade do |t|
          t.integer :widget_id, null: false
          t.integer :state, default: 0, null: false, limit: 2
          t.integer :retry_count, default: 0, limit: 2
          t.text :last_sync_failure
          t.datetime_with_timezone :retry_at
          t.datetime_with_timezone :last_synced_at
          t.datetime_with_timezone :created_at, null: false

          t.index :widget_id
          t.index :retry_at
          t.index :state
        end
      end
    end

    add_text_limit :widget_registry, :last_sync_failure, 255
  end

  def down
    drop_table :widget_registry
  end
end

Create ee/app/models/geo/widget_registry.rb:

# frozen_string_literal: true

class Geo::WidgetRegistry < Geo::BaseRegistry
  include Geo::ReplicableRegistry

  MODEL_CLASS = ::Widget
  MODEL_FOREIGN_KEY = :widget_id

  belongs_to :widget, class_name: 'Widget'
end

The method has_create_events? should return true in most of the cases. However, if the entity you add doesn't have the create event, don't add the method at all.

Update REGISTRY_CLASSES in ee/app/workers/geo/secondary/registry_consistency_worker.rb.

Create ee/spec/factories/geo/widget_registry.rb:

# frozen_string_literal: true

FactoryBot.define do
  factory :geo_widget_registry, class: 'Geo::WidgetRegistry' do
    widget
    state { Geo::WidgetRegistry.state_value(:pending) }

    trait :synced do
      state { Geo::WidgetRegistry.state_value(:synced) }
      last_synced_at { 5.days.ago }
    end

    trait :failed do
      state { Geo::WidgetRegistry.state_value(:failed) }
      last_synced_at { 1.day.ago }
      retry_count { 2 }
      last_sync_failure { 'Random error' }
    end

    trait :started do
      state { Geo::WidgetRegistry.state_value(:started) }
      last_synced_at { 1.day.ago }
      retry_count { 0 }
    end
  end
end

Create ee/spec/models/geo/widget_registry_spec.rb:

# frozen_string_literal: true

require 'spec_helper'

RSpec.describe Geo::WidgetRegistry, :geo, type: :model do
  let_it_be(:registry) { create(:geo_widget_registry) }

  specify 'factory is valid' do
    expect(registry).to be_valid
  end

  include_examples 'a Geo framework registry'

  describe '.find_registry_differences' do
    ... # To be implemented
  end
end

Widgets should now be replicated by Geo!

Verification

Add verification state fields to the widgets table so the Geo primary can track verification state:

# frozen_string_literal: true

class AddVerificationStateToWidgets < ActiveRecord::Migration[6.0]
  DOWNTIME = false

  def change
    add_column :widgets, :verification_retry_at, :datetime_with_timezone
    add_column :widgets, :verified_at, :datetime_with_timezone
    add_column :widgets, :verification_checksum, :binary, using: 'verification_checksum::bytea'
    add_column :widgets, :verification_failure, :string
    add_column :widgets, :verification_retry_count, :integer
  end
end

Add a partial index on verification_failure and verification_checksum to ensure re-verification can be performed efficiently:

# frozen_string_literal: true

class AddVerificationFailureIndexToWidgets < ActiveRecord::Migration[6.0]
  include Gitlab::Database::MigrationHelpers

  DOWNTIME = false

  disable_ddl_transaction!

  def up
    add_concurrent_index :widgets, :verification_failure, where: "(verification_failure IS NOT NULL)", name: "widgets_verification_failure_partial"
    add_concurrent_index :widgets, :verification_checksum, where: "(verification_checksum IS NOT NULL)", name: "widgets_verification_checksum_partial"
  end

  def down
    remove_concurrent_index :widgets, :verification_failure
    remove_concurrent_index :widgets, :verification_checksum
  end
end

To do: Add verification on secondaries. This should be done as part of Geo: Self Service Framework - First Implementation for Package File verification

Widgets should now be verified by Geo!

Metrics

Metrics are gathered by Geo::MetricsUpdateWorker, persisted in GeoNodeStatus for display in the UI, and sent to Prometheus.

Add fields widget_count, widget_checksummed_count, widget_checksum_failed_count, widget_synced_count, widget_failed_count, and widget_registry_count to GeoNodeStatus#RESOURCE_STATUS_FIELDS array in ee/app/models/geo_node_status.rb.
Add the same fields to GeoNodeStatus#PROMETHEUS_METRICS hash in ee/app/models/geo_node_status.rb.
Add the same fields to Sidekiq metrics table in doc/administration/monitoring/prometheus/gitlab_metrics.md.
Add the same fields to GET /geo_nodes/status example response in doc/api/geo_nodes.md.
Add the same fields to ee/spec/models/geo_node_status_spec.rb and ee/spec/factories/geo_node_statuses.rb.

Set widget_count in GeoNodeStatus#load_data_from_current_node:

self.widget_count = Geo::WidgetReplicator.primary_total_count

Add GeoNodeStatus#load_widgets_data to set widget_synced_count, widget_failed_count, and widget_registry_count:

def load_widget_data
  self.widget_synced_count = Geo::WidgetReplicator.synced_count
  self.widget_failed_count = Geo::WidgetReplicator.failed_count
  self.widget_registry_count = Geo::WidgetReplicator.registry_count
end

Call GeoNodeStatus#load_widgets_data in GeoNodeStatus#load_secondary_data.

Set widget_checksummed_count and widget_checksum_failed_count in GeoNodeStatus#load_verification_data:

self.widget_checksummed_count = Geo::WidgetReplicator.checksummed_count   self.widget_checksum_failed_count = Geo::WidgetReplicator.checksum_failed_count

Widget replication and verification metrics should now be available in the API, the Admin Area UI, and Prometheus!

GraphQL API

Add a new field to GeoNodeType in ee/app/graphql/types/geo/geo_node_type.rb:

field :widget_registries, ::Types::Geo::WidgetRegistryType.connection_type,
      null: true,
      resolver: ::Resolvers::Geo::WidgetRegistriesResolver,
      description: 'Find widget registries on this Geo node',
      feature_flag: :geo_self_service_framework

Add the new widget_registries field name to the expected_fields array in ee/spec/graphql/types/geo/geo_node_type_spec.rb.

Create ee/app/graphql/resolvers/geo/widget_registries_resolver.rb:

# frozen_string_literal: true

module Resolvers
  module Geo
    class WidgetRegistriesResolver < BaseResolver
      include RegistriesResolver
    end
  end
end

Create ee/spec/graphql/resolvers/geo/widget_registries_resolver_spec.rb:

# frozen_string_literal: true

require 'spec_helper'

RSpec.describe Resolvers::Geo::WidgetRegistriesResolver do
  it_behaves_like 'a Geo registries resolver', :geo_widget_registry
end

Create ee/app/finders/geo/widget_registry_finder.rb:

# frozen_string_literal: true

module Geo
  class WidgetRegistryFinder
    include FrameworkRegistryFinder
  end
end

Create ee/spec/finders/geo/widget_registry_finder_spec.rb:

# frozen_string_literal: true

require 'spec_helper'

RSpec.describe Geo::WidgetRegistryFinder do
  it_behaves_like 'a framework registry finder', :geo_widget_registry
end

Create ee/app/graphql/types/geo/widget_registry_type.rb:

# frozen_string_literal: true

module Types
  module Geo
    # rubocop:disable Graphql/AuthorizeTypes because it is included
    class WidgetRegistryType < BaseObject
      include ::Types::Geo::RegistryType

      graphql_name 'WidgetRegistry'
      description 'Represents the sync and verification state of a widget'

      field :widget_id, GraphQL::ID_TYPE, null: false, description: 'ID of the Widget'
    end
  end
end

Create ee/spec/graphql/types/geo/widget_registry_type_spec.rb:

# frozen_string_literal: true

require 'spec_helper'

RSpec.describe GitlabSchema.types['WidgetRegistry'] do
  it_behaves_like 'a Geo registry type'

  it 'has the expected fields (other than those included in RegistryType)' do
    expected_fields = %i[widget_id]

    expect(described_class).to have_graphql_fields(*expected_fields).at_least
  end
end

Add integration tests for providing Widget registry data to the frontend via the GraphQL API, by duplicating and modifying the following shared examples in ee/spec/requests/api/graphql/geo/registries_spec.rb:

it_behaves_like 'gets registries for', {
  field_name: 'widgetRegistries',
  registry_class_name: 'WidgetRegistry',
  registry_factory: :geo_widget_registry,
  registry_foreign_key_field_name: 'widgetId'
}

Individual widget synchronization and verification data should now be available via the GraphQL API!

Take care of replicating "update" events. Geo Framework does not currently support replicating "update" events because all entities added to the framework, by this time, are immutable. If this is the case for the entity you're going to add, please follow https://gitlab.com/gitlab-org/gitlab/-/issues/118743 and https://gitlab.com/gitlab-org/gitlab/-/issues/118745 as examples to add the new event type. Please also remove this notice when you've added it.

Admin UI

To do: This should be done as part of Geo: Implement frontend for Self-Service Framework replicables

Widget sync and verification data (aggregate and individual) should now be available in the Admin UI!

17 KiB Raw Blame History