debian-mirror-gitlab/lib/gitlab/database/tables_truncate.rb

# frozen_string_literal: true

module Gitlab
  module Database
    class TablesTruncate
      GITLAB_SCHEMAS_TO_IGNORE = %i[gitlab_geo].freeze

      def initialize(database_name:, min_batch_size:, logger: nil, until_table: nil, dry_run: false)
        @database_name = database_name
        @min_batch_size = min_batch_size
        @logger = logger
        @until_table = until_table
        @dry_run = dry_run
      end

      def execute
        raise "Cannot truncate legacy tables in single-db setup" if single_database_setup?
        raise "database is not supported" unless %w[main ci].include?(database_name)

        logger&.info "DRY RUN:" if dry_run

        schemas_for_connection = Gitlab::Database.gitlab_schemas_for_connection(connection)
        tables_to_truncate = Gitlab::Database::GitlabSchema.tables_to_schema.reject do |_, schema_name|
          GITLAB_SCHEMAS_TO_IGNORE.union(schemas_for_connection).include?(schema_name)
        end.keys

        Gitlab::Database::SharedModel.using_connection(connection) do
          Postgresql::DetachedPartition.find_each do |detached_partition|
            next if GITLAB_SCHEMAS_TO_IGNORE.union(schemas_for_connection).include?(detached_partition.table_schema)

            tables_to_truncate << detached_partition.fully_qualified_table_name
          end
        end

        tables_sorted = Gitlab::Database::TablesSortedByForeignKeys.new(connection, tables_to_truncate).execute
        # Checking if all the tables have the write-lock triggers
        # to make sure we are deleting the right tables on the right database.
        tables_sorted.flatten.each do |table_name|
          lock_writes_manager = Gitlab::Database::LockWritesManager.new(
            table_name: table_name,
            connection: connection,
            database_name: database_name,
            with_retries: true,
            logger: logger,
            dry_run: dry_run
          )

          unless lock_writes_manager.table_locked_for_writes?
            raise "Table '#{table_name}' is not locked for writes. Run the rake task gitlab:db:lock_writes first"
          end
        end

        if until_table
          table_index = tables_sorted.find_index { |tables_group| tables_group.include?(until_table) }
          raise "The table '#{until_table}' is not within the truncated tables" if table_index.nil?

          tables_sorted = tables_sorted[0..table_index]
        end

        # min_batch_size is the minimum number of new tables to truncate at each stage.
        # But in each stage we have also have to truncate the already truncated tables in the previous stages
        logger&.info "Truncating legacy tables for the database #{database_name}"
        truncate_tables_in_batches(tables_sorted)
      end

      private

      attr_accessor :database_name, :min_batch_size, :logger, :dry_run, :until_table

      def connection
        @connection ||= Gitlab::Database.database_base_models[database_name].connection
      end

      def truncate_tables_in_batches(tables_sorted)
        truncated_tables = []

        tables_sorted.flatten.each do |table|
          table_name_without_schema = ActiveRecord::ConnectionAdapters::PostgreSQL::Utils
            .extract_schema_qualified_name(table)
            .identifier

          sql_statement = "SELECT set_config('lock_writes.#{table_name_without_schema}', 'false', false)"
          logger&.info(sql_statement)
          connection.execute(sql_statement) unless dry_run

          # Temporarily unlocking writes on the attached partitions of the table.
          # Because in some cases they might have been locked for writes as well, when they used to be
          # normal tables before being converted into attached partitions.
          Gitlab::Database::SharedModel.using_connection(connection) do
            table_partitions = Gitlab::Database::PostgresPartition.for_parent_table(table_name_without_schema)
            table_partitions.each do |table_partition|
              partition_name_without_schema = ActiveRecord::ConnectionAdapters::PostgreSQL::Utils
                .extract_schema_qualified_name(table_partition.identifier)
                .identifier

              sql_statement = "SELECT set_config('lock_writes.#{partition_name_without_schema}', 'false', false)"
              logger&.info(sql_statement)
              connection.execute(sql_statement) unless dry_run
            end
          end
        end

        # We do the truncation in stages to avoid high IO
        # In each stage, we truncate the new tables along with the already truncated
        # tables before. That's because PostgreSQL doesn't allow to truncate any table (A)
        # without truncating any other table (B) that has a Foreign Key pointing to the table (A).
        # even if table (B) is empty, because it has been already truncated in a previous stage.
        tables_sorted.in_groups_of(min_batch_size, false).each do |tables_groups|
          new_tables_to_truncate = tables_groups.flatten
          logger&.info "= New tables to truncate: #{new_tables_to_truncate.join(', ')}"
          truncated_tables.push(*new_tables_to_truncate).tap(&:sort!)
          sql_statements = [
            "SET LOCAL statement_timeout = 0",
            "SET LOCAL lock_timeout = 0",
            "TRUNCATE TABLE #{truncated_tables.join(', ')} RESTRICT"
          ]

          sql_statements.each { |sql_statement| logger&.info(sql_statement) }

          next if dry_run

          connection.transaction do
            sql_statements.each { |sql_statement| connection.execute(sql_statement) }
          end
        end
      end

      def single_database_setup?
        return true unless Gitlab::Database.has_config?(:ci)

        ci_base_model = Gitlab::Database.database_base_models[:ci]
        !!Gitlab::Database.db_config_share_with(ci_base_model.connection_db_config)
      end
    end
  end
end