[#3259] Background replication of objects' attachments into media table.

This commit is contained in:
Ivan Tashkinov 2021-01-23 22:59:24 +03:00
parent 4ade12bf82
commit 4b10d15e62
9 changed files with 434 additions and 4 deletions

View File

@ -104,7 +104,8 @@ defmodule Pleroma.Application do
chat_child(chat_enabled?()) ++
[
Pleroma.Web.Endpoint,
Pleroma.Gopher.Server
Pleroma.Gopher.Server,
Pleroma.Migrators.MediaTableMigrator
]
# See http://elixir-lang.org/docs/stable/elixir/Supervisor.html

View File

@ -0,0 +1,46 @@
# Pleroma: A lightweight social networking server
# Copyright © 2017-2021 Pleroma Authors <https://pleroma.social/>
# SPDX-License-Identifier: AGPL-3.0-only
defmodule Pleroma.DataMigration do
use Ecto.Schema
alias Pleroma.DataMigration
alias Pleroma.DataMigration.State
alias Pleroma.Repo
import Ecto.Changeset
schema "data_migrations" do
field(:name, :string)
field(:state, State, default: :pending)
field(:feature_lock, :boolean, default: false)
field(:params, :map, default: %{})
field(:data, :map, default: %{})
timestamps()
end
def changeset(data_migration, params \\ %{}) do
data_migration
|> cast(params, [:name, :state, :feature_lock, :params, :data])
|> validate_required([:name])
|> unique_constraint(:name)
end
def update(data_migration, params \\ %{}) do
data_migration
|> changeset(params)
|> Repo.update()
end
def update_state(data_migration, new_state) do
update(data_migration, %{state: new_state})
end
def get_by_name(name) do
Repo.get_by(DataMigration, name: name)
end
def populate_media_table, do: get_by_name("populate_media_table")
end

View File

@ -17,3 +17,11 @@ defenum(Pleroma.FollowingRelationship.State,
follow_accept: 2,
follow_reject: 3
)
defenum(Pleroma.DataMigration.State,
pending: 1,
running: 2,
complete: 3,
failed: 4,
manual: 5
)

View File

@ -0,0 +1,299 @@
# Pleroma: A lightweight social networking server
# Copyright © 2017-2021 Pleroma Authors <https://pleroma.social/>
# SPDX-License-Identifier: AGPL-3.0-only
defmodule Pleroma.Migrators.MediaTableMigrator do
use GenServer
require Logger
import Ecto.Query
alias __MODULE__.State
alias Pleroma.Config
alias Pleroma.DataMigration
alias Pleroma.Media
alias Pleroma.Object
alias Pleroma.Repo
defdelegate state(), to: State, as: :get
defdelegate put_stat(key, value), to: State, as: :put
defdelegate increment_stat(key, increment), to: State, as: :increment
defdelegate data_migration(), to: DataMigration, as: :populate_media_table
@reg_name {:global, __MODULE__}
def whereis, do: GenServer.whereis(@reg_name)
def start_link(_) do
case whereis() do
nil ->
GenServer.start_link(__MODULE__, nil, name: @reg_name)
pid ->
{:ok, pid}
end
end
@impl true
def init(_) do
{:ok, nil, {:continue, :init_state}}
end
@impl true
def handle_continue(:init_state, _state) do
{:ok, _} = State.start_link(nil)
update_status(:init)
data_migration = data_migration()
manual_migrations = Config.get([:instance, :manual_data_migrations], [])
cond do
Config.get(:env) == :test ->
update_status(:noop)
is_nil(data_migration) ->
update_status(:halt, "Data migration does not exist.")
data_migration.state == :manual or data_migration.name in manual_migrations ->
update_status(:noop, "Data migration is in manual execution state.")
data_migration.state == :complete ->
handle_success(data_migration)
true ->
send(self(), :process_attachments)
end
{:noreply, nil}
end
@impl true
def handle_info(:process_attachments, state) do
data_migration = data_migration()
persistent_data = Map.take(data_migration.data, ["max_processed_id"])
{:ok, data_migration} =
DataMigration.update(data_migration, %{state: :running, data: persistent_data})
update_status(:running)
put_stat(:started_at, NaiveDateTime.utc_now())
Logger.info("Starting creating `media` records for objects' attachments...")
max_processed_id = data_migration.data["max_processed_id"] || 0
query()
|> where([object], object.id > ^max_processed_id)
|> Repo.chunk_stream(100, :batches, timeout: :infinity)
|> Stream.each(fn objects ->
object_ids = Enum.map(objects, & &1.id)
failed_ids =
objects
|> Enum.map(&process_object_attachments(&1))
|> Enum.filter(&(elem(&1, 0) == :error))
|> Enum.map(&elem(&1, 1))
for failed_id <- failed_ids do
_ =
Repo.query(
"INSERT INTO data_migration_failed_ids(data_migration_id, record_id) " <>
"VALUES ($1, $2) ON CONFLICT DO NOTHING;",
[data_migration.id, failed_id]
)
end
_ =
Repo.query(
"DELETE FROM data_migration_failed_ids " <>
"WHERE data_migration_id = $1 AND record_id = ANY($2)",
[data_migration.id, object_ids -- failed_ids]
)
max_object_id = Enum.at(object_ids, -1)
put_stat(:max_processed_id, max_object_id)
increment_stat(:processed_count, length(object_ids))
increment_stat(:failed_count, length(failed_ids))
put_stat(
:records_per_second,
state()[:processed_count] /
Enum.max([NaiveDateTime.diff(NaiveDateTime.utc_now(), state()[:started_at]), 1])
)
persist_stats(data_migration)
# A quick and dirty approach to controlling the load this background migration imposes
sleep_interval = Config.get([:populate_media_table, :sleep_interval_ms], 0)
Process.sleep(sleep_interval)
end)
|> Stream.run()
with 0 <- failures_count(data_migration.id) do
{:ok, data_migration} = DataMigration.update_state(data_migration, :complete)
handle_success(data_migration)
else
_ ->
_ = DataMigration.update_state(data_migration, :failed)
update_status(:failed, "Please check data_migration_failed_ids records.")
end
{:noreply, state}
end
def query do
from(
object in Object,
where:
fragment(
"(?)->'attachment' IS NOT NULL AND \
(?)->'attachment' != ANY(ARRAY['null'::jsonb, '[]'::jsonb])",
object.data,
object.data
),
select: %{
id: object.id,
attachment: fragment("(?)->'attachment'", object.data),
actor: fragment("(?)->'actor'", object.data)
}
)
end
defp process_object_attachments(object) do
attachments =
if Map.has_key?(object, :attachment), do: object.attachment, else: object.data["attachment"]
actor = if Map.has_key?(object, :actor), do: object.actor, else: object.data["actor"]
Repo.transaction(fn ->
with {_, true} <- {:any, Enum.any?(attachments || [], &is_nil(&1["id"]))},
updated_attachments =
Enum.map(attachments, fn attachment ->
if is_nil(attachment["id"]) do
with {:ok, media} <-
Media.create_from_object_data(attachment, %{
actor: actor,
object_id: object.id
}) do
Map.put(attachment, "id", media.id)
else
{:error, e} ->
error =
"ERROR: could not process attachment of object #{object.id}: " <>
"#{attachment["href"]}: #{inspect(e)}"
Logger.error(error)
Repo.rollback(object.id)
end
else
attachment
end
end),
{:ok, _} <-
Object.update_data(%Object{id: object.id}, %{"attachment" => updated_attachments}) do
object.id
else
{:any, false} ->
object.id
{:error, e} ->
error = "ERROR: could not update attachments of object #{object.id}: #{inspect(e)}"
Logger.error(error)
Repo.rollback(object.id)
end
end)
end
@doc "Approximate count for current iteration (including processed records count)"
def count(force \\ false, timeout \\ :infinity) do
stored_count = state()[:count]
if stored_count && !force do
stored_count
else
processed_count = state()[:processed_count] || 0
max_processed_id = data_migration().data["max_processed_id"] || 0
query = where(query(), [object], object.id > ^max_processed_id)
count = Repo.aggregate(query, :count, :id, timeout: timeout) + processed_count
put_stat(:count, count)
count
end
end
defp persist_stats(data_migration) do
runner_state = Map.drop(state(), [:status])
_ = DataMigration.update(data_migration, %{data: runner_state})
end
defp handle_success(_data_migration) do
update_status(:complete)
end
def failed_objects_query do
from(o in Object)
|> join(:inner, [o], dmf in fragment("SELECT * FROM data_migration_failed_ids"),
on: dmf.record_id == o.id
)
|> where([_o, dmf], dmf.data_migration_id == ^data_migration().id)
|> order_by([o], asc: o.id)
end
def failures_count(data_migration_id \\ nil) do
data_migration_id = data_migration_id || data_migration().id
with {:ok, %{rows: [[count]]}} <-
Repo.query(
"SELECT COUNT(record_id) FROM data_migration_failed_ids WHERE data_migration_id = $1;",
[data_migration_id]
) do
count
end
end
def retry_failed do
data_migration = data_migration()
failed_objects_query()
|> Repo.chunk_stream(100, :one)
|> Stream.each(fn object ->
with {:ok, _} <- process_object_attachments(object) do
_ =
Repo.query(
"DELETE FROM data_migration_failed_ids " <>
"WHERE data_migration_id = $1 AND record_id = $2",
[data_migration.id, object.id]
)
end
end)
|> Stream.run()
end
def force_continue do
send(whereis(), :process_attachments)
end
def force_restart do
{:ok, _} = DataMigration.update(data_migration(), %{state: :pending, data: %{}})
force_continue()
end
def force_complete do
{:ok, data_migration} = DataMigration.update_state(data_migration(), :complete)
handle_success(data_migration)
end
defp update_status(status, message \\ nil) do
put_stat(:status, status)
put_stat(:message, message)
end
end

View File

@ -0,0 +1,31 @@
# Pleroma: A lightweight social networking server
# Copyright © 2017-2021 Pleroma Authors <https://pleroma.social/>
# SPDX-License-Identifier: AGPL-3.0-only
defmodule Pleroma.Migrators.MediaTableMigrator.State do
use Agent
@init_state %{}
@reg_name {:global, __MODULE__}
def start_link(_) do
Agent.start_link(fn -> @init_state end, name: @reg_name)
end
def get do
Agent.get(@reg_name, & &1)
end
def put(key, value) do
Agent.update(@reg_name, fn state ->
Map.put(state, key, value)
end)
end
def increment(key, increment \\ 1) do
Agent.update(@reg_name, fn state ->
updated_value = (state[key] || 0) + increment
Map.put(state, key, updated_value)
end)
end
end

View File

@ -63,8 +63,8 @@ defmodule Pleroma.Repo do
iex> Pleroma.Repo.chunk_stream(Pleroma.Activity.Queries.by_actor(ap_id), 500, :batches)
"""
@spec chunk_stream(Ecto.Query.t(), integer(), atom()) :: Enumerable.t()
def chunk_stream(query, chunk_size, returns_as \\ :one) do
# We don't actually need start and end funcitons of resource streaming,
def chunk_stream(query, chunk_size, returns_as \\ :one, query_options \\ []) do
# We don't actually need start and end functions of resource streaming,
# but it seems to be the only way to not fetch records one-by-one and
# have individual records be the elements of the stream, instead of
# lists of records
@ -76,7 +76,7 @@ defmodule Pleroma.Repo do
|> order_by(asc: :id)
|> where([r], r.id > ^last_id)
|> limit(^chunk_size)
|> all()
|> all(query_options)
|> case do
[] ->
{:halt, last_id}

View File

@ -0,0 +1,17 @@
defmodule Pleroma.Repo.Migrations.CreateDataMigrations do
use Ecto.Migration
def change do
create_if_not_exists table(:data_migrations) do
add(:name, :string, null: false)
add(:state, :integer, default: 1)
add(:feature_lock, :boolean, default: false)
add(:params, :map, default: %{})
add(:data, :map, default: %{})
timestamps()
end
create_if_not_exists(unique_index(:data_migrations, [:name]))
end
end

View File

@ -0,0 +1,14 @@
defmodule Pleroma.Repo.Migrations.CreateDataMigrationFailedIds do
use Ecto.Migration
def change do
create_if_not_exists table(:data_migration_failed_ids, primary_key: false) do
add(:data_migration_id, references(:data_migrations), null: false)
add(:record_id, :bigint, null: false)
end
create_if_not_exists(
unique_index(:data_migration_failed_ids, [:data_migration_id, :record_id])
)
end
end

View File

@ -0,0 +1,14 @@
defmodule Pleroma.Repo.Migrations.DataMigrationCreatePopulateMediaTable do
use Ecto.Migration
def up do
dt = NaiveDateTime.utc_now()
execute(
"INSERT INTO data_migrations(name, inserted_at, updated_at) " <>
"VALUES ('populate_media_table', '#{dt}', '#{dt}') ON CONFLICT DO NOTHING;"
)
end
def down, do: :ok
end