Fix tag updates atomicity and deadlocks

lib/philomena/images.ex

@@ -96,22 +96,25 @@ defmodule Philomena.Images do
     |> Multi.insert(:image, image)
     |> Multi.run(:added_tag_count, fn repo, %{image: image} ->
       tag_ids = image.added_tags |> Enum.map(& &1.id)
-      tags = Tag |> where([t], t.id in ^tag_ids)
 
-      {count, nil} = repo.update_all(tags, inc: [images_count: 1])
+      count = Tags.update_image_counts(repo, 1, tag_ids)
 
       {:ok, count}
     end)
     |> maybe_subscribe_on(:image, attribution[:user], :watch_on_upload)
     |> Repo.transaction()
     |> case do
-      {:ok, %{image: image}} = result ->
-        async_upload(image, attrs["image"])
+      {:ok, %{image: image}} ->
+        upload_pid = async_upload(image, attrs["image"])
         reindex_image(image)
         Tags.reindex_tags(image.added_tags)
         maybe_approve_image(image, attribution[:user])
 
-        result
+        # Return the upload PID along with the created image so that the caller
+        # can control the lifecycle of the upload if needed. It's useful, for
+        # example for the seeding process to know when to delete the temp file
+        # used for uploading.
+        {:ok, %{image: image, upload_pid: upload_pid}}
 
       result ->
         result
@@ -138,6 +141,8 @@ defmodule Philomena.Images do
 
     # Free up the linked process
     send(linked_pid, :ready)
+
+    linked_pid
   end
 
   defp try_upload(image, retry_count) when retry_count < 100 do
@@ -666,9 +671,8 @@ defmodule Philomena.Images do
 
       repo, %{image: {_image, _added, removed_tags}} ->
         tag_ids = removed_tags |> Enum.map(& &1.id)
-        tags = Tag |> where([t], t.id in ^tag_ids)
 
-        {count, nil} = repo.update_all(tags, inc: [images_count: -1])
+        count = Tags.update_image_counts(repo, -1, tag_ids)
 
         {:ok, count}
     end)
@@ -955,9 +959,8 @@ defmodule Philomena.Images do
       # to way too much drift, and the index has to be
       # maintained.
       tag_ids = Enum.map(image.tags, & &1.id)
-      query = where(Tag, [t], t.id in ^tag_ids)
 
-      repo.update_all(query, inc: [images_count: -1])
+      Tags.update_image_counts(repo, -1, tag_ids)
 
       {:ok, image.tags}
     end)

lib/philomena/tags.ex

@@ -4,6 +4,7 @@ defmodule Philomena.Tags do
   """
 
   import Ecto.Query, warn: false
+  alias Ecto.Multi
   alias Philomena.Repo
 
   alias PhilomenaQuery.Search
@@ -24,6 +25,47 @@ defmodule Philomena.Tags do
   alias Philomena.DnpEntries.DnpEntry
   alias Philomena.Channels.Channel
 
+  # There is a really delicate nuance that must be known to avoid deadlocks in
+  # vectorized mutation queries such as `INSERT ON CONFLICT UPDATE`, `UPDATE`,
+  # `DELETE`, `SELECT FOR [NO KEY] UPDATE` that touch multiple records. Note that
+  # `INSERT ON CONFLICT DO NOTHING` doesn't lock the conflicting records, so this
+  # nuance doesn't apply in that case (https://dba.stackexchange.com/questions/322912/will-insert-on-conflict-do-nothing-lock-the-row-in-case-of-conflict)
+  #
+  # If a vectorized mutation is run without a consistent locking order of the records,
+  # it can end up with a deadlock where one transaction locks a set of records
+  # that overlap with the other transaction while the other transaction locks
+  # the other set that overlaps with the first transaction. Thus, both transactions
+  # wait for each other to release the locks on records they locked resulting in
+  # a deadlock.
+  #
+  # For raw `UPDATE/DELETE ... WHERE ... IN (...)` queries, the items inside `IN (...)`
+  # don't influence the order of locking. These queries also don't have an `ORDER BY`
+  # clause. Thus, this function returns a `SELECT [lock_type]` query that establishes a
+  # consistent order of records by primary keys that must be used with all vectorized
+  # mutation queries to avoid deadlocks. This query can be used as a subquery in
+  # the `WHERE` clause for the vectorized mutation.
+  #
+  # If no locking order is set, the deadlock can appear randomly and its probability
+  # increases with the amount of items in the vectorized mutation query and with
+  # the number of overlapping records in concurrent transactions.
+  #
+  # This phenomena was discovered when @MareStare was trying to parallelize
+  # the image creation process for seeding the images during development, where
+  # tons of image uploads are issued in parallel with many overlapping tags
+  # (https://github.com/philomena-dev/philomena/pull/481).
+  #
+  # Big thanks to this StackOverflow post for explanations:
+  # https://stackoverflow.com/questions/27262900/postgres-update-and-lock-ordering/27263824#27263824
+  defmacro vectorized_mutation_lock(lock_type, tag_ids) do
+    quote do
+      Tag
+      |> select([t], t.id)
+      |> lock(unquote(lock_type))
+      |> where([t], t.id in ^unquote(tag_ids))
+      |> order_by([t], t.id)
+    end
+  end
+
   @doc """
   Gets existing tags or creates new ones from a tag list string.
 
@@ -39,43 +81,72 @@ defmodule Philomena.Tags do
   """
   @spec get_or_create_tags(String.t()) :: list()
   def get_or_create_tags(tag_list) do
-    tag_names = Tag.parse_tag_list(tag_list)
-
-    existent_tags =
-      Tag
-      |> where([t], t.name in ^tag_names)
-      |> preload([:implied_tags, aliased_tag: :implied_tags])
-      |> Repo.all()
-      |> Enum.uniq_by(& &1.name)
-
-    existent_tag_names =
-      existent_tags
-      |> Map.new(&{&1.name, true})
+    case Tag.parse_tag_list(tag_list) do
+      [] -> []
+      tag_names -> get_or_create_non_empty_tags_list(tag_names)
+    end
+  end
 
-    nonexistent_tag_names =
+  @spec get_or_create_non_empty_tags_list(list(String.t())) :: list()
+  defp get_or_create_non_empty_tags_list(tag_names) do
+    tags =
       tag_names
-      |> Enum.reject(&existent_tag_names[&1])
-
-    # Now get rid of the aliases
-    existent_tags =
-      existent_tags
-      |> Enum.map(&(&1.aliased_tag || &1))
-
-    new_tags =
-      nonexistent_tag_names
-      |> Enum.map(fn name ->
-        {:ok, tag} =
-          %Tag{}
-          |> Tag.creation_changeset(%{name: name})
-          |> Repo.insert()
-
-        %{tag | implied_tags: []}
+      |> Enum.sort()
+      |> Enum.dedup()
+      |> Enum.map(fn tag_name ->
+        %Tag{}
+        |> Tag.creation_changeset(%{name: tag_name})
+        |> Ecto.Changeset.apply_changes()
+        |> Map.take([
+          :slug,
+          :name,
+          :category,
+          :images_count,
+          :description,
+          :short_description,
+          :namespace,
+          :name_in_namespace,
+          :image,
+          :image_format,
+          :image_mime_type,
+          :mod_notes
+        ])
+        |> Map.merge(%{
+          created_at: {:placeholder, :timestamp},
+          updated_at: {:placeholder, :timestamp}
+        })
       end)
 
+    %{new_tags: {_rows_affected, new_tags}, all_tags: all_tags} =
+      Multi.new()
+      |> Multi.insert_all(
+        :new_tags,
+        Tag,
+        tags,
+        placeholders: %{timestamp: DateTime.utc_now() |> DateTime.truncate(:second)},
+        on_conflict: :nothing,
+        returning: [:id]
+      )
+      |> Multi.all(
+        :all_tags,
+        Tag
+        |> where([t], t.name in ^tag_names)
+        |> distinct([t], t.name)
+        |> preload([:implied_tags, aliased_tag: :implied_tags])
+      )
+      |> Repo.transaction()
+      |> case do
+        {:ok, ok} ->
+          ok
+
+        result ->
+          raise "get_or_create_tags failed: #{inspect(result)}\ntag_names: #{inspect(tag_names)}"
+      end
+
     new_tags
     |> reindex_tags()
 
-    existent_tags ++ new_tags
+    all_tags
   end
 
   @doc """
@@ -545,9 +616,7 @@ defmodule Philomena.Tags do
 
         tag_ids = Enum.map(taggings, & &1.tag_id)
 
-        Tag
-        |> where([t], t.id in ^tag_ids)
-        |> Repo.update_all(inc: [images_count: 1])
+        update_image_counts(Repo, 1, tag_ids)
 
         tag_ids
       end)
@@ -557,6 +626,25 @@ defmodule Philomena.Tags do
     |> Repo.all()
   end
 
+  @doc """
+  Accepts IDs of tags and increments their `images_count` by 1.
+  """
+  @spec update_image_counts(term(), integer(), [integer()]) :: integer()
+  def update_image_counts(repo, diff, tag_ids)
+
+  def update_image_counts(_repo, _diff, []), do: 0
+
+  def update_image_counts(repo, diff, tag_ids) do
+    locked_tags = vectorized_mutation_lock("FOR NO KEY UPDATE", tag_ids)
+
+    {rows_affected, _} =
+      Tag
+      |> where([t], t.id in subquery(locked_tags))
+      |> repo.update_all(inc: [images_count: diff])
+
+    rows_affected
+  end
+
   @doc """
   Returns an `%Ecto.Changeset{}` for tracking tag changes.