Chunk the notification type backfill migration

Long-term we want that migration to be done entirely in SQL,
but for now this is a hotfix to not cause OOMs on large databases.

This is using a homegrown version of `Repo.stream`, it's worse in
terms of performance than the upstream since it doesn't use the same
prepared query for chunk queries, but unlike the upstream it supports
preloads.

lib/pleroma/migration_helper/notification_backfill.ex

@@ -18,7 +18,7 @@ defmodule Pleroma.MigrationHelper.NotificationBackfill do
       )
 
     query
-    |> Repo.all()
+    |> Repo.chunk_stream(100)
     |> Enum.each(fn notification ->
       type =
         notification.activity

lib/pleroma/repo.ex

@@ -8,6 +8,7 @@ defmodule Pleroma.Repo do
     adapter: Ecto.Adapters.Postgres,
     migration_timestamps: [type: :naive_datetime_usec]
 
+  import Ecto.Query
   require Logger
 
   defmodule Instrumenter do
@@ -78,6 +79,33 @@ defmodule Pleroma.Repo do
       :ok
     end
   end
+
+  def chunk_stream(query, chunk_size) do
+    # We don't actually need start and end funcitons 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
+    Stream.resource(
+      fn -> 0 end,
+      fn
+        last_id ->
+          query
+          |> order_by(asc: :id)
+          |> where([r], r.id > ^last_id)
+          |> limit(^chunk_size)
+          |> all()
+          |> case do
+            [] ->
+              {:halt, last_id}
+
+            records ->
+              last_id = List.last(records).id
+              {records, last_id}
+          end
+      end,
+      fn _ -> :ok end
+    )
+  end
 end
 
 defmodule Pleroma.Repo.UnappliedMigrationsError do