Clean up orphaned LFS file references during GC

At the moment, users have no way to clean up orphaned LFS objects in
their repositories. Ideally, this would happen automatically, but in
the meantime, making it part of project housekeeping - which can be
manually triggered, and is also used during project cleanup, which is
often used as part of the "remove LFS from this repository" flow -
solves a pressing need.

app/workers/git_garbage_collect_worker.rb

@@ -27,7 +27,10 @@ class GitGarbageCollectWorker # rubocop:disable Scalability/IdempotentWorker
 
     task = task.to_sym
 
-    ::Projects::GitDeduplicationService.new(project).execute if task == :gc
+    if task == :gc
+      ::Projects::GitDeduplicationService.new(project).execute
+      cleanup_orphan_lfs_file_references(project)
+    end
 
     gitaly_call(task, project.repository.raw_repository)
 
@@ -90,6 +93,13 @@ class GitGarbageCollectWorker # rubocop:disable Scalability/IdempotentWorker
     raise Gitlab::Git::CommandError.new(e)
   end
 
+  def cleanup_orphan_lfs_file_references(project)
+    return unless Feature.enabled?(:cleanup_lfs_during_gc, project)
+    return if Gitlab::Database.read_only? # GitGarbageCollectWorker may be run on a Geo secondary
+
+    ::Gitlab::Cleanup::OrphanLfsFileReferences.new(project, dry_run: false, logger: logger).run!
+  end
+
   def flush_ref_caches(project)
     project.repository.after_create_branch
     project.repository.branch_names

changelogs/unreleased/223621-prune-lfs-objects-during-git-gc.yml

+---
+title: Clean up orphaned LFS file references during GC
+merge_request: 38813
+author:
+type: added

spec/workers/git_garbage_collect_worker_spec.rb

@@ -7,7 +7,7 @@ require 'spec_helper'
 RSpec.describe GitGarbageCollectWorker do
   include GitHelpers
 
-  let(:project) { create(:project, :repository) }
+  let_it_be(:project) { create(:project, :repository) }
   let(:shell) { Gitlab::Shell.new }
   let!(:lease_uuid) { SecureRandom.uuid }
   let!(:lease_key) { "project_housekeeping:#{project.id}" }
@@ -114,6 +114,56 @@ RSpec.describe GitGarbageCollectWorker do
             subject.perform(*params)
           end
         end
+
+        context 'LFS object garbage collection' do
+          before do
+            stub_lfs_setting(enabled: true)
+          end
+
+          let_it_be(:lfs_reference) { create(:lfs_objects_project, project: project) }
+          let(:lfs_object) { lfs_reference.lfs_object }
+
+          context 'with cleanup_lfs_during_gc feature flag enabled' do
+            before do
+              stub_feature_flags(cleanup_lfs_during_gc: true)
+            end
+
+            it 'cleans up unreferenced LFS objects' do
+              expect_next_instance_of(Gitlab::Cleanup::OrphanLfsFileReferences) do |svc|
+                expect(svc.project).to eq(project)
+                expect(svc.dry_run).to be_falsy
+                expect(svc).to receive(:run!).and_call_original
+              end
+
+              subject.perform(*params)
+
+              expect(project.lfs_objects.reload).not_to include(lfs_object)
+            end
+
+            it 'does nothing if the database is read-only' do
+              expect(Gitlab::Database).to receive(:read_only?) { true }
+              expect_any_instance_of(Gitlab::Cleanup::OrphanLfsFileReferences).not_to receive(:run!)
+
+              subject.perform(*params)
+
+              expect(project.lfs_objects.reload).to include(lfs_object)
+            end
+          end
+
+          context 'with cleanup_lfs_during_gc feature flag disabled' do
+            before do
+              stub_feature_flags(cleanup_lfs_during_gc: false)
+            end
+
+            it 'does not clean up unreferenced LFS objects' do
+              expect_any_instance_of(Gitlab::Cleanup::OrphanLfsFileReferences).not_to receive(:run!)
+
+              subject.perform(*params)
+
+              expect(project.lfs_objects.reload).to include(lfs_object)
+            end
+          end
+        end
       end
 
       context 'when no lease can be obtained' do