Recover from all exceptions when stealing bg migration

It also makes it possible to gracefully retry a migration in order to
avoid problems like deadlocks.

lib/gitlab/background_migration.rb

@@ -7,6 +7,12 @@ module Gitlab
     # Begins stealing jobs from the background migrations queue, blocking the
     # caller until all jobs have been completed.
     #
+    # When a migration raises a StandardError is is going to be retries up to
+    # three times, for example, to recover from a deadlock.
+    #
+    # When Exception is being raised, it enqueues the migration again, and
+    # re-raises the exception.
+    #
     # steal_class - The name of the class for which to steal jobs.
     def self.steal(steal_class)
       enqueued = Sidekiq::Queue.new(self.queue)
@@ -20,22 +26,34 @@ module Gitlab
           next unless migration_class == steal_class
 
           begin
-            perform(migration_class, migration_args) if job.delete
-          rescue => e
-            Logger.new($stdout).warn(e.message)
+            perform(migration_class, migration_args, retries: 3) if job.delete
+          rescue StandardError
             next
+          rescue Exception
+            BackgroundMigrationWorker # enqueue this migration again
+              .perform_async(migration_class, migration_args)
+
+            raise
           end
         end
       end
     end
 
+    ##
+    # Performs a background migration. In case of `StandardError` being caught
+    # this will retry a migration up to three times.
+    #
     # class_name - The name of the background migration class as defined in the
     #              Gitlab::BackgroundMigration namespace.
     #
     # arguments - The arguments to pass to the background migration's "perform"
     #             method.
-    def self.perform(class_name, arguments)
+    def self.perform(class_name, arguments, retries: 1)
       const_get(class_name).new.perform(*arguments)
+    rescue => e
+      Rails.logger.warn("Retrying background migration #{class_name} " \
+                        "with #{arguments}")
+      (retries -= 1) > 0 ? retry : raise
     end
   end
 end

spec/lib/gitlab/background_migration_spec.rb

@@ -24,7 +24,8 @@ describe Gitlab::BackgroundMigration do
         it 'steals jobs from a queue' do
           expect(queue[0]).to receive(:delete).and_return(true)
 
-          expect(described_class).to receive(:perform).with('Foo', [10, 20])
+          expect(described_class).to receive(:perform)
+            .with('Foo', [10, 20], anything)
 
           described_class.steal('Foo')
         end
@@ -32,7 +33,7 @@ describe Gitlab::BackgroundMigration do
         it 'does not steal job that has already been taken' do
           expect(queue[0]).to receive(:delete).and_return(false)
 
-          expect(described_class).not_to receive(:perform).with('Foo', [10, 20])
+          expect(described_class).not_to receive(:perform)
 
           described_class.steal('Foo')
         end
@@ -57,17 +58,40 @@ describe Gitlab::BackgroundMigration do
         before do
           stub_const("#{described_class}::Foo", migration)
 
+          allow(queue[0]).to receive(:delete).and_return(true)
+          allow(queue[1]).to receive(:delete).and_return(true)
+        end
+
+        context 'when standard error is being raised' do
+          before do
             allow(migration).to receive(:perform).with(10, 20)
               .and_raise(StandardError, 'Migration error')
+          end
 
-          allow(queue[0]).to receive(:delete).and_return(true)
-          allow(queue[1]).to receive(:delete).and_return(true)
+          it 'recovers from an exception and retries the migration' do
+            expect(migration).to receive(:perform).with(10, 20)
+              .exactly(3).times.ordered
+            expect(migration).to receive(:perform).with(20, 30)
+              .once.ordered
+            expect(Rails.logger).to receive(:warn)
+              .with(/Retrying background migration/).exactly(3).times
+
+            described_class.steal('Foo')
           end
+        end
+
+        context 'when top level exception is being raised' do
+          it 'enqueues the migration again and reraises the error' do
+            allow(migration).to receive(:perform).with(10, 20)
+              .and_raise(Exception, 'Migration error').once
+
+            expect(BackgroundMigrationWorker).to receive(:perform_async)
+              .with('Foo', [10, 20]).once
 
-        it 'recovers from an exceptions and continues' do
-          expect(migration).to receive(:perform).twice
+            expect(Rails.logger).not_to receive(:warn)
             expect { described_class.steal('Foo') }
-            .to output(/Migration error/).to_stdout
+              .to raise_error(Exception)
+          end
         end
       end
     end
@@ -91,9 +115,9 @@ describe Gitlab::BackgroundMigration do
       it 'steals from the scheduled sets queue first' do
         Sidekiq::Testing.disable! do
           expect(described_class).to receive(:perform)
-            .with('Object', [1]).ordered
+            .with('Object', [1], anything).ordered
           expect(described_class).to receive(:perform)
-            .with('Object', [2]).ordered
+            .with('Object', [2], anything).ordered
 
           BackgroundMigrationWorker.perform_async('Object', [2])
           BackgroundMigrationWorker.perform_in(10.minutes, 'Object', [1])
@@ -105,17 +129,38 @@ describe Gitlab::BackgroundMigration do
   end
 
   describe '.perform' do
-    it 'performs a background migration' do
-      instance = double(:instance)
-      klass = double(:klass, new: instance)
+    let(:migration) { spy(:migration) }
 
-      expect(described_class).to receive(:const_get)
-        .with('Foo')
-        .and_return(klass)
+    before do
+      stub_const("#{described_class.name}::Foo", migration)
+    end
 
-      expect(instance).to receive(:perform).with(10, 20)
+    context 'when retries count is not specified' do
+      it 'performs a background migration' do
+        expect(migration).to receive(:perform).with(10, 20).once
 
         described_class.perform('Foo', [10, 20])
       end
     end
+
+    context 'when retries count is zero' do
+      it 'perform a background migration only once' do
+        expect(migration).to receive(:perform).with(10, 20)
+          .and_raise(StandardError).once
+
+        expect {  described_class.perform('Foo', [10, 20], retries: 0) }
+          .to raise_error(StandardError)
+      end
+    end
+
+    context 'when retries count is larger than zero' do
+      it 'retries a background migration when needed' do
+        expect(migration).to receive(:perform).with(10, 20)
+          .and_raise(StandardError).exactly(3).times
+
+        expect {  described_class.perform('Foo', [10, 20], retries: 3) }
+          .to raise_error(StandardError)
+      end
+    end
+  end
 end