Merge branch 'mwaw/use_postgres_hll_batch_counting_in_usage_ping' into 'master'

Apply Postgres HLL distributed batch counter to Usage Ping secure pipelines metrics

See merge request gitlab-org/gitlab!48233

config/feature_flags/development/postgres_hll_batch_counting.yml

+---
+name: postgres_hll_batch_counting
+introduced_by_url: https://gitlab.com/gitlab-org/gitlab/-/merge_requests/48233
+rollout_issue_url: https://gitlab.com/gitlab-org/gitlab/-/issues/285485
+milestone: '13.7'
+type: development
+group: group::product analytics
+default_enabled: false

ee/lib/ee/gitlab/usage_data.rb

@@ -396,16 +396,44 @@ module EE
         def count_secure_pipelines(time_period)
           return {} if time_period.blank?
 
-          start = ::Ci::Pipeline.minimum(:id)
-          finish = ::Ci::Pipeline.maximum(:id)
           pipelines_with_secure_jobs = {}
 
-          ::Security::Scan.scan_types.each do |name, scan_type|
-            relation = ::Ci::Build.joins(:security_scans)
-                                .where(status: 'success', retried: [nil, false])
-                                .where('security_scans.scan_type = ?', scan_type)
-                                .where(time_period)
-            pipelines_with_secure_jobs["#{name}_pipeline".to_sym] = distinct_count(relation, :commit_id, start: start, finish: finish, batch: false)
+          # HLL batch counting always iterate over pkey of
+          # given relation, while ordinary batch count
+          # iterated over counted attribute, one-to-many joins
+          # can break batch size limitation, and lead to
+          # time outing batch queries, to avoid that
+          # different join strategy is used for HLL counter
+          if ::Feature.enabled?(:postgres_hll_batch_counting)
+            relation = ::Security::Scan.where(time_period).group(:created_at)
+
+            start = relation.select('MIN(id) as min_id').order('min_id ASC').first&.min_id
+            finish = relation.select('MAX(id) as max_id').order('max_id DESC').first&.max_id
+
+            ::Security::Scan.scan_types.each do |name, scan_type|
+              relation = ::Security::Scan.joins(:build)
+                           .where(ci_builds: { status: 'success', retried: [nil, false] })
+                           .where('security_scans.scan_type = ?', scan_type)
+                           .where(security_scans: time_period)
+
+              pipelines_with_secure_jobs["#{name}_pipeline".to_sym] =
+                if start && finish
+                  estimate_batch_distinct_count(relation, :commit_id, batch_size: 1000, start: start, finish: finish)
+                else
+                  0
+                end
+            end
+          else
+            start = ::Ci::Pipeline.minimum(:id)
+            finish = ::Ci::Pipeline.maximum(:id)
+
+            ::Security::Scan.scan_types.each do |name, scan_type|
+              relation = ::Ci::Build.joins(:security_scans)
+                           .where(status: 'success', retried: [nil, false])
+                           .where('security_scans.scan_type = ?', scan_type)
+                           .where(time_period)
+              pipelines_with_secure_jobs["#{name}_pipeline".to_sym] = distinct_count(relation, :commit_id, start: start, finish: finish, batch: false)
+            end
           end
 
           pipelines_with_secure_jobs

lib/gitlab/database/postgres_hll_batch_distinct_counter.rb

@@ -28,13 +28,14 @@ module Gitlab
     #  for given implementation no higher value was reported (https://gitlab.com/gitlab-org/gitlab/-/merge_requests/45673#accuracy-estimation) than 5.3%
     #  for the most of a cases this value is lower. However, if the exact value is necessary other tools has to be used.
     class PostgresHllBatchDistinctCounter
+      ERROR_RATE = 4.9 # max encountered empirical error rate, used in tests
       FALLBACK = -1
-      MIN_REQUIRED_BATCH_SIZE = 1_250
-      MAX_ALLOWED_LOOPS = 10_000
+      MIN_REQUIRED_BATCH_SIZE = 750
       SLEEP_TIME_IN_SECONDS = 0.01 # 10 msec sleep
+      MAX_DATA_VOLUME = 4_000_000_000
 
       # Each query should take < 500ms https://gitlab.com/gitlab-org/gitlab/-/merge_requests/22705
-      DEFAULT_BATCH_SIZE = 100_000
+      DEFAULT_BATCH_SIZE = 10_000
 
       BIT_31_MASK = "B'0#{'1' * 31}'"
       BIT_9_MASK = "B'#{'0' * 23}#{'1' * 9}'"
@@ -49,7 +50,7 @@ module Gitlab
         SELECT (attr_hash_32_bits & #{BIT_9_MASK})::int AS bucket_num,
           (31 - floor(log(2, min((attr_hash_32_bits & #{BIT_31_MASK})::int))))::int as bucket_hash
         FROM hashed_attributes
-        GROUP BY 1 ORDER BY 1
+        GROUP BY 1
       SQL
 
       TOTAL_BUCKETS_NUMBER = 512
@@ -61,7 +62,7 @@ module Gitlab
 
       def unwanted_configuration?(finish, batch_size, start)
         batch_size <= MIN_REQUIRED_BATCH_SIZE ||
-          (finish - start) / batch_size >= MAX_ALLOWED_LOOPS ||
+          (finish - start) >= MAX_DATA_VOLUME ||
           start > finish
       end
 
@@ -101,7 +102,7 @@ module Gitlab
 
         num_uniques = (
           ((TOTAL_BUCKETS_NUMBER**2) * (0.7213 / (1 + 1.079 / TOTAL_BUCKETS_NUMBER))) /
-            (num_zero_buckets + hll_blob.values.sum { |bucket_hash, _| 2**(-1 * bucket_hash)} )
+            (num_zero_buckets + hll_blob.values.sum { |bucket_hash| 2**(-1 * bucket_hash)} )
         ).to_i
 
         if num_zero_buckets > 0 && num_uniques < 2.5 * TOTAL_BUCKETS_NUMBER

lib/gitlab/usage_data_queries.rb

@@ -25,6 +25,13 @@ module Gitlab
         relation.select(relation.all.table[column].sum).to_sql
       end
 
+      # For estimated distinct count use exact query instead of hll
+      # buckets query, because it can't be used to obtain estimations without
+      # supplementary ruby code present in Gitlab::Database::PostgresHllBatchDistinctCounter
+      def estimate_batch_distinct_count(relation, column = nil, *rest)
+        raw_sql(relation, column, :distinct)
+      end
+
       private
 
       def raw_sql(relation, column, distinct = nil)

lib/gitlab/utils/usage_data.rb

@@ -38,6 +38,7 @@ module Gitlab
       extend self
 
       FALLBACK = -1
+      DISTRIBUTED_HLL_FALLBACK = -2
 
       def count(relation, column = nil, batch: true, batch_size: nil, start: nil, finish: nil)
         if batch
@@ -59,6 +60,17 @@ module Gitlab
         FALLBACK
       end
 
+      def estimate_batch_distinct_count(relation, column = nil, batch_size: nil, start: nil, finish: nil)
+        Gitlab::Database::PostgresHllBatchDistinctCounter.new(relation, column).estimate_distinct_count(batch_size: batch_size, start: start, finish: finish)
+      rescue ActiveRecord::StatementInvalid
+        FALLBACK
+      # catch all rescue should be removed as a part of feature flag rollout issue
+      # https://gitlab.com/gitlab-org/gitlab/-/issues/285485
+      rescue StandardError => error
+        Gitlab::ErrorTracking.track_and_raise_for_dev_exception(error)
+        DISTRIBUTED_HLL_FALLBACK
+      end
+
       def sum(relation, column, batch_size: nil, start: nil, finish: nil)
         Gitlab::Database::BatchCount.batch_sum(relation, column, batch_size: batch_size, start: start, finish: finish)
       rescue ActiveRecord::StatementInvalid

spec/lib/gitlab/database/postgres_hll_batch_distinct_counter_spec.rb

@@ -3,9 +3,9 @@
 require 'spec_helper'
 
 RSpec.describe Gitlab::Database::PostgresHllBatchDistinctCounter do
-  let_it_be(:error_rate) { 4.9 } # HyperLogLog is a probabilistic algorithm, which provides estimated data, with given error margin
+  let_it_be(:error_rate) { described_class::ERROR_RATE } # HyperLogLog is a probabilistic algorithm, which provides estimated data, with given error margin
   let_it_be(:fallback) { ::Gitlab::Database::BatchCounter::FALLBACK }
-  let_it_be(:small_batch_size) { calculate_batch_size(::Gitlab::Database::BatchCounter::MIN_REQUIRED_BATCH_SIZE) }
+  let_it_be(:small_batch_size) { calculate_batch_size(described_class::MIN_REQUIRED_BATCH_SIZE) }
   let(:model) { Issue }
   let(:column) { :author_id }
 
@@ -118,8 +118,8 @@ RSpec.describe Gitlab::Database::PostgresHllBatchDistinctCounter do
           expect(described_class.new(model, column).estimate_distinct_count(start: 1, finish: 0)).to eq(fallback)
         end
 
-        it 'returns fallback if loops more than allowed' do
-          large_finish = Gitlab::Database::PostgresHllBatchDistinctCounter::MAX_ALLOWED_LOOPS * default_batch_size + 1
+        it 'returns fallback if data volume exceeds upper limit' do
+          large_finish = Gitlab::Database::PostgresHllBatchDistinctCounter::MAX_DATA_VOLUME + 1
           expect(described_class.new(model, column).estimate_distinct_count(start: 1, finish: large_finish)).to eq(fallback)
         end
 

spec/lib/gitlab/utils/usage_data_spec.rb

@@ -37,6 +37,36 @@ RSpec.describe Gitlab::Utils::UsageData do
     end
   end
 
+  describe '#estimate_batch_distinct_count' do
+    let(:relation) { double(:relation) }
+
+    it 'delegates counting to counter class instance' do
+      expect_next_instance_of(Gitlab::Database::PostgresHllBatchDistinctCounter, relation, 'column') do |instance|
+        expect(instance).to receive(:estimate_distinct_count)
+                              .with(batch_size: nil, start: nil, finish: nil)
+                              .and_return(5)
+      end
+
+      expect(described_class.estimate_batch_distinct_count(relation, 'column')).to eq(5)
+    end
+
+    it 'returns default fallback value when counting fails due to database error' do
+      stub_const("Gitlab::Utils::UsageData::FALLBACK", 15)
+      allow(Gitlab::Database::PostgresHllBatchDistinctCounter).to receive(:new).and_raise(ActiveRecord::StatementInvalid.new(''))
+
+      expect(described_class.estimate_batch_distinct_count(relation)).to eq(15)
+    end
+
+    it 'logs error and returns DISTRIBUTED_HLL_FALLBACK value when counting raises any error', :aggregate_failures do
+      error = StandardError.new('')
+      stub_const("Gitlab::Utils::UsageData::DISTRIBUTED_HLL_FALLBACK", 15)
+      allow(Gitlab::Database::PostgresHllBatchDistinctCounter).to receive(:new).and_raise(error)
+
+      expect(Gitlab::ErrorTracking).to receive(:track_and_raise_for_dev_exception).with(error)
+      expect(described_class.estimate_batch_distinct_count(relation)).to eq(15)
+    end
+  end
+
   describe '#sum' do
     let(:relation) { double(:relation) }