compact+store+tools: Filter for duplicate block data only within a co…

…mpaction group and in parallel (thanos-io#4963) * Move group key function to the metadata package. Deduplicate within compaction groups and in parallel. Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * cleanup after rebase Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * formatting Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * Reduce mutex operations in DeduplicateFilter Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * fix tests after last update from main Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * formatting Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * Move fetcher modifiers before filters so we can compute correct compactions groups during deduplication Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * Merge the notion of filters and modifiers so we can strip replica labels before deduplication. Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * formatting Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * fix tests Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * remove duplicate filter list entry Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * FIx new calls to NewStoreGW Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * add CHANGELOG Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * Apply suggestions from code review Co-authored-by: Bartlomiej Plotka <bwplotka@gmail.com> Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> * comment suggested in review Signed-off-by: Nathan Berkley <nberkley@tripadvisor.com> Co-authored-by: Bartlomiej Plotka <bwplotka@gmail.com> Signed-off-by: Nicholaswang <wzhever@gmail.com>
Nicholaswang · Mar 6, 2022 · 129e507 · 129e507
1 parent 5ae2208
commit 129e507
Show file tree

Hide file tree

Showing 25 changed files with 171 additions and 239 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -12,6 +12,7 @@ We use *breaking :warning:* to mark changes that are not backward compatible (re
 
 ### Added
 
+- [#4963](https://github.com/thanos-io/thanos/pull/4963) Compactor, Store, Tools: Loading block metadata now only filters out duplicates within a source (or compaction group if replica labels are configured), and does so in parallel over sources.
 - [#5089](https://github.com/thanos-io/thanos/pull/5089) S3: Create an empty map in the case SSE-KMS is used and no KMSEncryptionContext is passed.
 - [#4970](https://github.com/thanos-io/thanos/pull/4970) Added a new flag `exclude-delete` to `tools bucket ls`, which excludes blocks marked for deletion.
 - [#4903](https://github.com/thanos-io/thanos/pull/4903) Compactor: Added tracing support for compaction.

diff --git a/cmd/thanos/compact.go b/cmd/thanos/compact.go
@@ -228,7 +228,7 @@ func runCompact(
 	// The delay of deleteDelay/2 is added to ensure we fetch blocks that are meant to be deleted but do not have a replacement yet.
 	// This is to make sure compactor will not accidentally perform compactions with gap instead.
 	ignoreDeletionMarkFilter := block.NewIgnoreDeletionMarkFilter(logger, bkt, deleteDelay/2, conf.blockMetaFetchConcurrency)
-	duplicateBlocksFilter := block.NewDeduplicateFilter()
+	duplicateBlocksFilter := block.NewDeduplicateFilter(conf.blockMetaFetchConcurrency)
 	noCompactMarkerFilter := compact.NewGatherNoCompactionMarkFilter(logger, bkt, conf.blockMetaFetchConcurrency)
 	labelShardedMetaFilter := block.NewLabelShardedMetaFilter(relabelConfig)
 	consistencyDelayMetaFilter := block.NewConsistencyDelayMetaFilter(logger, conf.consistencyDelay, extprom.WrapRegistererWithPrefix("thanos_", reg))
@@ -271,9 +271,10 @@ func runCompact(
 				labelShardedMetaFilter,
 				consistencyDelayMetaFilter,
 				ignoreDeletionMarkFilter,
+				block.NewReplicaLabelRemover(logger, conf.dedupReplicaLabels),
 				duplicateBlocksFilter,
 				noCompactMarkerFilter,
-			}, []block.MetadataModifier{block.NewReplicaLabelRemover(logger, conf.dedupReplicaLabels)},
+			},
 		)
 		cf.UpdateOnChange(func(blocks []metadata.Meta, err error) {
 			compactorView.Set(blocks, err)
@@ -440,7 +441,7 @@ func runCompact(
 			}
 
 			for _, meta := range sy.Metas() {
-				groupKey := compact.DefaultGroupKey(meta.Thanos)
+				groupKey := meta.Thanos.GroupKey()
 				downsampleMetrics.downsamples.WithLabelValues(groupKey)
 				downsampleMetrics.downsampleFailures.WithLabelValues(groupKey)
 			}
@@ -532,7 +533,7 @@ func runCompact(
 
 		// Separate fetcher for global view.
 		// TODO(bwplotka): Allow Bucket UI to visualize the state of the block as well.
-		f := baseMetaFetcher.NewMetaFetcher(extprom.WrapRegistererWithPrefix("thanos_bucket_ui", reg), nil, nil, "component", "globalBucketUI")
+		f := baseMetaFetcher.NewMetaFetcher(extprom.WrapRegistererWithPrefix("thanos_bucket_ui", reg), nil, "component", "globalBucketUI")
 		f.UpdateOnChange(func(blocks []metadata.Meta, err error) {
 			global.Set(blocks, err)
 			api.SetGlobal(blocks, err)

diff --git a/cmd/thanos/downsample.go b/cmd/thanos/downsample.go
@@ -24,7 +24,6 @@ import (
 
 	"github.com/thanos-io/thanos/pkg/block"
 	"github.com/thanos-io/thanos/pkg/block/metadata"
-	"github.com/thanos-io/thanos/pkg/compact"
 	"github.com/thanos-io/thanos/pkg/compact/downsample"
 	"github.com/thanos-io/thanos/pkg/component"
 	"github.com/thanos-io/thanos/pkg/errutil"
@@ -86,8 +85,8 @@ func RunDownsample(
 	}
 
 	metaFetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix("thanos_", reg), []block.MetadataFilter{
-		block.NewDeduplicateFilter(),
-	}, nil)
+		block.NewDeduplicateFilter(block.FetcherConcurrency),
+	})
 	if err != nil {
 		return errors.Wrap(err, "create meta fetcher")
 	}
@@ -121,7 +120,7 @@ func RunDownsample(
 			}
 
 			for _, meta := range metas {
-				groupKey := compact.DefaultGroupKey(meta.Thanos)
+				groupKey := meta.Thanos.GroupKey()
 				metrics.downsamples.WithLabelValues(groupKey)
 				metrics.downsampleFailures.WithLabelValues(groupKey)
 			}
@@ -252,10 +251,11 @@ func downsampleBucket(
 					errMsg = "downsampling to 60 min"
 				}
 				if err := processDownsampling(workerCtx, logger, bkt, m, dir, resolution, hashFunc, metrics); err != nil {
-					metrics.downsampleFailures.WithLabelValues(compact.DefaultGroupKey(m.Thanos)).Inc()
+					metrics.downsampleFailures.WithLabelValues(m.Thanos.GroupKey()).Inc()
 					errCh <- errors.Wrap(err, errMsg)
+
 				}
-				metrics.downsamples.WithLabelValues(compact.DefaultGroupKey(m.Thanos)).Inc()
+				metrics.downsamples.WithLabelValues(m.Thanos.GroupKey()).Inc()
 			}
 		}()
 	}
@@ -377,7 +377,7 @@ func processDownsampling(
 	downsampleDuration := time.Since(begin)
 	level.Info(logger).Log("msg", "downsampled block",
 		"from", m.ULID, "to", id, "duration", downsampleDuration, "duration_ms", downsampleDuration.Milliseconds())
-	metrics.downsampleDuration.WithLabelValues(compact.DefaultGroupKey(m.Thanos)).Observe(downsampleDuration.Seconds())
+	metrics.downsampleDuration.WithLabelValues(m.Thanos.GroupKey()).Observe(downsampleDuration.Seconds())
 
 	if err := block.VerifyIndex(logger, filepath.Join(resdir, block.IndexFilename), m.MinTime, m.MaxTime); err != nil {
 		return errors.Wrap(err, "output block index not valid")

diff --git a/cmd/thanos/main_test.go b/cmd/thanos/main_test.go
@@ -22,7 +22,6 @@ import (
 
 	"github.com/thanos-io/thanos/pkg/block"
 	"github.com/thanos-io/thanos/pkg/block/metadata"
-	"github.com/thanos-io/thanos/pkg/compact"
 	"github.com/thanos-io/thanos/pkg/compact/downsample"
 	"github.com/thanos-io/thanos/pkg/objstore"
 	"github.com/thanos-io/thanos/pkg/testutil"
@@ -154,8 +153,8 @@ func TestRegression4960_Deadlock(t *testing.T) {
 	testutil.Ok(t, err)
 
 	metrics := newDownsampleMetrics(prometheus.NewRegistry())
-	testutil.Equals(t, 0.0, promtest.ToFloat64(metrics.downsamples.WithLabelValues(compact.DefaultGroupKey(meta.Thanos))))
-	metaFetcher, err := block.NewMetaFetcher(nil, block.FetcherConcurrency, bkt, "", nil, nil, nil)
+	testutil.Equals(t, 0.0, promtest.ToFloat64(metrics.downsamples.WithLabelValues(meta.Thanos.GroupKey())))
+	metaFetcher, err := block.NewMetaFetcher(nil, block.FetcherConcurrency, bkt, "", nil, nil)
 	testutil.Ok(t, err)
 
 	metas, _, err := metaFetcher.Fetch(ctx)
@@ -194,14 +193,14 @@ func TestCleanupDownsampleCacheFolder(t *testing.T) {
 	testutil.Ok(t, err)
 
 	metrics := newDownsampleMetrics(prometheus.NewRegistry())
-	testutil.Equals(t, 0.0, promtest.ToFloat64(metrics.downsamples.WithLabelValues(compact.DefaultGroupKey(meta.Thanos))))
-	metaFetcher, err := block.NewMetaFetcher(nil, block.FetcherConcurrency, bkt, "", nil, nil, nil)
+	testutil.Equals(t, 0.0, promtest.ToFloat64(metrics.downsamples.WithLabelValues(meta.Thanos.GroupKey())))
+	metaFetcher, err := block.NewMetaFetcher(nil, block.FetcherConcurrency, bkt, "", nil, nil)
 	testutil.Ok(t, err)
 
 	metas, _, err := metaFetcher.Fetch(ctx)
 	testutil.Ok(t, err)
 	testutil.Ok(t, downsampleBucket(ctx, logger, metrics, bkt, metas, dir, 1, metadata.NoneFunc))
-	testutil.Equals(t, 1.0, promtest.ToFloat64(metrics.downsamples.WithLabelValues(compact.DefaultGroupKey(meta.Thanos))))
+	testutil.Equals(t, 1.0, promtest.ToFloat64(metrics.downsamples.WithLabelValues(meta.Thanos.GroupKey())))
 
 	_, err = os.Stat(dir)
 	testutil.Assert(t, os.IsNotExist(err), "index cache dir should not exist at the end of execution")

diff --git a/cmd/thanos/store.go b/cmd/thanos/store.go
@@ -304,8 +304,8 @@ func runStore(
 			block.NewLabelShardedMetaFilter(relabelConfig),
 			block.NewConsistencyDelayMetaFilter(logger, time.Duration(conf.consistencyDelay), extprom.WrapRegistererWithPrefix("thanos_", reg)),
 			ignoreDeletionMarkFilter,
-			block.NewDeduplicateFilter(),
-		}, nil)
+			block.NewDeduplicateFilter(conf.blockMetaFetchConcurrency),
+		})
 	if err != nil {
 		return errors.Wrap(err, "meta fetcher")
 	}

diff --git a/cmd/thanos/tools_bucket.go b/cmd/thanos/tools_bucket.go
@@ -330,7 +330,7 @@ func registerBucketVerify(app extkingpin.AppClause, objStoreConfig *extflag.Path
 			return err
 		}
 
-		fetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix(extpromPrefix, reg), nil, nil)
+		fetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix(extpromPrefix, reg), nil)
 		if err != nil {
 			return err
 		}
@@ -386,7 +386,7 @@ func registerBucketLs(app extkingpin.AppClause, objStoreConfig *extflag.PathOrCo
 			ignoreDeletionMarkFilter := block.NewIgnoreDeletionMarkFilter(logger, bkt, 0, block.FetcherConcurrency)
 			filters = append(filters, ignoreDeletionMarkFilter)
 		}
-		fetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix(extpromPrefix, reg), filters, nil)
+		fetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix(extpromPrefix, reg), filters)
 		if err != nil {
 			return err
 		}
@@ -486,7 +486,7 @@ func registerBucketInspect(app extkingpin.AppClause, objStoreConfig *extflag.Pat
 			return err
 		}
 
-		fetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix(extpromPrefix, reg), nil, nil)
+		fetcher, err := block.NewMetaFetcher(logger, block.FetcherConcurrency, bkt, "", extprom.WrapRegistererWithPrefix(extpromPrefix, reg), nil)
 		if err != nil {
 			return err
 		}
@@ -632,8 +632,8 @@ func registerBucketWeb(app extkingpin.AppClause, objStoreConfig *extflag.PathOrC
 			[]block.MetadataFilter{
 				block.NewTimePartitionMetaFilter(filterConf.MinTime, filterConf.MaxTime),
 				block.NewLabelShardedMetaFilter(relabelConfig),
-				block.NewDeduplicateFilter(),
-			}, nil)
+				block.NewDeduplicateFilter(block.FetcherConcurrency),
+			})
 		if err != nil {
 			return err
 		}
@@ -786,7 +786,7 @@ func registerBucketCleanup(app extkingpin.AppClause, objStoreConfig *extflag.Pat
 		// The delay of deleteDelay/2 is added to ensure we fetch blocks that are meant to be deleted but do not have a replacement yet.
 		// This is to make sure compactor will not accidentally perform compactions with gap instead.
 		ignoreDeletionMarkFilter := block.NewIgnoreDeletionMarkFilter(logger, bkt, tbc.deleteDelay/2, block.FetcherConcurrency)
-		duplicateBlocksFilter := block.NewDeduplicateFilter()
+		duplicateBlocksFilter := block.NewDeduplicateFilter(block.FetcherConcurrency)
 		blocksCleaner := compact.NewBlocksCleaner(logger, bkt, ignoreDeletionMarkFilter, tbc.deleteDelay, stubCounter, stubCounter)
 
 		ctx := context.Background()
@@ -803,7 +803,7 @@ func registerBucketCleanup(app extkingpin.AppClause, objStoreConfig *extflag.Pat
 					block.NewConsistencyDelayMetaFilter(logger, tbc.consistencyDelay, extprom.WrapRegistererWithPrefix(extpromPrefix, reg)),
 					ignoreDeletionMarkFilter,
 					duplicateBlocksFilter,
-				}, []block.MetadataModifier{block.NewReplicaLabelRemover(logger, make([]string, 0))},
+				},
 			)
 			sy, err = compact.NewMetaSyncer(
 				logger,
@@ -1312,7 +1312,7 @@ func registerBucketRetention(app extkingpin.AppClause, objStoreConfig *extflag.P
 		// The delay of deleteDelay/2 is added to ensure we fetch blocks that are meant to be deleted but do not have a replacement yet.
 		// This is to make sure compactor will not accidentally perform compactions with gap instead.
 		ignoreDeletionMarkFilter := block.NewIgnoreDeletionMarkFilter(logger, bkt, tbc.deleteDelay/2, block.FetcherConcurrency)
-		duplicateBlocksFilter := block.NewDeduplicateFilter()
+		duplicateBlocksFilter := block.NewDeduplicateFilter(block.FetcherConcurrency)
 		stubCounter := promauto.With(nil).NewCounter(prometheus.CounterOpts{})
 
 		var sy *compact.Syncer
@@ -1327,7 +1327,7 @@ func registerBucketRetention(app extkingpin.AppClause, objStoreConfig *extflag.P
 					block.NewConsistencyDelayMetaFilter(logger, tbc.consistencyDelay, extprom.WrapRegistererWithPrefix(extpromPrefix, reg)),
 					duplicateBlocksFilter,
 					ignoreDeletionMarkFilter,
-				}, []block.MetadataModifier{block.NewReplicaLabelRemover(logger, make([]string, 0))},
+				},
 			)
 			sy, err = compact.NewMetaSyncer(
 				logger,