compactor: Add offline deduplication

CHANGELOG.md

@@ -38,6 +38,7 @@ We use _breaking :warning:_ to mark changes that are not backward compatible (re
 - [#4211](https://github.com/thanos-io/thanos/pull/4211) Add TLS and basic authentication to Thanos APIs
 - [#4249](https://github.com/thanos-io/thanos/pull/4249) UI: add dark theme
 - [#3707](https://github.com/thanos-io/thanos/pull/3707) Tools: Added `--rewrite.to-relabel-config` to bucket rewrite tool to support series relabel from given blocks.
+- [#4239](https://github.com/thanos-io/thanos/pull/4239) Add penalty based deduplication mode for compactor.
 
 ### Fixed
 

cmd/thanos/compact.go

@@ -23,13 +23,16 @@ import (
 	"github.com/prometheus/client_golang/prometheus/promauto"
 	"github.com/prometheus/common/model"
 	"github.com/prometheus/common/route"
+	"github.com/prometheus/prometheus/storage"
 	"github.com/prometheus/prometheus/tsdb"
+
 	blocksAPI "github.com/thanos-io/thanos/pkg/api/blocks"
 	"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/dedup"
 	"github.com/thanos-io/thanos/pkg/extflag"
 	"github.com/thanos-io/thanos/pkg/extkingpin"
 	"github.com/thanos-io/thanos/pkg/extprom"
@@ -302,9 +305,20 @@ func runCompact(
 			cancel()
 		}
 	}()
+
+	var mergeFunc storage.VerticalChunkSeriesMergeFunc
+	switch conf.dedupFunc {
+	case compact.DedupAlgorithmPenalty:
+		mergeFunc = dedup.NewChunkSeriesMerger()
+
+		if len(conf.dedupReplicaLabels) == 0 {
+			return errors.New("penalty based deduplication needs at least one replica label specified")
+		}
+	}
+
 	// Instantiate the compactor with different time slices. Timestamps in TSDB
 	// are in milliseconds.
-	comp, err := tsdb.NewLeveledCompactor(ctx, reg, logger, levels, downsample.NewPool(), nil)
+	comp, err := tsdb.NewLeveledCompactor(ctx, reg, logger, levels, downsample.NewPool(), mergeFunc)
 	if err != nil {
 		return errors.Wrap(err, "create compactor")
 	}
@@ -564,6 +578,7 @@ type compactConfig struct {
 	maxBlockIndexSize                              units.Base2Bytes
 	hashFunc                                       string
 	enableVerticalCompaction                       bool
+	dedupFunc                                      string
 }
 
 func (cc *compactConfig) registerFlag(cmd extkingpin.FlagClause) {
@@ -623,14 +638,19 @@ func (cc *compactConfig) registerFlag(cmd extkingpin.FlagClause) {
 		Default("48h").SetValue(&cc.deleteDelay)
 
 	cmd.Flag("compact.enable-vertical-compaction", "Experimental. When set to true, compactor will allow overlaps and perform **irreversible** vertical compaction. See https://thanos.io/tip/components/compact.md/#vertical-compactions to read more."+
-		"Please note that this uses a NAIVE algorithm for merging (no smart replica deduplication, just chaining samples together)."+
+		"Please note that by default this uses a NAIVE algorithm for merging. If you need a different deduplication algorithm (e.g one that works well with Prometheus replicas), please set it via --deduplication.func."+
 		"NOTE: This flag is ignored and (enabled) when --deduplication.replica-label flag is set.").
 		Hidden().Default("false").BoolVar(&cc.enableVerticalCompaction)
 
+	cmd.Flag("deduplication.func", "Experimental. Deduplication algorithm for merging overlapping blocks. "+
+		"Possible values are: \"\", \"penalty\". If no value is specified, the default compact deduplication merger is used, which performs 1:1 deduplication for samples. "+
+		"When set to penalty, penalty based deduplication algorithm will be used. At least one replica label has to be set via --deduplication.replica-label flag.").
+		Default("").EnumVar(&cc.dedupFunc, compact.DedupAlgorithmPenalty, "")
+
 	cmd.Flag("deduplication.replica-label", "Label to treat as a replica indicator of blocks that can be deduplicated (repeated flag). This will merge multiple replica blocks into one. This process is irreversible."+
 		"Experimental. When it is set to true, compactor will ignore the given labels so that vertical compaction can merge the blocks."+
-		"Please note that this uses a NAIVE algorithm for merging (no smart replica deduplication, just chaining samples together)."+
-		"This works well for deduplication of blocks with **precisely the same samples** like produced by Receiver replication.").
+		"Please note that by default this uses a NAIVE algorithm for merging which works well for deduplication of blocks with **precisely the same samples** like produced by Receiver replication."+
+		"If you need a different deduplication algorithm (e.g one that works well with Prometheus replicas), please set it via --deduplication.func.").
 		Hidden().StringsVar(&cc.dedupReplicaLabels)
 
 	// TODO(bwplotka): This is short term fix for https://github.com/thanos-io/thanos/issues/1424, replace with vertical block sharding https://github.com/thanos-io/thanos/pull/3390.

docs/components/compact.md

@@ -89,8 +89,8 @@ which compacts overlapping blocks into single one. This is mainly used for **bac
 
 In Thanos, it works similarly, but on bigger scale and using external labels for grouping as explained in [Compaction section](#compaction).
 
-In both systems, series with the same labels are merged together. Merging samples is **naive**. It works by deduplicating samples within
-exactly the same timestamps. Otherwise samples are added in sorted by time order.
+In both systems, series with the same labels are merged together. In prometheus, merging samples is **naive**. It works by deduplicating samples within
+exactly the same timestamps. Otherwise samples are added in sorted by time order. Thanos also support a new penalty based samples merger and it is explained in [Deduplication](#Vertical Compaction Use Cases).
 
 > **NOTE:** Both Prometheus and Thanos default behaviour is to fail compaction if any overlapping blocks are spotted. (For Thanos, within the same external labels).
 
@@ -101,12 +101,12 @@ There can be few valid use cases for vertical compaction:
 * Races between multiple compactions, for example multiple compactors or between compactor and Prometheus compactions. While this will have extra
 computation overhead for Compactor it's safe to enable vertical compaction for this case.
 * Backfilling. If you want to add blocks of data to any stream where there is existing data already there for the time range, you will need enabled vertical compaction.
-* Offline deduplication of series. It's very common to have the same data replicated into multiple streams. We can distinguish two common series duplications, `one-to-one` and `realistic`:
-  * `one-to-one` duplication is when same series (series with the same labels from different blocks) for the same range have **exactly** the same samples: Same values and timestamps.
+* Offline deduplication of series. It's very common to have the same data replicated into multiple streams. We can distinguish two common series deduplications, `one-to-one` and `penalty`:
+  * `one-to-one` deduplication is when same series (series with the same labels from different blocks) for the same range have **exactly** the same samples: Same values and timestamps.
 This is very common while using [Receivers](../components/receive.md) with replication greater than 1 as receiver replication copies exactly the same timestamps and values to different receive instances.
-  * `realistic` duplication is when same series data is **logically duplicated**. For example, it comes from the same application, but scraped by two different Prometheus-es. Ideally
+  * `penalty` deduplication is when same series data is **logically duplicated**. For example, it comes from the same application, but scraped by two different Prometheus-es. Ideally
 this requires more complex deduplication algorithms. For example one that is used to [deduplicate on the fly on the Querier](query.md#run-time-deduplication-of-ha-groups). This is common
-case when Prometheus HA replicas are used. [Offline deduplication for this is in progress](https://github.com/thanos-io/thanos/issues/1014).
+case when Prometheus HA replicas are used. You can enable this deduplication via `--deduplication.func=penalty` flag.
 
 #### Vertical Compaction Risks
 
@@ -143,6 +143,8 @@ external_labels: {cluster="us1", receive="true", environment="staging"}
 
 On next compaction multiple streams' blocks will be compacted into one.
 
+If you need a different deduplication algorithm, use `deduplication.func` flag. The default value is the original `one-to-one` deduplication.
+
 ## Enforcing Retention of Data
 
 By default, there is NO retention set for object storage data. This means that you store data for unlimited time, which is a valid and recommended way of running Thanos.
@@ -349,6 +351,15 @@ Flags:
                                 consistency-delay and 48h0m0s will be removed.
       --data-dir="./data"       Data directory in which to cache blocks and
                                 process compactions.
+      --deduplication.func=     Experimental. Deduplication algorithm for
+                                merging overlapping blocks. Possible values are:
+                                "", "penalty". If no value is specified, the
+                                default compact deduplication merger is used,
+                                which performs 1:1 deduplication for samples.
+                                When set to penalty, penalty based deduplication
+                                algorithm will be used. At least one replica
+                                label has to be set via
+                                --deduplication.replica-label flag.
       --delete-delay=48h        Time before a block marked for deletion is
                                 deleted from bucket. If delete-delay is non
                                 zero, blocks will be marked for deletion and

go.mod

@@ -53,6 +53,7 @@ require (
 	github.com/prometheus/common v0.23.0
 	github.com/prometheus/exporter-toolkit v0.5.1
 	github.com/prometheus/prometheus v1.8.2-0.20210519120135-d95b0972505f
+	github.com/stretchr/testify v1.7.0
 	github.com/uber/jaeger-client-go v2.28.0+incompatible
 	github.com/uber/jaeger-lib v2.4.1+incompatible
 	github.com/weaveworks/common v0.0.0-20210419092856-009d1eebd624

pkg/compact/compact.go

@@ -41,6 +41,12 @@ const (
 	ResolutionLevel1h  = ResolutionLevel(downsample.ResLevel2)
 )
 
+const (
+	// DedupAlgorithmPenalty is the penalty based compactor series merge algorithm.
+	// This is the same as the online deduplication of querier except counter reset handling.
+	DedupAlgorithmPenalty = "penalty"
+)
+
 // Syncer synchronizes block metas from a bucket into a local directory.
 // It sorts them into compaction groups based on equal label sets.
 type Syncer struct {

pkg/compact/compact_e2e_test.go

@@ -22,9 +22,12 @@ import (
 	"github.com/prometheus/client_golang/prometheus/promauto"
 	promtest "github.com/prometheus/client_golang/prometheus/testutil"
 	"github.com/prometheus/prometheus/pkg/labels"
+	"github.com/prometheus/prometheus/storage"
 	"github.com/prometheus/prometheus/tsdb"
+
 	"github.com/thanos-io/thanos/pkg/block"
 	"github.com/thanos-io/thanos/pkg/block/metadata"
+	"github.com/thanos-io/thanos/pkg/dedup"
 	"github.com/thanos-io/thanos/pkg/objstore"
 	"github.com/thanos-io/thanos/pkg/objstore/objtesting"
 	"github.com/thanos-io/thanos/pkg/testutil"
@@ -167,7 +170,16 @@ func MetricCount(c prometheus.Collector) int {
 	return mCount
 }
 
-func TestGroup_Compact_e2e(t *testing.T) {
+func TestGroupCompactE2E(t *testing.T) {
+	testGroupCompactE2e(t, nil)
+}
+
+// Penalty based merger should get the same result as the blocks don't have overlap.
+func TestGroupCompactPenaltyDedupE2E(t *testing.T) {
+	testGroupCompactE2e(t, dedup.NewChunkSeriesMerger())
+}
+
+func testGroupCompactE2e(t *testing.T, mergeFunc storage.VerticalChunkSeriesMergeFunc) {
 	objtesting.ForeachStore(t, func(t *testing.T, bkt objstore.Bucket) {
 		ctx, cancel := context.WithTimeout(context.Background(), 120*time.Second)
 		defer cancel()
@@ -194,7 +206,7 @@ func TestGroup_Compact_e2e(t *testing.T) {
 		sy, err := NewMetaSyncer(nil, nil, bkt, metaFetcher, duplicateBlocksFilter, ignoreDeletionMarkFilter, blocksMarkedForDeletion, garbageCollectedBlocks, 5)
 		testutil.Ok(t, err)
 
-		comp, err := tsdb.NewLeveledCompactor(ctx, reg, logger, []int64{1000, 3000}, nil, nil)
+		comp, err := tsdb.NewLeveledCompactor(ctx, reg, logger, []int64{1000, 3000}, nil, mergeFunc)
 		testutil.Ok(t, err)
 
 		planner := NewTSDBBasedPlanner(logger, []int64{1000, 3000})