block: Handle gracefully extra duplicated chunk produced by TSDB issue. #375

bwplotka · 2018-06-13T23:03:17Z

TSDB issue: prometheus-junkyard/tsdb#347

How we handle it?

segregate this issue in special stat entry in verifier.
auto-fix broken plan block before thanos compaction.
adding repair job to run offline batch repair for indivdual blocks.

NOTE: At this point we have no power of fixing the bug when someone uses local compaction ):

Fixes #354

PTAL @BenoitKnecht
Signed-off-by: Bartek Plotka [email protected]

BenoitKnecht · 2018-06-14T06:38:52Z

Looks good! The tests are failing, but it looks like it's due to the changes you made to some function signatures, nothing critical. I just have a few minor comments.

BenoitKnecht · 2018-06-14T06:39:35Z

cmd/thanos/bucket.go

+			for i, bid := range *verifyIDWhitelist {
+				id, err := ulid.Parse(bid)
+				if err != nil {
+					return errors.Wrap(err, "not valid ULID found in --id-whitelist flag")


"not valid" → "invalid"

BenoitKnecht · 2018-06-14T06:40:22Z

pkg/block/index.go

+
+	n := i.OutsideChunks - (i.CompleteOutsideChunks + i.Issue347OutsideChunks)
+	if n > 0 {
+		errMsg += fmt.Sprintf("found %d chunks non-completely outside the block time range.", n)


"non-completely" → "partially"

BenoitKnecht · 2018-06-14T06:41:10Z

pkg/block/index.go

+// GatherIndexIssueStats returns useful counters as well as outsider chunks (chunks outside of block time range) that
+// helps to assess index health.
+// It considers https://github.com/prometheus/tsdb/issues/347 as something that Thanos can handle.
+// See Stats.AcceptableOutsiders for details.


I guess you renamed Stats.AcceptableOutsiders to Stats.Issue347OutsideChunks.

BenoitKnecht · 2018-06-14T06:44:57Z

pkg/block/index.go

 				if c.MinTime > maxTime || c.MaxTime < minTime {
-					stats.CompleteOutsiders++
+					stats.CompleteOutsideChunks++
+				} else if c.MinTime == maxTime && c.MaxTime > maxTime {


I don't know if it can actually happen, but in theory a chunk with a single value could have c.MinTime == c.MaxTime, and if it's also at the very end of the block (c.MinTime == maxTime), it wouldn't be picked up by this condition.

I think you just need to test if c.MinTime == maxTime.

Can agree with it, but single sample chunk is very sad chunk (: But can happen probably.

BenoitKnecht · 2018-06-14T06:54:54Z

pkg/block/index.go

+		return true, nil
+	}
+	return false, nil
+}


I can see why you would collect separate stats for outside chunks and complete outside chunks, but when it comes to repairing blocks, they're really the same thing.

I mean if you have two intervals, block = [a, b) and chunk = [c, d], they don't overlap at all even if b == c. So you could have a single ignore function:

func IgnoreOutsideChunk(mint, maxt int64, _ *chunks.Meta, curr *chunks.Meta) (bool, error) { if curr.MinTime >= maxt || curr.MaxTime < mint { return true, nil } return false, nil }

If you don't merge those two functions, then I think you should change the condition in IgnoreIssue347OutsideChunk() to

if curr.MinTime == maxt { return true, nil }

Yea will do latter. Can agree that "complete" and "near" is kind of the same -> we can kill them just fine, but there is no upstream issue causing "complete". with issue347 we know what is that root cause and we know that it will not be fixed until it lands and land into Prometheus AND Prometheus go released. This will be quite a long time. With complete and duplicates even though they safe to repair they only occurred once Thanos had an issue, so the only explicit manual command should be fixing these, nothing automated.

BenoitKnecht · 2018-06-14T07:00:44Z

pkg/compact/compact.go

+			return false, errors.Wrapf(err, "gather index issues for block %s", pdir)
+		}
+
+		if err = stats.CriticalErr(); err == nil {


You probably want to use if err := ... here, since you're in a different scope. Same for a few more if conditions below.

also err != nil?

stefan-improbable · 2018-06-14T07:41:26Z

cmd/thanos/bucket.go

+				for _, bid := range whilelistIDs {
+					if id.Compare(bid) == 0 {
+						found = true
+					}


return true here and get rid of found variable

stefan-improbable · 2018-06-14T07:46:04Z

pkg/block/index.go

-	ExactSum        int
+	if i.OutOfOrderSeries > 0 {
+		errMsg += fmt.Sprintf(
+			"%d/%d series have an average of %.3f out-of-order chunks. "+


nit: the sentence should be "x, y, z" as opposed to "x. y. z". You should probably have errMsg []string and join it with ", " in the end.

stefan-improbable · 2018-06-14T08:01:04Z

pkg/block/index.go

-			repl = append(repl, c)
-			continue
+			if ignore {
+				break


would labelling outer loop with OUTER_LOOP: and continue OUTER_LOOP work here?

stefan-improbable · 2018-06-14T08:04:50Z

pkg/block/index.go

+		return false, nil
+	}
+
+	if curr.MinTime > prev.MaxTime {


Can you have curr.Mintime = prev.MaxTime?

no. Chunks never overlaps. Blocks overlaps with miliesecond always but means [t1, t2) (TSDB known inconsistency). But good question.

stefan-improbable · 2018-06-14T08:08:44Z

pkg/block/index.go

+		)
+
+		for _, ignoreChkFn := range ignoreChkFns {
+			ignore, err = ignoreChkFn(mint, maxt, last, &c)


some ignore check functions will fail if last is nil, is this ok?

yes - it is all handled.

stefan-improbable · 2018-06-14T08:10:07Z

pkg/compact/compact.go

@@ -532,35 +534,35 @@ func (cg *Group) areBlocksOverlapping(include *block.Meta, excludeDirs ...string
 	return nil
 }

-func (cg *Group) compact(ctx context.Context, dir string, comp tsdb.Compactor) (compID ulid.ULID, err error) {
+func (cg *Group) compact(ctx context.Context, dir string, comp tsdb.Compactor) (nothingToCompact bool, err error) {


you don't seem to be utilizing named returns

stefan-improbable · 2018-06-14T08:13:52Z

pkg/compact/compact.go

+			return false, errors.Wrapf(err, "gather index issues for block %s", pdir)
+		}
+
+		if err = stats.CriticalErr(); err == nil {


also err != nil?

stefan-improbable · 2018-06-14T08:16:37Z

pkg/compact/compact.go

+				return false, errors.Wrapf(err, "deleting old block %s failed. You need to delete this block manually", id)
+			}
+
+			return false, nil


shouldn't we continue with execution here?

I think we can remove this return as we would want to continue with the rest of the pdir := range plan

domgreen

Looks good overall, i think it would be good to make the Compact section of code more clear ... also changes as requested by @BenoitKnecht + @stefan-improbable

domgreen · 2018-06-14T08:43:53Z

cmd/thanos/bucket.go

+
+			idMatcher = func(id ulid.ULID) bool {
+				found := false
+				for _, bid := range whilelistIDs {


Thought - would it be better to do this using a map? map[ulid.ULID]struct{} then see if the ulid is present in the map? Not a big deal anyway.

you cannot simply do map[ulid.ULID]struct{} (ulid.ULID is a byte slice) but we cannot do map[string] - good point!

domgreen · 2018-06-14T08:56:02Z

cmd/thanos/compact.go

 					// We keep going through the outer loop until no group has any work left.
-					if id != (ulid.ULID{}) {
+					if !nothingToCompact {


Nit - !nothingToCompact reads weird would prefer !completed or !compactionCompleted

domgreen · 2018-06-14T09:00:47Z

pkg/compact/compact.go

@@ -440,23 +442,23 @@ func (cg *Group) Resolution() int64 {

 // Compact plans and runs a single compaction against the group. The compacted result
 // is uploaded into the bucket the blocks were retrieved from.
-func (cg *Group) Compact(ctx context.Context, dir string, comp tsdb.Compactor) (ulid.ULID, error) {
+func (cg *Group) Compact(ctx context.Context, dir string, comp tsdb.Compactor) (bool, error) {


Would be good to add to comment what the bool represents. Will returns true if you should continue compaction on subsequent groups.

domgreen · 2018-06-14T09:14:26Z

pkg/compact/compact.go

 	if err != nil {
 		cg.compactionFailures.Inc()
 	}
 	cg.compactions.Inc()

-	return id, err
+	return continueCompact, err


Abit confused here ... continueCompaction here turns into nothingToCompact in thanos/compact.go and in the inner compact below you return nothingToCompact as true if its done ... so this then gets assinged to continueCompact ... maybe im missing something but the vars are abit confusing.

Can we just change it all to continueCompact? Think this would read a lot better and help keep it simple.

domgreen · 2018-06-14T09:23:20Z

pkg/compact/compact.go

+		}
+
+		if err = stats.Issue347OutsideChunksErr(); err != nil {
+			level.Info(cg.logger).Log("msg", "stopping compaction and repairing invalid plan block", "err", err.Error(), "broken_block_id", pdir)


Nit - sentence

Actually looks like this meets convention so ignore 👍

domgreen · 2018-06-14T09:30:49Z

pkg/compact/compact.go

+				return false, errors.Wrapf(err, "deleting old block %s failed. You need to delete this block manually", id)
+			}
+
+			return false, nil


I think we can remove this return as we would want to continue with the rest of the pdir := range plan

bwplotka · 2018-06-14T10:49:55Z

Thanks all for the detailed review! Apologise for so many nits.. That always happens on 2 AM PRs ^^

Addressed all PTAL again @BenoitKnecht @stefan-improbable @domgreen

On green CI, before landing, I am going to test the fixed compaction flow on our testing cluster.

stefan-improbable · 2018-06-14T10:53:20Z

pkg/block/index.go

+// Issue347OutsideChunksErr returns error if stats indicates issue347 block issue, that is repaired explicitly before compaction (on plan block).
+func (i Stats) Issue347OutsideChunksErr() error {
+	if i.Issue347OutsideChunks > 0 {
+		return errors.Errorf("Found %d chunks outside the block time range introduced by https://github.com/prometheus/tsdb/issues/347", i.Issue347OutsideChunks)


nit: found with small f in the beginning due to style guide

stefan-improbable · 2018-06-14T10:53:46Z

pkg/block/index.go

+	var errMsg string
+
+	if err := i.CriticalErr(); err != nil {
+		errMsg += err.Error()


nit: use strings.join ", " similarly to CriticalErr

stefan-improbable · 2018-06-14T10:56:55Z

@Bplotka is it possible to unit test compaction and the block error detection/repair logic?

domgreen · 2018-06-14T11:07:16Z

cmd/thanos/compact.go

 					}
+					return errors.Wrap(err, "compaction")


Could we change this section so we do not indent the normal flow?

id, err := g.Compact(ctx, compactDir, comp) if err != nil { if compact.IsIssue347Error(err) { err = compact.RepairIssue347(ctx, logger, bkt, err) if err == nil { done = false continue } } return errors.Wrap(err, "compaction") } // If the returned ID has a zero value, the group had no blocks to be compacted. // We keep going through the outer loop until no group has any work left. if id != (ulid.ULID{}) { done = false }

but that will increase the indentation deepness. It's a tradeoff. What do you prefer?

I prefer less indentation on standard flow but it is not a blocker, just a preference.

bwplotka · 2018-06-14T11:35:37Z

@stefan-improbable yes, and we have these.

I added compaction upstream test though: prometheus-junkyard/tsdb#349 (we use vanilla TSDB compactor internally)

bwplotka · 2018-06-14T11:58:40Z

@stefan-improbable actually we have even e2e tests, sorry forgot abou these: https://github.com/improbable-eng/thanos/blob/master/pkg/compact/compact_e2e_test.go

bwplotka · 2018-06-14T15:26:51Z

PTAL again -> Added notion of ThanosMeta.Source

domgreen

LGTM

domgreen · 2018-06-15T09:15:46Z

I like the addition of ThanosMeta.Source 👍

bwplotka · 2018-06-15T11:46:55Z

I am going to leave it for a day on our beta cluster, but seems to be good to go 👍

stefan-improbable · 2018-06-15T11:40:42Z

pkg/block/block.go

-	IndexFilename = "index"
-	// ChunksDirname is the known dir name for chunks with compressed samples.
-	ChunksDirname = "chunks"
+	// Real upload source of the block.


Nit: Source is ..., comment should start with object name due to govet or go lint conventions.

stefan-improbable · 2018-06-15T11:44:47Z

pkg/block/index.go

-	ExactSum        int
+	if i.OutOfOrderSeries > 0 {
+		errMsg = append(errMsg, fmt.Sprintf(
+			"%d/%d series have an average of %.3f out-of-order chunks. "+


Nit: comment should be "%d/%d series have an ... chunks: %.3f of these ... time range)" (replace first period with colon and don't sentence in period so that Join(", ") works properly.

stefan-improbable · 2018-06-15T11:45:04Z

pkg/block/index.go

+	}
+
+	if len(errMsg) > 0 {
+		return errors.New(strings.Join(errMsg, ","))


Nit: replace "," with ", "

stefan-improbable · 2018-06-15T11:45:14Z

pkg/block/index.go

-	CompleteOutsiders int
+	n := i.OutsideChunks - (i.CompleteOutsideChunks + i.Issue347OutsideChunks)
+	if n > 0 {
+		errMsg = append(errMsg, fmt.Sprintf("found %d chunks non-completely outside the block time range.", n))


no period at the end of error message

stefan-improbable · 2018-06-15T11:45:22Z

pkg/block/index.go

+	}
+
+	if i.CompleteOutsideChunks > 0 {
+		errMsg = append(errMsg, fmt.Sprintf("found %d chunks completely outside the block time range.", i.CompleteOutsideChunks))


nit: no period at the end of error message

stefan-improbable · 2018-06-15T11:45:36Z

pkg/block/index.go

-		return errors.Errorf("No chunks are out of order, but found some outsider blocks. (Blocks that outside of block time range): %d. Complete: %d",
-			i.Outsiders, i.CompleteOutsiders)
+	if len(errMsg) > 0 {
+		return errors.New(strings.Join(errMsg, ","))


nit: replace "," with ", "

stefan-improbable · 2018-06-15T11:50:09Z

pkg/block/index.go

+	cb := crc32.Checksum(curr.Chunk.Bytes(), castagnoli)
+
+	if ca != cb {
+		return false, errors.Errorf("non-sequential chunks not equal: %x and %x", ca, cb)


Printing ca and cb is not useful (or is it)?. Printing (last.MinTime, last.MaxTime), (curr.MinTime, curr.MaxTime) is probably better. Or is there a way to indentify bad chunks so that users can debug them further?

I think this is enough. You have block ID so you can go through chunks and see on your own.

stefan-improbable · 2018-06-15T11:57:14Z

pkg/compact/compact.go

+		// - compactor created blocks
+		// NOTE: It is not safe to miss "old" block (even that it is newly created) in sync step. Compactor needs to aware of ALL old blocks.
+		// TODO(bplotka): https://github.com/improbable-eng/thanos/issues/377
+		if ulid.Now()-id.Time() < uint64(c.syncDelay/time.Millisecond) &&


Why don't you use time.Now().UTC()?

instead of what? We want to compare ulids which are not same as Time.Now() (they are not that precise)

I meant using time.Now().UTC() - id.Time(). The functions that "conveniently" call another functions are just introducing an unnecessary indirection.

I disagree here. time.Now().UTC() will not work.

we would endup doing ulid.Timestamp(time.Now().UTC()) which is rewritting function that is exactly for this. so would vote for leaving ulid.Now()

stefan-improbable · 2018-06-15T12:00:26Z

pkg/compact/compact.go

-		// avoid races when a block is only partially uploaded. This relates only to level 1 blocks.
-		// NOTE: It is not safe to miss compacted block in sync step. Compactor needs to aware of ALL old blocks.
-		if meta.Compaction.Level == 1 && ulid.Now()-id.Time() < uint64(c.syncDelay/time.Millisecond) {
+		// avoid races when a block is only partially uploaded. This relates to all blocks despite:


nit: despite is probably incorrect word here

TSDB issue: prometheus-junkyard/tsdb#347 How we handle it? - segregate this issue in special stat entry in verifier. - auto-fix broken plan block before thanos compaction. - adding repair job to run offline batch repair for indivdual blocks. NOTE: At this point we have no power of fixing the bug when someone uses local compaction ): Fixes #354 Signed-off-by: Bartek Plotka <[email protected]>

Use iterator for local Series call

bwplotka requested review from domgreen, fabxc and stefan-improbable June 13, 2018 23:03

BenoitKnecht reviewed Jun 14, 2018

View reviewed changes

stefan-improbable reviewed Jun 14, 2018

View reviewed changes

domgreen requested changes Jun 14, 2018

View reviewed changes

bwplotka force-pushed the issue354-outsider-handling branch 2 times, most recently from 90a2b8d to 90228b3 Compare June 14, 2018 10:47

stefan-improbable reviewed Jun 14, 2018

View reviewed changes

bwplotka force-pushed the issue354-outsider-handling branch from 90228b3 to 1dda32e Compare June 14, 2018 11:10

domgreen reviewed Jun 14, 2018

View reviewed changes

bwplotka force-pushed the issue354-outsider-handling branch from 1dda32e to 40fccef Compare June 14, 2018 15:23

bwplotka force-pushed the issue354-outsider-handling branch 2 times, most recently from f35be41 to 76c8191 Compare June 14, 2018 15:46

domgreen approved these changes Jun 15, 2018

View reviewed changes

bwplotka mentioned this pull request Jun 15, 2018

compact: Get rid of syncDelay and handle eventual consistency & partial upload differently. #377

Closed

bwplotka force-pushed the issue354-outsider-handling branch from 76c8191 to 52faa10 Compare June 15, 2018 11:32

stefan-improbable reviewed Jun 15, 2018

View reviewed changes

bwplotka force-pushed the issue354-outsider-handling branch from 52faa10 to 4ff9c61 Compare June 15, 2018 12:54

stefan-improbable approved these changes Jun 15, 2018

View reviewed changes

bwplotka merged commit 5b66b72 into master Jun 15, 2018

bwplotka deleted the issue354-outsider-handling branch December 11, 2018 13:27

fpetkovski added a commit to fpetkovski/thanos that referenced this pull request Oct 17, 2024

Merge pull request thanos-io#375 from Shopify/shopify-series-iterator

6b098d3

Use iterator for local Series call

block: Handle gracefully extra duplicated chunk produced by TSDB issue. #375

block: Handle gracefully extra duplicated chunk produced by TSDB issue. #375

Conversation

bwplotka commented Jun 13, 2018

BenoitKnecht commented Jun 14, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bwplotka Jun 14, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

domgreen left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bwplotka commented Jun 14, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

stefan-improbable commented Jun 14, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bwplotka commented Jun 14, 2018 • edited Loading

bwplotka commented Jun 14, 2018

bwplotka commented Jun 14, 2018

domgreen left a comment

Choose a reason for hiding this comment

domgreen commented Jun 15, 2018

bwplotka commented Jun 15, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bwplotka Jun 15, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bwplotka Jun 14, 2018 •

edited

Loading

bwplotka commented Jun 14, 2018 •

edited

Loading

bwplotka Jun 15, 2018 •

edited

Loading