make sure access to chunk is always serialized

Author: Dieterbe

mdata/aggmetric.go

@@ -339,6 +339,8 @@ func (a *AggMetric) addAggregators(ts uint32, val float64) {
 	}
 }
 
+// pushToCache adds the chunk into the cache if it is hot
+// assumes lock held by caller
 func (a *AggMetric) pushToCache(c *chunk.Chunk) {
 	if a.cachePusher == nil {
 		return
@@ -348,14 +350,14 @@ func (a *AggMetric) pushToCache(c *chunk.Chunk) {
 	if a.Key.Archive != 0 {
 		intervalHint = a.Key.Archive.Span()
 	}
-	go a.cachePusher.AddIfHot(
-		a.Key,
-		0,
-		chunk.NewBareIterGen(c.T0, intervalHint, c.Encode(a.ChunkSpan)),
-	)
+
+	itergen := chunk.NewBareIterGen(c.Series.T0, intervalHint, c.Encode(a.ChunkSpan))
+	go a.cachePusher.AddIfHot(a.Key, 0, itergen)
 }
 
 // write a chunk to persistent storage. This should only be called while holding a.Lock()
+// never persist a chunk that may receive further updates!
+// (because the stores will read out chunk data on the unlocked chunk)
 func (a *AggMetric) persist(pos int) {
 	chunk := a.Chunks[pos]
 	pre := time.Now()

mdata/chunk/chunk.go

@@ -9,6 +9,13 @@ import (
 )
 
 // Chunk is a chunk of data. not concurrency safe.
+// last check that the methods are being called safely by Dieter on 20/11/2018
+// checked: String, Push, Finish, Encode
+// for the most part, confirming serialized access is easy by tracking all callers.
+// The main exception is the ChunkWriteRequest mechanism. any CWR created is processed
+// "later", and chunk data (the series, etc) will be accessed via Encode() and reads of
+// some of the members. But it can be proven
+// that we only call NewChunkWriteRequest() on chunks that are no longer being modified.
 type Chunk struct {
 	Series    tsz.SeriesLong
 	NumPoints uint32
@@ -48,6 +55,7 @@ func (c *Chunk) Finish() {
 // Encode encodes the chunk
 // note: chunks don't know their own span, the caller/owner manages that,
 // so for formats that encode it, it needs to be passed in.
+// the returned value contains no references to the chunk. data is copied.
 func (c *Chunk) Encode(span uint32) []byte {
 	return encode(span, FormatGoTszLongWithSpan, c.Series.Bytes())
 }

mdata/chunk/encode.go

@@ -7,6 +7,7 @@ import (
 )
 
 // encode is a helper function to encode a chunk of data into various formats
+// input data is copied
 func encode(span uint32, format Format, data []byte) []byte {
 	switch format {
 	case FormatStandardGoTszWithSpan, FormatGoTszLongWithSpan: