func (bt *bucketTraverser) apply(f func(fv []*protocol.FieldValue, node *Node) error) error {
for !bt.reachedEnd() {
timestamp := &protocol.FieldValue{Int64Value: protocol.Int64(bt.current)}
defaultChildNode := &Node{states: make([]interface{}, bt.aggregators)}
err := bt.trie.TraverseLevel(bt.levels, func(v []*protocol.FieldValue, node *Node) error {
childNode := node.GetChildNode(timestamp)
if childNode == nil {
childNode = defaultChildNode
}
return f(append(v, timestamp), childNode)
})
if err != nil {
return err
}
bt.current += bt.step
}
return nil
}
// We have three types of queries:
// 1. time() without fill
// 2. time() with fill
// 3. no time()
//
// For (1) we flush as soon as a new bucket start, the prefix tree
// keeps track of the other group by columns without the time
// bucket. We reset the trie once the series is yielded. For (2), we
// keep track of all group by columns with time being the last level
// in the prefix tree. At the end of the query we step through [start
// time, end time] in self.duration steps and get the state from the
// prefix tree, using default values for groups without state in the
// prefix tree. For the last case we keep the groups in the prefix
// tree and on close() we loop through the groups and flush their
// values with a timestamp equal to now()
func (self *AggregatorEngine) aggregateValuesForSeries(series *protocol.Series) (bool, error) {
for _, aggregator := range self.aggregators {
if err := aggregator.InitializeFieldsMetadata(series); err != nil {
return false, err
}
}
seriesState := self.getSeriesState(series.GetName())
currentRange := seriesState.pointsRange
includeTimestampInGroup := self.duration != nil && self.isFillQuery
var group []*protocol.FieldValue
if !includeTimestampInGroup {
group = make([]*protocol.FieldValue, len(self.elems))
} else {
group = make([]*protocol.FieldValue, len(self.elems)+1)
}
for _, point := range series.Points {
currentRange.UpdateRange(point)
// this is a groupby with time() and no fill, flush as soon as we
// start a new bucket
if self.duration != nil && !self.isFillQuery {
timestamp := self.getTimestampFromPoint(point)
// this is the timestamp aggregator
if seriesState.started && seriesState.lastTimestamp != timestamp {
self.runAggregatesForTable(series.GetName())
}
seriesState.lastTimestamp = timestamp
seriesState.started = true
}
// get the group this point belongs to
for idx, elem := range self.elems {
// TODO: create an index from fieldname to index
// TODO: We shouldn't rely on GetValue() to do arithmetic
// operations. Instead we should cascade the arithmetic engine
// with the aggregator engine and possibly add another
// arithmetic engine to be able to do arithmetics on the
// resulting aggregated data.
value, err := GetValue(elem, series.Fields, point)
if err != nil {
return false, err
}
group[idx] = value
}
// if this is a fill() query, add the timestamp at the end
if includeTimestampInGroup {
timestamp := self.getTimestampFromPoint(point)
group[len(self.elems)] = &protocol.FieldValue{Int64Value: protocol.Int64(timestamp)}
}
// update the state of the given group
node := seriesState.trie.GetNode(group)
var err error
log4go.Debug("Aggregating for group %v", group)
for idx, aggregator := range self.aggregators {
log4go.Debug("Aggregating value for %T for group %v and state %v", aggregator, group, node.states[idx])
node.states[idx], err = aggregator.AggregatePoint(node.states[idx], point)
if err != nil {
return false, err
}
}
}
return true, nil
}
func (self *QueryEngine) runAggregatesForTable(table string) {
// TODO: if this is a fill query, step through [start,end] in duration
// steps and flush the groups for the given bucket
self.calculateSummariesForTable(table)
state := self.getSeriesState(table)
trie := state.trie
points := make([]*protocol.Point, 0, trie.CountLeafNodes())
f := func(group []*protocol.FieldValue, node *Node) error {
points = append(points, self.getValuesForGroup(table, group, node)...)
return nil
}
var err error
if self.duration != nil && self.fillWithZero {
timestampRange := state.pointsRange
// TODO: DRY this
if self.query.Ascending {
bucket := self.getTimestampBucket(uint64(timestampRange.startTime))
for bucket <= timestampRange.endTime {
timestamp := &protocol.FieldValue{Int64Value: protocol.Int64(bucket)}
defaultChildNode := &Node{states: make([]interface{}, len(self.aggregators))}
err = trie.TraverseLevel(len(self.elems), func(v []*protocol.FieldValue, node *Node) error {
childNode := node.GetChildNode(timestamp)
if childNode == nil {
childNode = defaultChildNode
}
return f(append(v, timestamp), childNode)
})
bucket += self.duration.Nanoseconds() / 1000
}
} else {
bucket := self.getTimestampBucket(uint64(timestampRange.endTime))
for {
timestamp := &protocol.FieldValue{Int64Value: protocol.Int64(bucket)}
defaultChildNode := &Node{states: make([]interface{}, len(self.aggregators))}
err = trie.TraverseLevel(len(self.elems), func(v []*protocol.FieldValue, node *Node) error {
childNode := node.GetChildNode(timestamp)
if childNode == nil {
childNode = defaultChildNode
}
return f(append(v, timestamp), childNode)
})
if bucket <= timestampRange.startTime {
break
}
bucket -= self.duration.Nanoseconds() / 1000
}
}
} else {
err = trie.Traverse(f)
}
if err != nil {
panic(err)
}
trie.Clear()
self.aggregateYield(&protocol.Series{
Name: &table,
Fields: self.fields,
Points: points,
})
}
// We have three types of queries:
// 1. time() without fill
// 2. time() with fill
// 3. no time()
//
// For (1) we flush as soon as a new bucket start, the prefix tree
// keeps track of the other group by columns without the time
// bucket. We reset the trie once the series is yielded. For (2), we
// keep track of all group by columns with time being the last level
// in the prefix tree. At the end of the query we step through [start
// time, end time] in self.duration steps and get the state from the
// prefix tree, using default values for groups without state in the
// prefix tree. For the last case we keep the groups in the prefix
// tree and on close() we loop through the groups and flush their
// values with a timestamp equal to now()
func (self *QueryEngine) aggregateValuesForSeries(series *protocol.Series) error {
for _, aggregator := range self.aggregators {
if err := aggregator.InitializeFieldsMetadata(series); err != nil {
return err
}
}
seriesState := self.getSeriesState(series.GetName())
currentRange := seriesState.pointsRange
includeTimestampInGroup := self.duration != nil && self.fillWithZero
var group []*protocol.FieldValue
if !includeTimestampInGroup {
group = make([]*protocol.FieldValue, len(self.elems))
} else {
group = make([]*protocol.FieldValue, len(self.elems)+1)
}
for _, point := range series.Points {
currentRange.UpdateRange(point)
// this is a groupby with time() and no fill, flush as soon as we
// start a new bucket
if self.duration != nil && !self.fillWithZero {
timestamp := self.getTimestampFromPoint(point)
// this is the timestamp aggregator
if seriesState.started && seriesState.lastTimestamp != timestamp {
self.runAggregatesForTable(series.GetName())
}
seriesState.lastTimestamp = timestamp
seriesState.started = true
}
// get the group this point belongs to
for idx, elem := range self.elems {
// TODO: create an index from fieldname to index
value, err := GetValue(elem, series.Fields, point)
if err != nil {
return err
}
group[idx] = value
}
// if this is a fill() query, add the timestamp at the end
if includeTimestampInGroup {
timestamp := self.getTimestampFromPoint(point)
group[len(self.elems)] = &protocol.FieldValue{Int64Value: protocol.Int64(timestamp)}
}
// update the state of the given group
node := seriesState.trie.GetNode(group)
var err error
for idx, aggregator := range self.aggregators {
node.states[idx], err = aggregator.AggregatePoint(node.states[idx], point)
if err != nil {
return err
}
}
}
return nil
}
func (self *TrieTestSuite) TestTrie(c *C) {
trie := NewTrie(2, 1)
firstValue := []*protocol.FieldValue{
{StringValue: protocol.String("some_value")},
{Int64Value: protocol.Int64(1)},
}
firstNode := trie.GetNode(firstValue)
c.Assert(firstNode, NotNil)
c.Assert(trie.GetNode(firstValue), DeepEquals, firstNode)
c.Assert(trie.CountLeafNodes(), Equals, 1)
secondValue := []*protocol.FieldValue{
{StringValue: protocol.String("some_value")},
{Int64Value: protocol.Int64(2)},
}
secondNode := trie.GetNode(secondValue)
c.Assert(secondNode, NotNil)
c.Assert(trie.GetNode(secondValue), DeepEquals, secondNode)
c.Assert(trie.CountLeafNodes(), Equals, 2)
thirdValue := []*protocol.FieldValue{
{StringValue: protocol.String("another_value")},
{Int64Value: protocol.Int64(1)},
}
thirdNode := trie.GetNode(thirdValue)
c.Assert(thirdNode, NotNil)
c.Assert(trie.GetNode(thirdValue), DeepEquals, thirdNode)
c.Assert(trie.CountLeafNodes(), Equals, 3)
nodes := 0
orderValues := [][]*protocol.FieldValue{thirdValue, firstValue, secondValue}
c.Assert(trie.Traverse(func(v []*protocol.FieldValue, _ *Node) error {
c.Assert(v, DeepEquals, orderValues[nodes])
nodes++
return nil
}), IsNil)
c.Assert(nodes, Equals, trie.CountLeafNodes())
// make sure TraverseLevel work as expected
ns := []*Node{}
c.Assert(trie.TraverseLevel(0, func(_ []*protocol.FieldValue, n *Node) error {
ns = append(ns, n)
return nil
}), IsNil)
c.Assert(ns, HasLen, 1) // should return the root node only
c.Assert(ns[0].value, IsNil)
ns = []*Node{}
c.Assert(trie.TraverseLevel(1, func(_ []*protocol.FieldValue, n *Node) error {
ns = append(ns, n)
return nil
}), IsNil)
c.Assert(ns, HasLen, 2) // should return the root node only
c.Assert(ns[0].value.GetStringValue(), Equals, "another_value")
c.Assert(ns[1].value.GetStringValue(), Equals, "some_value")
c.Assert(ns[0].GetChildNode(&protocol.FieldValue{Int64Value: protocol.Int64(1)}).isLeaf, Equals, true)
c.Assert(ns[0].GetChildNode(&protocol.FieldValue{Int64Value: protocol.Int64(2)}), IsNil)
c.Assert(ns[1].GetChildNode(&protocol.FieldValue{Int64Value: protocol.Int64(1)}).isLeaf, Equals, true)
c.Assert(ns[1].GetChildNode(&protocol.FieldValue{Int64Value: protocol.Int64(2)}).isLeaf, Equals, true)
}