iterator.go

package query

import (
	"context"
	"encoding/binary"
	"errors"
	"fmt"
	"io"
	"regexp"
	"time"

	"github.com/gogo/protobuf/proto"
	"github.com/influxdata/influxdb/v2"
	internal "github.com/influxdata/influxdb/v2/influxql/query/internal"
	"github.com/influxdata/influxql"
)

// ErrUnknownCall is returned when operating on an unknown function call.
var ErrUnknownCall = errors.New("unknown call")

const (
	// secToNs is the number of nanoseconds in a second.
	secToNs = int64(time.Second)
)

// Iterator represents a generic interface for all Iterators.
// Most iterator operations are done on the typed sub-interfaces.
type Iterator interface {
	Stats() IteratorStats
	Close() error
}

// Iterators represents a list of iterators.
type Iterators []Iterator

// Stats returns the aggregation of all iterator stats.
func (a Iterators) Stats() IteratorStats {
	var stats IteratorStats
	for _, itr := range a {
		stats.Add(itr.Stats())
	}
	return stats
}

// Close closes all iterators.
func (a Iterators) Close() error {
	for _, itr := range a {
		itr.Close()
	}
	return nil
}

// filterNonNil returns a slice of iterators that removes all nil iterators.
func (a Iterators) filterNonNil() []Iterator {
	other := make([]Iterator, 0, len(a))
	for _, itr := range a {
		if itr == nil {
			continue
		}
		other = append(other, itr)
	}
	return other
}

// dataType determines what slice type this set of iterators should be.
// An iterator type is chosen by looking at the first element in the slice
// and then returning the data type for that iterator.
func (a Iterators) dataType() influxql.DataType {
	if len(a) == 0 {
		return influxql.Unknown
	}

	switch a[0].(type) {
	case FloatIterator:
		return influxql.Float
	case IntegerIterator:
		return influxql.Integer
	case UnsignedIterator:
		return influxql.Unsigned
	case StringIterator:
		return influxql.String
	case BooleanIterator:
		return influxql.Boolean
	default:
		return influxql.Unknown
	}
}

// coerce forces an array of iterators to be a single type.
// Iterators that are not of the same type as the first element in the slice
// will be closed and dropped.
func (a Iterators) coerce() interface{} {
	typ := a.dataType()
	switch typ {
	case influxql.Float:
		return newFloatIterators(a)
	case influxql.Integer:
		return newIntegerIterators(a)
	case influxql.Unsigned:
		return newUnsignedIterators(a)
	case influxql.String:
		return newStringIterators(a)
	case influxql.Boolean:
		return newBooleanIterators(a)
	}
	return a
}

// Merge combines all iterators into a single iterator.
// A sorted merge iterator or a merge iterator can be used based on opt.
func (a Iterators) Merge(opt IteratorOptions) (Iterator, error) {
	// Check if this is a call expression.
	call, ok := opt.Expr.(*influxql.Call)

	// Merge into a single iterator.
	if !ok && opt.MergeSorted() {
		itr := NewSortedMergeIterator(a, opt)
		if itr != nil && opt.InterruptCh != nil {
			itr = NewInterruptIterator(itr, opt.InterruptCh)
		}
		return itr, nil
	}

	// We do not need an ordered output so use a merge iterator.
	itr := NewMergeIterator(a, opt)
	if itr == nil {
		return nil, nil
	}

	if opt.InterruptCh != nil {
		itr = NewInterruptIterator(itr, opt.InterruptCh)
	}

	if !ok {
		// This is not a call expression so do not use a call iterator.
		return itr, nil
	}

	// When merging the count() function, use sum() to sum the counted points.
	if call.Name == "count" {
		opt.Expr = &influxql.Call{
			Name: "sum",
			Args: call.Args,
		}
	}
	return NewCallIterator(itr, opt)
}

// NewMergeIterator returns an iterator to merge itrs into one.
// Inputs must either be merge iterators or only contain a single name/tag in
// sorted order. The iterator will output all points by window, name/tag, then
// time. This iterator is useful when you need all of the points for an
// interval.
func NewMergeIterator(inputs []Iterator, opt IteratorOptions) Iterator {
	inputs = Iterators(inputs).filterNonNil()
	if n := len(inputs); n == 0 {
		return nil
	} else if n == 1 {
		return inputs[0]
	}

	// Aggregate functions can use a more relaxed sorting so that points
	// within a window are grouped. This is much more efficient.
	switch inputs := Iterators(inputs).coerce().(type) {
	case []FloatIterator:
		return newFloatMergeIterator(inputs, opt)
	case []IntegerIterator:
		return newIntegerMergeIterator(inputs, opt)
	case []UnsignedIterator:
		return newUnsignedMergeIterator(inputs, opt)
	case []StringIterator:
		return newStringMergeIterator(inputs, opt)
	case []BooleanIterator:
		return newBooleanMergeIterator(inputs, opt)
	default:
		panic(fmt.Sprintf("unsupported merge iterator type: %T", inputs))
	}
}

// NewParallelMergeIterator returns an iterator that breaks input iterators
// into groups and processes them in parallel.
func NewParallelMergeIterator(inputs []Iterator, opt IteratorOptions, parallelism int) Iterator {
	inputs = Iterators(inputs).filterNonNil()
	if len(inputs) == 0 {
		return nil
	} else if len(inputs) == 1 {
		return inputs[0]
	}

	// Limit parallelism to the number of inputs.
	if len(inputs) < parallelism {
		parallelism = len(inputs)
	}

	// Determine the number of inputs per output iterator.
	n := len(inputs) / parallelism

	// Group iterators together.
	outputs := make([]Iterator, parallelism)
	for i := range outputs {
		var slice []Iterator
		if i < len(outputs)-1 {
			slice = inputs[i*n : (i+1)*n]
		} else {
			slice = inputs[i*n:]
		}

		outputs[i] = newParallelIterator(NewMergeIterator(slice, opt))
	}

	// Merge all groups together.
	return NewMergeIterator(outputs, opt)
}

// NewSortedMergeIterator returns an iterator to merge itrs into one.
// Inputs must either be sorted merge iterators or only contain a single
// name/tag in sorted order. The iterator will output all points by name/tag,
// then time. This iterator is useful when you need all points for a name/tag
// to be in order.
func NewSortedMergeIterator(inputs []Iterator, opt IteratorOptions) Iterator {
	inputs = Iterators(inputs).filterNonNil()
	if len(inputs) == 0 {
		return nil
	} else if len(inputs) == 1 {
		return inputs[0]
	}

	switch inputs := Iterators(inputs).coerce().(type) {
	case []FloatIterator:
		return newFloatSortedMergeIterator(inputs, opt)
	case []IntegerIterator:
		return newIntegerSortedMergeIterator(inputs, opt)
	case []UnsignedIterator:
		return newUnsignedSortedMergeIterator(inputs, opt)
	case []StringIterator:
		return newStringSortedMergeIterator(inputs, opt)
	case []BooleanIterator:
		return newBooleanSortedMergeIterator(inputs, opt)
	default:
		panic(fmt.Sprintf("unsupported sorted merge iterator type: %T", inputs))
	}
}

// newParallelIterator returns an iterator that runs in a separate goroutine.
func newParallelIterator(input Iterator) Iterator {
	if input == nil {
		return nil
	}

	switch itr := input.(type) {
	case FloatIterator:
		return newFloatParallelIterator(itr)
	case IntegerIterator:
		return newIntegerParallelIterator(itr)
	case UnsignedIterator:
		return newUnsignedParallelIterator(itr)
	case StringIterator:
		return newStringParallelIterator(itr)
	case BooleanIterator:
		return newBooleanParallelIterator(itr)
	default:
		panic(fmt.Sprintf("unsupported parallel iterator type: %T", itr))
	}
}

// NewLimitIterator returns an iterator that limits the number of points per grouping.
func NewLimitIterator(input Iterator, opt IteratorOptions) Iterator {
	switch input := input.(type) {
	case FloatIterator:
		return newFloatLimitIterator(input, opt)
	case IntegerIterator:
		return newIntegerLimitIterator(input, opt)
	case UnsignedIterator:
		return newUnsignedLimitIterator(input, opt)
	case StringIterator:
		return newStringLimitIterator(input, opt)
	case BooleanIterator:
		return newBooleanLimitIterator(input, opt)
	default:
		panic(fmt.Sprintf("unsupported limit iterator type: %T", input))
	}
}

// NewFilterIterator returns an iterator that filters the points based on the
// condition. This iterator is not nearly as efficient as filtering points
// within the query engine and is only used when filtering subqueries.
func NewFilterIterator(input Iterator, cond influxql.Expr, opt IteratorOptions) Iterator {
	if input == nil {
		return nil
	}

	switch input := input.(type) {
	case FloatIterator:
		return newFloatFilterIterator(input, cond, opt)
	case IntegerIterator:
		return newIntegerFilterIterator(input, cond, opt)
	case UnsignedIterator:
		return newUnsignedFilterIterator(input, cond, opt)
	case StringIterator:
		return newStringFilterIterator(input, cond, opt)
	case BooleanIterator:
		return newBooleanFilterIterator(input, cond, opt)
	default:
		panic(fmt.Sprintf("unsupported filter iterator type: %T", input))
	}
}

// NewTagSubsetIterator will strip each of the points to a subset of the tag key values
// for each point it processes.
func NewTagSubsetIterator(input Iterator, opt IteratorOptions) Iterator {
	if input == nil {
		return nil
	}

	switch input := input.(type) {
	case FloatIterator:
		return newFloatTagSubsetIterator(input, opt)
	case IntegerIterator:
		return newIntegerTagSubsetIterator(input, opt)
	case UnsignedIterator:
		return newUnsignedTagSubsetIterator(input, opt)
	case StringIterator:
		return newStringTagSubsetIterator(input, opt)
	case BooleanIterator:
		return newBooleanTagSubsetIterator(input, opt)
	default:
		panic(fmt.Sprintf("unsupported tag subset iterator type: %T", input))
	}
}

// NewDedupeIterator returns an iterator that only outputs unique points.
// This iterator maintains a serialized copy of each row so it is inefficient
// to use on large datasets. It is intended for small datasets such as meta queries.
func NewDedupeIterator(input Iterator) Iterator {
	if input == nil {
		return nil
	}

	switch input := input.(type) {
	case FloatIterator:
		return newFloatDedupeIterator(input)
	case IntegerIterator:
		return newIntegerDedupeIterator(input)
	case UnsignedIterator:
		return newUnsignedDedupeIterator(input)
	case StringIterator:
		return newStringDedupeIterator(input)
	case BooleanIterator:
		return newBooleanDedupeIterator(input)
	default:
		panic(fmt.Sprintf("unsupported dedupe iterator type: %T", input))
	}
}

// NewFillIterator returns an iterator that fills in missing points in an aggregate.
func NewFillIterator(input Iterator, expr influxql.Expr, opt IteratorOptions) Iterator {
	switch input := input.(type) {
	case FloatIterator:
		return newFloatFillIterator(input, expr, opt)
	case IntegerIterator:
		return newIntegerFillIterator(input, expr, opt)
	case UnsignedIterator:
		return newUnsignedFillIterator(input, expr, opt)
	case StringIterator:
		return newStringFillIterator(input, expr, opt)
	case BooleanIterator:
		return newBooleanFillIterator(input, expr, opt)
	default:
		panic(fmt.Sprintf("unsupported fill iterator type: %T", input))
	}
}

// NewIntervalIterator returns an iterator that sets the time on each point to the interval.
func NewIntervalIterator(input Iterator, opt IteratorOptions) Iterator {
	switch input := input.(type) {
	case FloatIterator:
		return newFloatIntervalIterator(input, opt)
	case IntegerIterator:
		return newIntegerIntervalIterator(input, opt)
	case UnsignedIterator:
		return newUnsignedIntervalIterator(input, opt)
	case StringIterator:
		return newStringIntervalIterator(input, opt)
	case BooleanIterator:
		return newBooleanIntervalIterator(input, opt)
	default:
		panic(fmt.Sprintf("unsupported interval iterator type: %T", input))
	}
}

// NewInterruptIterator returns an iterator that will stop producing output
// when the passed-in channel is closed.
func NewInterruptIterator(input Iterator, closing <-chan struct{}) Iterator {
	switch input := input.(type) {
	case FloatIterator:
		return newFloatInterruptIterator(input, closing)
	case IntegerIterator:
		return newIntegerInterruptIterator(input, closing)
	case UnsignedIterator:
		return newUnsignedInterruptIterator(input, closing)
	case StringIterator:
		return newStringInterruptIterator(input, closing)
	case BooleanIterator:
		return newBooleanInterruptIterator(input, closing)
	default:
		panic(fmt.Sprintf("unsupported interrupt iterator type: %T", input))
	}
}

// IteratorScanner is used to scan the results of an iterator into a map.
type IteratorScanner interface {
	// Peek retrieves information about the next point. It returns a timestamp, the name, and the tags.
	Peek() (int64, string, Tags)

	// ScanAt will take a time, name, and tags and scan the point that matches those into the map.
	ScanAt(ts int64, name string, tags Tags, values map[string]interface{})

	// Stats returns the IteratorStats from the Iterator.
	Stats() IteratorStats

	// Err returns an error that was encountered while scanning.
	Err() error

	io.Closer
}

// SkipDefault is a sentinel value to tell the IteratorScanner to skip setting the
// default value if none was present. This causes the map to use the previous value
// if it was previously set.
var SkipDefault = interface{}(0)

// NewIteratorScanner produces an IteratorScanner for the Iterator.
func NewIteratorScanner(input Iterator, keys []influxql.VarRef, defaultValue interface{}) IteratorScanner {
	switch input := input.(type) {
	case FloatIterator:
		return newFloatIteratorScanner(input, keys, defaultValue)
	case IntegerIterator:
		return newIntegerIteratorScanner(input, keys, defaultValue)
	case UnsignedIterator:
		return newUnsignedIteratorScanner(input, keys, defaultValue)
	case StringIterator:
		return newStringIteratorScanner(input, keys, defaultValue)
	case BooleanIterator:
		return newBooleanIteratorScanner(input, keys, defaultValue)
	default:
		panic(fmt.Sprintf("unsupported type for iterator scanner: %T", input))
	}
}

// DrainIterator reads and discards all points from itr.
func DrainIterator(itr Iterator) {
	defer itr.Close()
	switch itr := itr.(type) {
	case FloatIterator:
		for p, _ := itr.Next(); p != nil; p, _ = itr.Next() {
		}
	case IntegerIterator:
		for p, _ := itr.Next(); p != nil; p, _ = itr.Next() {
		}
	case UnsignedIterator:
		for p, _ := itr.Next(); p != nil; p, _ = itr.Next() {
		}
	case StringIterator:
		for p, _ := itr.Next(); p != nil; p, _ = itr.Next() {
		}
	case BooleanIterator:
		for p, _ := itr.Next(); p != nil; p, _ = itr.Next() {
		}
	default:
		panic(fmt.Sprintf("unsupported iterator type for draining: %T", itr))
	}
}

// DrainIterators reads and discards all points from itrs.
func DrainIterators(itrs []Iterator) {
	defer Iterators(itrs).Close()
	for {
		var hasData bool

		for _, itr := range itrs {
			switch itr := itr.(type) {
			case FloatIterator:
				if p, _ := itr.Next(); p != nil {
					hasData = true
				}
			case IntegerIterator:
				if p, _ := itr.Next(); p != nil {
					hasData = true
				}
			case UnsignedIterator:
				if p, _ := itr.Next(); p != nil {
					hasData = true
				}
			case StringIterator:
				if p, _ := itr.Next(); p != nil {
					hasData = true
				}
			case BooleanIterator:
				if p, _ := itr.Next(); p != nil {
					hasData = true
				}
			default:
				panic(fmt.Sprintf("unsupported iterator type for draining: %T", itr))
			}
		}

		// Exit once all iterators return a nil point.
		if !hasData {
			break
		}
	}
}

// NewReaderIterator returns an iterator that streams from a reader.
func NewReaderIterator(ctx context.Context, r io.Reader, typ influxql.DataType, stats IteratorStats) Iterator {
	switch typ {
	case influxql.Float:
		return newFloatReaderIterator(ctx, r, stats)
	case influxql.Integer:
		return newIntegerReaderIterator(ctx, r, stats)
	case influxql.Unsigned:
		return newUnsignedReaderIterator(ctx, r, stats)
	case influxql.String:
		return newStringReaderIterator(ctx, r, stats)
	case influxql.Boolean:
		return newBooleanReaderIterator(ctx, r, stats)
	default:
		return &nilFloatReaderIterator{r: r}
	}
}

// IteratorCreator is an interface to create Iterators.
type IteratorCreator interface {
	// Creates a simple iterator for use in an InfluxQL query.
	CreateIterator(ctx context.Context, source *influxql.Measurement, opt IteratorOptions) (Iterator, error)

	// Determines the potential cost for creating an iterator.
	IteratorCost(ctx context.Context, source *influxql.Measurement, opt IteratorOptions) (IteratorCost, error)
}

// IteratorOptions is an object passed to CreateIterator to specify creation options.
type IteratorOptions struct {
	// OrgID is the organization for which this query is being executed.
	OrgID influxdb.ID

	// Expression to iterate for.
	// This can be VarRef or a Call.
	Expr influxql.Expr

	// Auxiliary tags or values to also retrieve for the point.
	Aux []influxql.VarRef

	// Data sources from which to receive data. This is only used for encoding
	// measurements over RPC and is no longer used in the open source version.
	Sources []influxql.Source

	// Group by interval and tags.
	Interval   Interval
	Dimensions []string            // The final dimensions of the query (stays the same even in subqueries).
	GroupBy    map[string]struct{} // Dimensions to group points by in intermediate iterators.
	Location   *time.Location

	// Fill options.
	Fill      influxql.FillOption
	FillValue interface{}

	// Condition to filter by.
	Condition influxql.Expr

	// Time range for the iterator.
	StartTime int64
	EndTime   int64

	// Sorted in time ascending order if true.
	Ascending bool

	// Limits the number of points per series.
	Limit, Offset int

	// Limits the number of series.
	SLimit, SOffset int

	// Removes the measurement name. Useful for meta queries.
	StripName bool

	// Removes duplicate rows from raw queries.
	Dedupe bool

	// Determines if this is a query for raw data or an aggregate/selector.
	Ordered bool

	// Limits on the creation of iterators.
	MaxSeriesN int

	// If this channel is set and is closed, the iterator should try to exit
	// and close as soon as possible.
	InterruptCh <-chan struct{}

	// Authorizer can limit access to data
	Authorizer Authorizer
}

// newIteratorOptionsStmt creates the iterator options from stmt.
func newIteratorOptionsStmt(stmt *influxql.SelectStatement, sopt SelectOptions) (opt IteratorOptions, err error) {
	// Determine time range from the condition.
	valuer := &influxql.NowValuer{Location: stmt.Location}
	condition, timeRange, err := influxql.ConditionExpr(stmt.Condition, valuer)
	if err != nil {
		return IteratorOptions{}, err
	}

	if !timeRange.Min.IsZero() {
		opt.StartTime = timeRange.Min.UnixNano()
	} else {
		opt.StartTime = influxql.MinTime
	}
	if !timeRange.Max.IsZero() {
		opt.EndTime = timeRange.Max.UnixNano()
	} else {
		opt.EndTime = influxql.MaxTime
	}
	opt.Location = stmt.Location

	// Determine group by interval.
	interval, err := stmt.GroupByInterval()
	if err != nil {
		return opt, err
	}
	// Set duration to zero if a negative interval has been used.
	if interval < 0 {
		interval = 0
	} else if interval > 0 {
		opt.Interval.Offset, err = stmt.GroupByOffset()
		if err != nil {
			return opt, err
		}
	}
	opt.Interval.Duration = interval

	// Always request an ordered output for the top level iterators.
	// The emitter will always emit points as ordered.
	opt.Ordered = true

	// Determine dimensions.
	opt.GroupBy = make(map[string]struct{}, len(opt.Dimensions))
	for _, d := range stmt.Dimensions {
		if d, ok := d.Expr.(*influxql.VarRef); ok {
			opt.Dimensions = append(opt.Dimensions, d.Val)
			opt.GroupBy[d.Val] = struct{}{}
		}
	}

	opt.Condition = condition
	opt.Ascending = stmt.TimeAscending()
	opt.Dedupe = stmt.Dedupe
	opt.StripName = stmt.StripName

	opt.Fill, opt.FillValue = stmt.Fill, stmt.FillValue
	if opt.Fill == influxql.NullFill && stmt.Target != nil {
		// Set the fill option to none if a target has been given.
		// Null values will get ignored when being written to the target
		// so fill(null) wouldn't write any null values to begin with.
		opt.Fill = influxql.NoFill
	}
	opt.Limit, opt.Offset = stmt.Limit, stmt.Offset
	opt.SLimit, opt.SOffset = stmt.SLimit, stmt.SOffset
	opt.MaxSeriesN = sopt.MaxSeriesN
	opt.OrgID = sopt.OrgID

	return opt, nil
}

func newIteratorOptionsSubstatement(ctx context.Context, stmt *influxql.SelectStatement, opt IteratorOptions) (IteratorOptions, error) {
	subOpt, err := newIteratorOptionsStmt(stmt, SelectOptions{
		OrgID:      opt.OrgID,
		MaxSeriesN: opt.MaxSeriesN,
	})
	if err != nil {
		return IteratorOptions{}, err
	}

	if subOpt.StartTime < opt.StartTime {
		subOpt.StartTime = opt.StartTime
	}
	if subOpt.EndTime > opt.EndTime {
		subOpt.EndTime = opt.EndTime
	}
	if !subOpt.Interval.IsZero() && subOpt.EndTime == influxql.MaxTime {
		if now := ctx.Value(nowKey); now != nil {
			subOpt.EndTime = now.(time.Time).UnixNano()
		}
	}
	// Propagate the dimensions to the inner subquery.
	subOpt.Dimensions = opt.Dimensions
	for d := range opt.GroupBy {
		subOpt.GroupBy[d] = struct{}{}
	}
	subOpt.InterruptCh = opt.InterruptCh

	// Extract the time range and condition from the condition.
	valuer := &influxql.NowValuer{Location: stmt.Location}
	cond, t, err := influxql.ConditionExpr(stmt.Condition, valuer)
	if err != nil {
		return IteratorOptions{}, err
	}
	subOpt.Condition = cond
	// If the time range is more constrained, use it instead. A less constrained time
	// range should be ignored.
	if !t.Min.IsZero() && t.MinTimeNano() > opt.StartTime {
		subOpt.StartTime = t.MinTimeNano()
	}
	if !t.Max.IsZero() && t.MaxTimeNano() < opt.EndTime {
		subOpt.EndTime = t.MaxTimeNano()
	}

	// Propagate the SLIMIT and SOFFSET from the outer query.
	subOpt.SLimit += opt.SLimit
	subOpt.SOffset += opt.SOffset

	// Propagate the ordering from the parent query.
	subOpt.Ascending = opt.Ascending

	// If the inner query uses a null fill option and is not a raw query,
	// switch it to none so we don't hit an unnecessary penalty from the
	// fill iterator. Null values will end up getting stripped by an outer
	// query anyway so there's no point in having them here. We still need
	// all other types of fill iterators because they can affect the result
	// of the outer query. We also do not do this for raw queries because
	// there is no fill iterator for them and fill(none) doesn't work with
	// raw queries.
	if !stmt.IsRawQuery && subOpt.Fill == influxql.NullFill {
		subOpt.Fill = influxql.NoFill
	}

	// Inherit the ordering method from the outer query.
	subOpt.Ordered = opt.Ordered

	// If there is no interval for this subquery, but the outer query has an
	// interval, inherit the parent interval.
	interval, err := stmt.GroupByInterval()
	if err != nil {
		return IteratorOptions{}, err
	} else if interval == 0 {
		subOpt.Interval = opt.Interval
	}
	return subOpt, nil
}

// MergeSorted returns true if the options require a sorted merge.
func (opt IteratorOptions) MergeSorted() bool {
	return opt.Ordered
}

// SeekTime returns the time the iterator should start from.
// For ascending iterators this is the start time, for descending iterators it's the end time.
func (opt IteratorOptions) SeekTime() int64 {
	if opt.Ascending {
		return opt.StartTime
	}
	return opt.EndTime
}

// StopTime returns the time the iterator should end at.
// For ascending iterators this is the end time, for descending iterators it's the start time.
func (opt IteratorOptions) StopTime() int64 {
	if opt.Ascending {
		return opt.EndTime
	}
	return opt.StartTime
}

// Window returns the time window [start,end) that t falls within.
func (opt IteratorOptions) Window(t int64) (start, end int64) {
	if opt.Interval.IsZero() {
		return opt.StartTime, opt.EndTime + 1
	}

	// Subtract the offset to the time so we calculate the correct base interval.
	t -= int64(opt.Interval.Offset)

	// Retrieve the zone offset for the start time.
	var zone int64
	if opt.Location != nil {
		_, zone = opt.Zone(t)
	}

	// Truncate time by duration.
	dt := (t + zone) % int64(opt.Interval.Duration)
	if dt < 0 {
		// Negative modulo rounds up instead of down, so offset
		// with the duration.
		dt += int64(opt.Interval.Duration)
	}

	// Find the start time.
	if influxql.MinTime+dt >= t {
		start = influxql.MinTime
	} else {
		start = t - dt
	}

	// Look for the start offset again because the first time may have been
	// after the offset switch. Now that we are at midnight in UTC, we can
	// lookup the zone offset again to get the real starting offset.
	if opt.Location != nil {
		_, startOffset := opt.Zone(start)
		// Do not adjust the offset if the offset change is greater than or
		// equal to the duration.
		if o := zone - startOffset; o != 0 && abs(o) < int64(opt.Interval.Duration) {
			start += o
		}
	}
	start += int64(opt.Interval.Offset)

	// Find the end time.
	if dt := int64(opt.Interval.Duration) - dt; influxql.MaxTime-dt <= t {
		end = influxql.MaxTime
	} else {
		end = t + dt
	}

	// Retrieve the zone offset for the end time.
	if opt.Location != nil {
		_, endOffset := opt.Zone(end)
		// Adjust the end time if the offset is different from the start offset.
		// Only apply the offset if it is smaller than the duration.
		// This prevents going back in time and creating time windows
		// that don't make any sense.
		if o := zone - endOffset; o != 0 && abs(o) < int64(opt.Interval.Duration) {
			// If the offset is greater than 0, that means we are adding time.
			// Added time goes into the previous interval because the clocks
			// move backwards. If the offset is less than 0, then we are skipping
			// time. Skipped time comes after the switch so if we have a time
			// interval that lands on the switch, it comes from the next
			// interval and not the current one. For this reason, we need to know
			// when the actual switch happens by seeing if the time switch is within
			// the current interval. We calculate the zone offset with the offset
			// and see if the value is the same. If it is, we apply the
			// offset.
			if o > 0 {
				end += o
			} else if _, z := opt.Zone(end + o); z == endOffset {
				end += o
			}
		}
	}
	end += int64(opt.Interval.Offset)
	return
}

// DerivativeInterval returns the time interval for the derivative function.
func (opt IteratorOptions) DerivativeInterval() Interval {
	// Use the interval on the derivative() call, if specified.
	if expr, ok := opt.Expr.(*influxql.Call); ok && len(expr.Args) == 2 {
		return Interval{Duration: expr.Args[1].(*influxql.DurationLiteral).Val}
	}

	// Otherwise use the group by interval, if specified.
	if opt.Interval.Duration > 0 {
		return Interval{Duration: opt.Interval.Duration}
	}

	return Interval{Duration: time.Second}
}

// ElapsedInterval returns the time interval for the elapsed function.
func (opt IteratorOptions) ElapsedInterval() Interval {
	// Use the interval on the elapsed() call, if specified.
	if expr, ok := opt.Expr.(*influxql.Call); ok && len(expr.Args) == 2 {
		return Interval{Duration: expr.Args[1].(*influxql.DurationLiteral).Val}
	}

	return Interval{Duration: time.Nanosecond}
}

// IntegralInterval returns the time interval for the integral function.
func (opt IteratorOptions) IntegralInterval() Interval {
	// Use the interval on the integral() call, if specified.
	if expr, ok := opt.Expr.(*influxql.Call); ok && len(expr.Args) == 2 {
		return Interval{Duration: expr.Args[1].(*influxql.DurationLiteral).Val}
	}

	return Interval{Duration: time.Second}
}

// GetDimensions retrieves the dimensions for this query.
func (opt IteratorOptions) GetDimensions() []string {
	if len(opt.GroupBy) > 0 {
		dimensions := make([]string, 0, len(opt.GroupBy))
		for dim := range opt.GroupBy {
			dimensions = append(dimensions, dim)
		}
		return dimensions
	}
	return opt.Dimensions
}

// Zone returns the zone information for the given time. The offset is in nanoseconds.
func (opt *IteratorOptions) Zone(ns int64) (string, int64) {
	if opt.Location == nil {
		return "", 0
	}

	t := time.Unix(0, ns).In(opt.Location)
	name, offset := t.Zone()
	return name, secToNs * int64(offset)
}

// MarshalBinary encodes opt into a binary format.
func (opt *IteratorOptions) MarshalBinary() ([]byte, error) {
	return proto.Marshal(encodeIteratorOptions(opt))
}

// UnmarshalBinary decodes from a binary format in to opt.
func (opt *IteratorOptions) UnmarshalBinary(buf []byte) error {
	var pb internal.IteratorOptions
	if err := proto.Unmarshal(buf, &pb); err != nil {
		return err
	}

	other, err := decodeIteratorOptions(&pb)
	if err != nil {
		return err
	}
	*opt = *other

	return nil
}

func encodeIteratorOptions(opt *IteratorOptions) *internal.IteratorOptions {
	pb := &internal.IteratorOptions{
		Interval:   encodeInterval(opt.Interval),
		Dimensions: opt.Dimensions,
		Fill:       proto.Int32(int32(opt.Fill)),
		StartTime:  proto.Int64(opt.StartTime),
		EndTime:    proto.Int64(opt.EndTime),
		Ascending:  proto.Bool(opt.Ascending),
		Limit:      proto.Int64(int64(opt.Limit)),
		Offset:     proto.Int64(int64(opt.Offset)),
		SLimit:     proto.Int64(int64(opt.SLimit)),
		SOffset:    proto.Int64(int64(opt.SOffset)),
		StripName:  proto.Bool(opt.StripName),
		Dedupe:     proto.Bool(opt.Dedupe),
		MaxSeriesN: proto.Int64(int64(opt.MaxSeriesN)),
		Ordered:    proto.Bool(opt.Ordered),
	}

	// Set expression, if set.
	if opt.Expr != nil {
		pb.Expr = proto.String(opt.Expr.String())
	}

	// Set the location, if set.
	if opt.Location != nil {
		pb.Location = proto.String(opt.Location.String())
	}

	// Convert and encode aux fields as variable references.
	if opt.Aux != nil {
		pb.Fields = make([]*internal.VarRef, len(opt.Aux))
		pb.Aux = make([]string, len(opt.Aux))
		for i, ref := range opt.Aux {
			pb.Fields[i] = encodeVarRef(ref)
			pb.Aux[i] = ref.Val
		}
	}

	// Encode group by dimensions from a map.
	if opt.GroupBy != nil {
		dimensions := make([]string, 0, len(opt.GroupBy))
		for dim := range opt.GroupBy {
			dimensions = append(dimensions, dim)
		}
		pb.GroupBy = dimensions
	}

	// Convert and encode sources to measurements.
	if opt.Sources != nil {
		sources := make([]*internal.Measurement, len(opt.Sources))
		for i, source := range opt.Sources {
			mm := source.(*influxql.Measurement)
			sources[i] = encodeMeasurement(mm)
		}
		pb.Sources = sources
	}

	// Fill value can only be a number. Set it if available.
	if v, ok := opt.FillValue.(float64); ok {
		pb.FillValue = proto.Float64(v)
	}

	// Set condition, if set.
	if opt.Condition != nil {
		pb.Condition = proto.String(opt.Condition.String())
	}

	return pb
}

func decodeIteratorOptions(pb *internal.IteratorOptions) (*IteratorOptions, error) {
	opt := &IteratorOptions{
		Interval:   decodeInterval(pb.GetInterval()),
		Dimensions: pb.GetDimensions(),
		Fill:       influxql.FillOption(pb.GetFill()),
		StartTime:  pb.GetStartTime(),
		EndTime:    pb.GetEndTime(),
		Ascending:  pb.GetAscending(),
		Limit:      int(pb.GetLimit()),
		Offset:     int(pb.GetOffset()),
		SLimit:     int(pb.GetSLimit()),
		SOffset:    int(pb.GetSOffset()),
		StripName:  pb.GetStripName(),
		Dedupe:     pb.GetDedupe(),
		MaxSeriesN: int(pb.GetMaxSeriesN()),
		Ordered:    pb.GetOrdered(),
	}

	// Set expression, if set.
	if pb.Expr != nil {
		expr, err := influxql.ParseExpr(pb.GetExpr())
		if err != nil {
			return nil, err
		}
		opt.Expr = expr
	}

	if pb.Location != nil {
		loc, err := time.LoadLocation(pb.GetLocation())
		if err != nil {
			return nil, err
		}
		opt.Location = loc
	}

	// Convert and decode variable references.
	if fields := pb.GetFields(); fields != nil {
		opt.Aux = make([]influxql.VarRef, len(fields))
		for i, ref := range fields {
			opt.Aux[i] = decodeVarRef(ref)
		}
	} else if aux := pb.GetAux(); aux != nil {
		opt.Aux = make([]influxql.VarRef, len(aux))
		for i, name := range aux {
			opt.Aux[i] = influxql.VarRef{Val: name}
		}
	}

	// Convert and decode sources to measurements.
	if pb.Sources != nil {
		sources := make([]influxql.Source, len(pb.GetSources()))
		for i, source := range pb.GetSources() {
			mm, err := decodeMeasurement(source)
			if err != nil {
				return nil, err
			}
			sources[i] = mm
		}
		opt.Sources = sources
	}

	// Convert group by dimensions to a map.
	if pb.GroupBy != nil {
		dimensions := make(map[string]struct{}, len(pb.GroupBy))
		for _, dim := range pb.GetGroupBy() {
			dimensions[dim] = struct{}{}
		}
		opt.GroupBy = dimensions
	}

	// Set the fill value, if set.
	if pb.FillValue != nil {
		opt.FillValue = pb.GetFillValue()
	}

	// Set condition, if set.
	if pb.Condition != nil {
		expr, err := influxql.ParseExpr(pb.GetCondition())
		if err != nil {
			return nil, err
		}
		opt.Condition = expr
	}

	return opt, nil
}

func encodeMeasurement(mm *influxql.Measurement) *internal.Measurement {
	pb := &internal.Measurement{
		Database:        proto.String(mm.Database),
		RetentionPolicy: proto.String(mm.RetentionPolicy),
		Name:            proto.String(mm.Name),
		SystemIterator:  proto.String(mm.SystemIterator),
		IsTarget:        proto.Bool(mm.IsTarget),
	}
	if mm.Regex != nil {
		pb.Regex = proto.String(mm.Regex.Val.String())
	}
	return pb
}

func decodeMeasurement(pb *internal.Measurement) (*influxql.Measurement, error) {
	mm := &influxql.Measurement{
		Database:        pb.GetDatabase(),
		RetentionPolicy: pb.GetRetentionPolicy(),
		Name:            pb.GetName(),
		SystemIterator:  pb.GetSystemIterator(),
		IsTarget:        pb.GetIsTarget(),
	}

	if pb.Regex != nil {
		regex, err := regexp.Compile(pb.GetRegex())
		if err != nil {
			return nil, fmt.Errorf("invalid binary measurement regex: value=%q, err=%s", pb.GetRegex(), err)
		}
		mm.Regex = &influxql.RegexLiteral{Val: regex}
	}

	return mm, nil
}

// Interval represents a repeating interval for a query.
type Interval struct {
	Duration time.Duration
	Offset   time.Duration
}

// IsZero returns true if the interval has no duration.
func (i Interval) IsZero() bool { return i.Duration == 0 }

func encodeInterval(i Interval) *internal.Interval {
	return &internal.Interval{
		Duration: proto.Int64(i.Duration.Nanoseconds()),
		Offset:   proto.Int64(i.Offset.Nanoseconds()),
	}
}

func decodeInterval(pb *internal.Interval) Interval {
	return Interval{
		Duration: time.Duration(pb.GetDuration()),
		Offset:   time.Duration(pb.GetOffset()),
	}
}

func encodeVarRef(ref influxql.VarRef) *internal.VarRef {
	return &internal.VarRef{
		Val:  proto.String(ref.Val),
		Type: proto.Int32(int32(ref.Type)),
	}
}

func decodeVarRef(pb *internal.VarRef) influxql.VarRef {
	return influxql.VarRef{
		Val:  pb.GetVal(),
		Type: influxql.DataType(pb.GetType()),
	}
}

type nilFloatIterator struct{}

func (*nilFloatIterator) Stats() IteratorStats       { return IteratorStats{} }
func (*nilFloatIterator) Close() error               { return nil }
func (*nilFloatIterator) Next() (*FloatPoint, error) { return nil, nil }

type nilFloatReaderIterator struct {
	r io.Reader
}

func (*nilFloatReaderIterator) Stats() IteratorStats { return IteratorStats{} }
func (itr *nilFloatReaderIterator) Close() error {
	if r, ok := itr.r.(io.ReadCloser); ok {
		itr.r = nil
		return r.Close()
	}
	return nil
}
func (*nilFloatReaderIterator) Next() (*FloatPoint, error) { return nil, nil }

// IteratorStats represents statistics about an iterator.
// Some statistics are available immediately upon iterator creation while
// some are derived as the iterator processes data.
type IteratorStats struct {
	SeriesN int // series represented
	PointN  int // points returned
}

// Add aggregates fields from s and other together. Overwrites s.
func (s *IteratorStats) Add(other IteratorStats) {
	s.SeriesN += other.SeriesN
	s.PointN += other.PointN
}

func encodeIteratorStats(stats *IteratorStats) *internal.IteratorStats {
	return &internal.IteratorStats{
		SeriesN: proto.Int64(int64(stats.SeriesN)),
		PointN:  proto.Int64(int64(stats.PointN)),
	}
}

func decodeIteratorStats(pb *internal.IteratorStats) IteratorStats {
	return IteratorStats{
		SeriesN: int(pb.GetSeriesN()),
		PointN:  int(pb.GetPointN()),
	}
}

// IteratorCost contains statistics retrieved for explaining what potential
// cost may be incurred by instantiating an iterator.
type IteratorCost struct {
	// The total number of shards that are touched by this query.
	NumShards int64

	// The total number of non-unique series that are accessed by this query.
	// This number matches the number of cursors created by the query since
	// one cursor is created for every series.
	NumSeries int64

	// CachedValues returns the number of cached values that may be read by this
	// query.
	CachedValues int64

	// The total number of non-unique files that may be accessed by this query.
	// This will count the number of files accessed by each series so files
	// will likely be double counted.
	NumFiles int64

	// The number of blocks that had the potential to be accessed.
	BlocksRead int64

	// The amount of data that can be potentially read.
	BlockSize int64
}

// Combine combines the results of two IteratorCost structures into one.
func (c IteratorCost) Combine(other IteratorCost) IteratorCost {
	return IteratorCost{
		NumShards:    c.NumShards + other.NumShards,
		NumSeries:    c.NumSeries + other.NumSeries,
		CachedValues: c.CachedValues + other.CachedValues,
		NumFiles:     c.NumFiles + other.NumFiles,
		BlocksRead:   c.BlocksRead + other.BlocksRead,
		BlockSize:    c.BlockSize + other.BlockSize,
	}
}

// floatFastDedupeIterator outputs unique points where the point has a single aux field.
type floatFastDedupeIterator struct {
	input FloatIterator
	m     map[fastDedupeKey]struct{} // lookup of points already sent
}

// newFloatFastDedupeIterator returns a new instance of floatFastDedupeIterator.
func newFloatFastDedupeIterator(input FloatIterator) *floatFastDedupeIterator {
	return &floatFastDedupeIterator{
		input: input,
		m:     make(map[fastDedupeKey]struct{}),
	}
}

// Stats returns stats from the input iterator.
func (itr *floatFastDedupeIterator) Stats() IteratorStats { return itr.input.Stats() }

// Close closes the iterator and all child iterators.
func (itr *floatFastDedupeIterator) Close() error { return itr.input.Close() }

// Next returns the next unique point from the input iterator.
func (itr *floatFastDedupeIterator) Next() (*FloatPoint, error) {
	for {
		// Read next point.
		// Skip if there are not any aux fields.
		p, err := itr.input.Next()
		if p == nil || err != nil {
			return nil, err
		} else if len(p.Aux) == 0 {
			continue
		}

		// If the point has already been output then move to the next point.
		key := fastDedupeKey{name: p.Name}
		key.values[0] = p.Aux[0]
		if len(p.Aux) > 1 {
			key.values[1] = p.Aux[1]
		}
		if _, ok := itr.m[key]; ok {
			continue
		}

		// Otherwise mark it as emitted and return point.
		itr.m[key] = struct{}{}
		return p, nil
	}
}

type fastDedupeKey struct {
	name   string
	values [2]interface{}
}

func abs(v int64) int64 {
	sign := v >> 63
	return (v ^ sign) - sign
}

// IteratorEncoder is an encoder for encoding an iterator's points to w.
type IteratorEncoder struct {
	w io.Writer

	// Frequency with which stats are emitted.
	StatsInterval time.Duration
}

// NewIteratorEncoder encodes an iterator's points to w.
func NewIteratorEncoder(w io.Writer) *IteratorEncoder {
	return &IteratorEncoder{
		w: w,

		StatsInterval: DefaultStatsInterval,
	}
}

// EncodeIterator encodes and writes all of itr's points to the underlying writer.
func (enc *IteratorEncoder) EncodeIterator(itr Iterator) error {
	switch itr := itr.(type) {
	case FloatIterator:
		return enc.encodeFloatIterator(itr)
	case IntegerIterator:
		return enc.encodeIntegerIterator(itr)
	case StringIterator:
		return enc.encodeStringIterator(itr)
	case BooleanIterator:
		return enc.encodeBooleanIterator(itr)
	default:
		panic(fmt.Sprintf("unsupported iterator for encoder: %T", itr))
	}
}

// encode a stats object in the point stream.
func (enc *IteratorEncoder) encodeStats(stats IteratorStats) error {
	buf, err := proto.Marshal(&internal.Point{
		Name: proto.String(""),
		Tags: proto.String(""),
		Time: proto.Int64(0),
		Nil:  proto.Bool(false),

		Stats: encodeIteratorStats(&stats),
	})
	if err != nil {
		return err
	}

	if err = binary.Write(enc.w, binary.BigEndian, uint32(len(buf))); err != nil {
		return err
	}
	if _, err = enc.w.Write(buf); err != nil {
		return err
	}
	return nil
}