func (self *ShardData) QueryResponseBufferSize(querySpec *parser.QuerySpec, batchPointSize int) int {
groupByTime := querySpec.GetGroupByInterval()
if groupByTime == nil {
// If the group by time is nil, we shouldn't have to use a buffer since the shards should be queried sequentially.
// However, set this to something high just to be safe.
log.Debug("BUFFER SIZE: 1000")
return 1000
}
tickCount := int(self.shardNanoseconds / uint64(*groupByTime))
if tickCount < 10 {
tickCount = 100
} else if tickCount > 1000 {
// cap this because each response should have up to this number of points in it.
tickCount = tickCount / batchPointSize
// but make sure it's at least 1k
if tickCount < 1000 {
tickCount = 1000
}
}
columnCount := querySpec.GetGroupByColumnCount()
if columnCount > 1 {
// we don't really know the cardinality for any column up front. This is a just a multiplier so we'll see how this goes.
// each response can have many points, so having a buffer of the ticks * 100 should be safe, but we'll see.
tickCount = tickCount * 100
}
log.Debug("BUFFER SIZE: %d", tickCount)
return tickCount
}
func (self *Coordinator) getShardsAndProcessor(querySpec *parser.QuerySpec, writer engine.Processor) ([]*cluster.ShardData, engine.Processor, error) {
shards, err := self.clusterConfiguration.GetShardsForQuery(querySpec)
if err != nil {
return nil, nil, err
}
shouldAggregateLocally := shards.ShouldAggregateLocally(querySpec)
q := querySpec.SelectQuery()
if q == nil {
return shards, writer, nil
}
if !shouldAggregateLocally {
// if we should aggregate in the coordinator (i.e. aggregation
// isn't happening locally at the shard level), create an engine
shardIds := make([]uint32, len(shards))
for i, s := range shards {
shardIds[i] = s.Id()
}
writer, err = engine.NewQueryEngine(writer, q, shardIds)
if err != nil {
log.Error(err)
log.Debug("Coordinator processor chain: %s", engine.ProcessorChain(writer))
}
return shards, writer, err
}
// if we have a query with limit, then create an engine, or we can
// make the passthrough limit aware
writer = engine.NewPassthroughEngineWithLimit(writer, 100, q.Limit)
return shards, writer, nil
}
func (self *ClusterConfiguration) getShardsToMatchQuery(querySpec *parser.QuerySpec) ([]*ShardData, error) {
self.shardLock.RLock()
defer self.shardLock.RUnlock()
seriesNames, fromRegex := querySpec.TableNamesAndRegex()
db := querySpec.Database()
if fromRegex != nil {
seriesNames = self.MetaStore.GetSeriesForDatabaseAndRegex(db, fromRegex)
}
uniqueShards := make(map[uint32]*ShardData)
for _, name := range seriesNames {
if fs := self.MetaStore.GetFieldsForSeries(db, name); len(fs) == 0 {
return nil, fmt.Errorf("Couldn't find series: %s", name)
}
space := self.getShardSpaceToMatchSeriesName(db, name)
if space == nil {
continue
}
for _, shard := range space.shards {
uniqueShards[shard.id] = shard
}
}
shards := make([]*ShardData, 0, len(uniqueShards))
for _, shard := range uniqueShards {
shards = append(shards, shard)
}
SortShardsByTimeDescending(shards)
return shards, nil
}
func (self *Permissions) CheckQueryPermissions(user common.User, db string, querySpec *parser.QuerySpec) (ok bool, err common.AuthorizationError) {
switch querySpec.Query().Type() {
case parser.Delete:
return self.AuthorizeDeleteQuery(user, db)
case parser.Select:
return self.AuthorizeSelectQuery(user, db, querySpec)
default:
return true, ""
}
}
func (self *Shard) hasReadAccess(querySpec *parser.QuerySpec) bool {
for series := range querySpec.SeriesValuesAndColumns() {
if _, isRegex := series.GetCompiledRegex(); !isRegex {
if !querySpec.HasReadAccess(series.Name) {
return false
}
}
}
return true
}
func (self *Coordinator) expandRegex(spec *parser.QuerySpec) {
q := spec.SelectQuery()
if q == nil {
return
}
f := func(r *regexp.Regexp) []string {
return self.clusterConfiguration.MetaStore.GetSeriesForDatabaseAndRegex(spec.Database(), r)
}
parser.RewriteMergeQuery(q, f)
}
func (self *ClusterConfiguration) GetShardsForQuery(querySpec *parser.QuerySpec) (Shards, error) {
shards, err := self.getShardsToMatchQuery(querySpec)
if err != nil {
return nil, err
}
log.Debug("Querying %d shards for query", len(shards))
shards = self.getShardRange(querySpec, shards)
if querySpec.IsAscending() {
SortShardsByTimeAscending(shards)
}
return shards, nil
}
func (self *Shard) executeQueryForSeries(querySpec *parser.QuerySpec, name string, columns []string, processor engine.Processor) error {
if querySpec.IsSinglePointQuery() {
log.Debug("Running single query for series %s", name)
return self.executeSinglePointQuery(querySpec, name, columns, processor)
}
var pi *PointIterator
var err error
columns, pi, err = self.getPointIteratorForSeries(querySpec, name, columns)
if err != nil {
return err
}
defer pi.Close()
query := querySpec.SelectQuery()
aliases := query.GetTableAliases(name)
seriesOutgoing := &protocol.Series{Name: protocol.String(name), Fields: columns, Points: make([]*protocol.Point, 0, self.pointBatchSize)}
for pi.Valid() {
p := pi.Point()
seriesOutgoing.Points = append(seriesOutgoing.Points, p)
if len(seriesOutgoing.Points) >= self.pointBatchSize {
ok, err := yieldToProcessor(seriesOutgoing, processor, aliases)
if !ok || err != nil {
log.Debug("Stopping processing.")
if err != nil {
log.Error("Error while processing data: %v", err)
return err
}
return nil
}
seriesOutgoing = &protocol.Series{Name: protocol.String(name), Fields: columns, Points: make([]*protocol.Point, 0, self.pointBatchSize)}
}
pi.Next()
}
if err := pi.Error(); err != nil {
return err
}
//Yield remaining data
if ok, err := yieldToProcessor(seriesOutgoing, processor, aliases); !ok || err != nil {
log.Debug("Stopping processing remaining points...")
if err != nil {
log.Error("Error while processing data: %v", err)
return err
}
}
log.Debug("Finished running query %s", query.GetQueryString())
return nil
}
func (self *Permissions) AuthorizeSelectQuery(user common.User, db string, querySpec *parser.QuerySpec) (ok bool, err common.AuthorizationError) {
// if this isn't a regex query do the permission check here
fromClause := querySpec.SelectQuery().GetFromClause()
for _, n := range fromClause.Names {
if _, ok := n.Name.GetCompiledRegex(); ok {
break
} else if name := n.Name.Name; !user.HasReadAccess(name) {
return false, common.NewAuthorizationError("User doesn't have read access to %s", name)
}
}
return true, ""
}
func (self *Shard) executeSinglePointQuery(querySpec *parser.QuerySpec, name string, columns []string, p engine.Processor) error {
fields, err := self.getFieldsForSeries(querySpec.Database(), name, columns)
if err != nil {
log.Error("Error looking up fields for %s: %s", name, err)
return err
}
query := querySpec.SelectQuery()
fieldCount := len(fields)
fieldNames := make([]string, 0, fieldCount)
point := &protocol.Point{Values: make([]*protocol.FieldValue, 0, fieldCount)}
timestamp := common.TimeToMicroseconds(query.GetStartTime())
sequenceNumber, err := query.GetSinglePointQuerySequenceNumber()
if err != nil {
return err
}
// set the timestamp and sequence number
point.SequenceNumber = &sequenceNumber
point.SetTimestampInMicroseconds(timestamp)
for _, field := range fields {
sk := newStorageKey(field.Id, timestamp, sequenceNumber)
data, err := self.db.Get(sk.bytes())
if err != nil {
return err
}
if data == nil {
continue
}
fieldValue := &protocol.FieldValue{}
err = proto.Unmarshal(data, fieldValue)
if err != nil {
return err
}
fieldNames = append(fieldNames, field.Name)
point.Values = append(point.Values, fieldValue)
}
result := &protocol.Series{Name: &name, Fields: fieldNames, Points: []*protocol.Point{point}}
if len(result.Points) > 0 {
_, err := p.Yield(result)
return err
}
return nil
}
func (self *ShardData) getProcessor(querySpec *parser.QuerySpec, processor engine.Processor) (engine.Processor, error) {
switch qt := querySpec.Query().Type(); qt {
case parser.Delete, parser.DropSeries:
return NilProcessor{}, nil
case parser.Select:
// continue
default:
panic(fmt.Errorf("Unexpected query type: %s", qt))
}
if querySpec.IsSinglePointQuery() {
return engine.NewPassthroughEngine(processor, 1), nil
}
query := querySpec.SelectQuery()
var err error
// We should aggregate at the shard level
if self.ShouldAggregateLocally(querySpec) {
log.Debug("creating a query engine")
processor, err = engine.NewQueryEngine(processor, query, nil)
if err != nil {
return nil, err
}
goto addFilter
}
// we shouldn't limit the queries if they have aggregates and aren't
// aggregated locally, otherwise the aggregation result which happen
// in the coordinator will get partial data and will be incorrect
if query.HasAggregates() {
log.Debug("creating a passthrough engine")
processor = engine.NewPassthroughEngine(processor, 1000)
goto addFilter
}
// This is an optimization so we don't send more data that we should
// over the wire. The coordinator has its own Passthrough which does
// the final limit.
if l := query.Limit; l > 0 {
log.Debug("creating a passthrough engine with limit")
processor = engine.NewPassthroughEngineWithLimit(processor, 1000, query.Limit)
}
addFilter:
if query := querySpec.SelectQuery(); query != nil && query.GetFromClause().Type != parser.FromClauseInnerJoin {
// Joins do their own filtering since we need to get all
// points before filtering. This is due to the fact that some
// where expressions will be difficult to compute before the
// points are joined together, think where clause with
// left.column = 'something' or right.column =
// 'something_else'. We can't filter the individual series
// separately. The filtering happens in merge.go:55
processor = engine.NewFilteringEngine(query, processor)
}
return processor, nil
}
func (self *Coordinator) runDeleteQuery(querySpec *parser.QuerySpec, p engine.Processor) error {
if err := self.clusterConfiguration.CreateCheckpoint(); err != nil {
return err
}
querySpec.RunAgainstAllServersInShard = true
return self.runQuerySpec(querySpec, p)
}
func (self *Shard) executeArrayQuery(querySpec *parser.QuerySpec, processor engine.Processor) error {
seriesAndColumns := querySpec.SelectQuery().GetReferencedColumns()
for series, columns := range seriesAndColumns {
if regex, ok := series.GetCompiledRegex(); ok {
seriesNames := self.metaStore.GetSeriesForDatabaseAndRegex(querySpec.Database(), regex)
for _, name := range seriesNames {
if !querySpec.HasReadAccess(name) {
continue
}
err := self.executeQueryForSeries(querySpec, name, columns, processor)
if err != nil {
return err
}
}
} else {
err := self.executeQueryForSeries(querySpec, series.Name, columns, processor)
if err != nil {
return err
}
}
}
return nil
}
func (self *Shard) Query(querySpec *parser.QuerySpec, processor engine.Processor) error {
self.closeLock.RLock()
defer self.closeLock.RUnlock()
if self.closed {
return fmt.Errorf("Shard is closed")
}
if querySpec.IsListSeriesQuery() {
return fmt.Errorf("List series queries should never come to the shard")
} else if querySpec.IsDeleteFromSeriesQuery() {
return self.executeDeleteQuery(querySpec, processor)
}
if !self.hasReadAccess(querySpec) {
return errors.New("User does not have access to one or more of the series requested.")
}
switch t := querySpec.SelectQuery().FromClause.Type; t {
case parser.FromClauseArray:
log.Debug("Shard %s: running a regular query", self.db.Path())
return self.executeArrayQuery(querySpec, processor)
case parser.FromClauseMerge, parser.FromClauseInnerJoin:
log.Debug("Shard %s: running a merge query", self.db.Path())
return self.executeMergeQuery(querySpec, processor, t)
default:
panic(fmt.Errorf("Unknown from clause type %s", t))
}
}
func (self *Shard) executeMergeQuery(querySpec *parser.QuerySpec, processor engine.Processor, t parser.FromClauseType) error {
seriesAndColumns := querySpec.SelectQuery().GetReferencedColumns()
iterators := make([]*PointIterator, len(seriesAndColumns))
streams := make([]engine.StreamQuery, len(iterators))
i := 0
var err error
for s, c := range seriesAndColumns {
c, iterators[i], err = self.getPointIteratorForSeries(querySpec, s.Name, c)
if err != nil {
log.Error(err)
return err
}
defer iterators[i].Close()
aliases := querySpec.SelectQuery().GetTableAliases(s.Name)
if len(aliases) > 1 {
return fmt.Errorf("Cannot have the same table joined more than once")
}
streams[i] = PointIteratorStream{
pi: iterators[i],
name: aliases[0],
fields: c,
}
i++
}
h := engine.NewSeriesHeap(querySpec.IsAscending())
merger := engine.NewCME("Shard", streams, h, processor, t == parser.FromClauseMerge)
if _, err := merger.Update(); err != nil {
return err
}
return nil
}
func (self *Shard) executeDeleteQuery(querySpec *parser.QuerySpec, processor engine.Processor) error {
query := querySpec.DeleteQuery()
series := query.GetFromClause()
database := querySpec.Database()
if series.Type != parser.FromClauseArray {
return fmt.Errorf("Merge and Inner joins can't be used with a delete query: %v", series.Type)
}
for _, name := range series.Names {
var err error
if regex, ok := name.Name.GetCompiledRegex(); ok {
err = self.deleteRangeOfRegex(database, regex, query.GetStartTime(), query.GetEndTime())
} else {
err = self.deleteRangeOfSeries(database, name.Name.Name, query.GetStartTime(), query.GetEndTime())
}
if err != nil {
return err
}
}
self.db.Compact()
return nil
}
func (self *Coordinator) shouldQuerySequentially(shards cluster.Shards, querySpec *parser.QuerySpec) bool {
// if the query isn't a select, then it doesn't matter
if querySpec.SelectQuery() == nil {
return false
}
// if the query is a regex, we can't predic the number of responses
// we get back
if querySpec.IsRegex() {
return true
}
groupByClause := querySpec.SelectQuery().GetGroupByClause()
// if there's no group by clause, then we're returning raw points
// with some math done on them, thus we can't predict the number of
// points
if groupByClause == nil {
return true
}
// if there's a group by clause but no group by interval, we can't
// predict the cardinality of the columns used in the group by
// interval, thus we can't predict the number of responses returned
// from the shard
if querySpec.GetGroupByInterval() == nil {
return true
}
// if there's a group by time and other columns, then the previous
// logic holds
if len(groupByClause.Elems) > 1 {
return true
}
if !shards.ShouldAggregateLocally(querySpec) {
return true
}
for _, shard := range shards {
bufferSize := shard.QueryResponseBufferSize(querySpec, self.config.StoragePointBatchSize)
// if the number of repsonses is too big, do a sequential querying
if bufferSize > self.config.ClusterMaxResponseBufferSize {
return true
}
}
// parallel querying only if we're querying a single series, with
// group by time only
return false
}
func (self *Coordinator) runDropSeriesQuery(querySpec *parser.QuerySpec) error {
user := querySpec.User()
db := querySpec.Database()
series := querySpec.Query().DropSeriesQuery.GetTableName()
if ok, err := self.permissions.AuthorizeDropSeries(user, db, series); !ok {
return err
}
err := self.raftServer.DropSeries(db, series)
if err != nil {
return err
}
return nil
}
func (self *Coordinator) runListSeriesQuery(querySpec *parser.QuerySpec, p engine.Processor) error {
allSeries := self.clusterConfiguration.MetaStore.GetSeriesForDatabase(querySpec.Database())
matchingSeries := allSeries
q := querySpec.Query().GetListSeriesQuery()
if q.HasRegex() {
matchingSeries = nil
regex := q.GetRegex()
for _, s := range allSeries {
if !regex.MatchString(s) {
continue
}
matchingSeries = append(matchingSeries, s)
}
}
name := "list_series_result"
var fields []string
points := make([]*protocol.Point, len(matchingSeries))
if q.IncludeSpaces {
fields = []string{"name", "space"}
spaces := self.clusterConfiguration.GetShardSpacesForDatabase(querySpec.Database())
for i, s := range matchingSeries {
spaceName := ""
for _, sp := range spaces {
if sp.MatchesSeries(s) {
spaceName = sp.Name
break
}
}
fieldValues := []*protocol.FieldValue{
{StringValue: proto.String(s)},
{StringValue: proto.String(spaceName)},
}
points[i] = &protocol.Point{Values: fieldValues}
}
} else {
fields = []string{"name"}
for i, s := range matchingSeries {
fieldValues := []*protocol.FieldValue{
{StringValue: proto.String(s)},
}
points[i] = &protocol.Point{Values: fieldValues}
}
}
seriesResult := &protocol.Series{Name: &name, Fields: fields, Points: points}
_, err := p.Yield(seriesResult)
return err
}
// Returns true if we can aggregate the data locally per shard,
// i.e. the group by interval lines up with the shard duration and
// there are no joins or merges
func (self *ShardData) ShouldAggregateLocally(querySpec *parser.QuerySpec) bool {
f := querySpec.GetFromClause()
if f != nil && (f.Type == parser.FromClauseInnerJoin || f.Type == parser.FromClauseMerge) {
return false
}
groupByInterval := querySpec.GetGroupByInterval()
if groupByInterval == nil {
if querySpec.HasAggregates() {
return false
}
return true
}
return (self.shardDuration%*groupByInterval == 0) && !querySpec.GroupByIrregularInterval
}
func (self *ShardData) createRequest(querySpec *parser.QuerySpec) *p.Request {
queryString := querySpec.GetQueryStringWithTimeCondition()
user := querySpec.User()
userName := user.GetName()
database := querySpec.Database()
isDbUser := !user.IsClusterAdmin()
return &p.Request{
Type: p.Request_QUERY.Enum(),
ShardId: &self.id,
Query: &queryString,
UserName: &userName,
Database: &database,
IsDbUser: &isDbUser,
}
}
func (self *ShardData) logAndHandleDeleteQuery(querySpec *parser.QuerySpec, response chan<- *p.Response) {
queryString := querySpec.GetQueryStringWithTimeCondition()
request := self.createRequest(querySpec)
request.Query = &queryString
self.LogAndHandleDestructiveQuery(querySpec, request, response, false)
}
func (self *Shard) getPointIteratorForSeries(querySpec *parser.QuerySpec, name string, columns []string) ([]string, *PointIterator, error) {
fields, err := self.getFieldsForSeries(querySpec.Database(), name, columns)
if err != nil {
log.Error("Error looking up fields for %s: %s", name, err)
return nil, nil, err
}
startTime := querySpec.GetStartTime()
endTime := querySpec.GetEndTime()
query := querySpec.SelectQuery()
iterators := self.getIterators(fields, startTime, endTime, query.Ascending)
pi := NewPointIterator(iterators, fields, querySpec.GetStartTime(), querySpec.GetEndTime(), query.Ascending)
columns = make([]string, len(fields))
for i := range fields {
columns[i] = fields[i].Name
}
return columns, pi, nil
}
func (self *ShardData) Query(querySpec *parser.QuerySpec, response chan<- *p.Response) {
log.Debug("QUERY: shard %d, query '%s'", self.Id(), querySpec.GetQueryStringWithTimeCondition())
defer common.RecoverFunc(querySpec.Database(), querySpec.GetQueryStringWithTimeCondition(), func(err interface{}) {
response <- &p.Response{
Type: p.Response_ERROR.Enum(),
ErrorMessage: p.String(fmt.Sprintf("%s", err)),
}
})
// This is only for queries that are deletes or drops. They need to be sent everywhere as opposed to just the local or one of the remote shards.
// But this boolean should only be set to true on the server that receives the initial query.
if querySpec.RunAgainstAllServersInShard {
if querySpec.IsDeleteFromSeriesQuery() {
self.logAndHandleDeleteQuery(querySpec, response)
} else if querySpec.IsDropSeriesQuery() {
self.logAndHandleDropSeriesQuery(querySpec, response)
}
}
if self.IsLocal {
var processor engine.Processor = NewResponseChannelProcessor(NewResponseChannelWrapper(response))
var err error
processor = NewShardIdInserterProcessor(self.Id(), processor)
processor, err = self.getProcessor(querySpec, processor)
if err != nil {
response <- &p.Response{
Type: p.Response_ERROR.Enum(),
ErrorMessage: p.String(err.Error()),
}
log.Error("Error while creating engine: %s", err)
return
}
shard, err := self.store.GetOrCreateShard(self.id)
if err != nil {
response <- &p.Response{
Type: p.Response_ERROR.Enum(),
ErrorMessage: p.String(err.Error()),
}
log.Error("Error while getting shards: %s", err)
return
}
defer self.store.ReturnShard(self.id)
log.Debug("Processor chain: %s\n", engine.ProcessorChain(processor))
err = shard.Query(querySpec, processor)
// if we call Close() in case of an error it will mask the error
if err != nil {
response <- &p.Response{
Type: p.Response_ERROR.Enum(),
ErrorMessage: p.String(err.Error()),
}
return
}
processor.Close()
response <- &p.Response{Type: p.Response_END_STREAM.Enum()}
return
}
if server := self.randomHealthyServer(); server != nil {
log.Debug("Querying server %d for shard %d", server.GetId(), self.Id())
request := self.createRequest(querySpec)
server.MakeRequest(request, response)
return
}
message := fmt.Sprintf("No servers up to query shard %d", self.id)
response <- &p.Response{
Type: p.Response_ERROR.Enum(),
ErrorMessage: &message,
}
log.Error(message)
}
