window_op.go 7.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283
  1. package operators
  2. import (
  3. "fmt"
  4. "github.com/benbjohnson/clock"
  5. "github.com/emqx/kuiper/common"
  6. "github.com/emqx/kuiper/xsql"
  7. "github.com/emqx/kuiper/xstream/api"
  8. "github.com/emqx/kuiper/xstream/nodes"
  9. "math"
  10. "time"
  11. )
  12. type WindowConfig struct {
  13. Type xsql.WindowType
  14. Length int
  15. Interval int //If interval is not set, it is equals to Length
  16. }
  17. type WindowOperator struct {
  18. input chan interface{}
  19. outputs map[string]chan<- interface{}
  20. name string
  21. ticker *clock.Ticker //For processing time only
  22. window *WindowConfig
  23. interval int
  24. triggerTime int64
  25. isEventTime bool
  26. statManager nodes.StatManager
  27. watermarkGenerator *WatermarkGenerator //For event time only
  28. }
  29. func NewWindowOp(name string, w *xsql.Window, isEventTime bool, lateTolerance int64, streams []string, bufferLength int) (*WindowOperator, error) {
  30. o := new(WindowOperator)
  31. o.input = make(chan interface{}, bufferLength)
  32. o.outputs = make(map[string]chan<- interface{})
  33. o.name = name
  34. o.isEventTime = isEventTime
  35. if w != nil {
  36. o.window = &WindowConfig{
  37. Type: w.WindowType,
  38. Length: w.Length.Val,
  39. Interval: w.Interval.Val,
  40. }
  41. } else {
  42. o.window = &WindowConfig{
  43. Type: xsql.NOT_WINDOW,
  44. }
  45. }
  46. if isEventTime {
  47. //Create watermark generator
  48. if w, err := NewWatermarkGenerator(o.window, lateTolerance, streams, o.input); err != nil {
  49. return nil, err
  50. } else {
  51. o.watermarkGenerator = w
  52. }
  53. } else {
  54. switch o.window.Type {
  55. case xsql.NOT_WINDOW:
  56. case xsql.TUMBLING_WINDOW:
  57. o.ticker = common.GetTicker(o.window.Length)
  58. o.interval = o.window.Length
  59. case xsql.HOPPING_WINDOW:
  60. o.ticker = common.GetTicker(o.window.Interval)
  61. o.interval = o.window.Interval
  62. case xsql.SLIDING_WINDOW:
  63. o.interval = o.window.Length
  64. case xsql.SESSION_WINDOW:
  65. o.ticker = common.GetTicker(o.window.Length)
  66. o.interval = o.window.Interval
  67. default:
  68. return nil, fmt.Errorf("unsupported window type %d", o.window.Type)
  69. }
  70. }
  71. return o, nil
  72. }
  73. func (o *WindowOperator) GetName() string {
  74. return o.name
  75. }
  76. func (o *WindowOperator) AddOutput(output chan<- interface{}, name string) error {
  77. if _, ok := o.outputs[name]; !ok {
  78. o.outputs[name] = output
  79. } else {
  80. return fmt.Errorf("fail to add output %s, operator %s already has an output of the same name", name, o.name)
  81. }
  82. return nil
  83. }
  84. func (o *WindowOperator) GetInput() (chan<- interface{}, string) {
  85. return o.input, o.name
  86. }
  87. // Exec is the entry point for the executor
  88. // input: *xsql.Tuple from preprocessor
  89. // output: xsql.WindowTuplesSet
  90. func (o *WindowOperator) Exec(ctx api.StreamContext, errCh chan<- error) {
  91. log := ctx.GetLogger()
  92. log.Debugf("Window operator %s is started", o.name)
  93. if len(o.outputs) <= 0 {
  94. go func() { errCh <- fmt.Errorf("no output channel found") }()
  95. return
  96. }
  97. stats, err := nodes.NewStatManager("op", ctx)
  98. if err != nil {
  99. go func() { errCh <- err }()
  100. return
  101. }
  102. o.statManager = stats
  103. if o.isEventTime {
  104. go o.execEventWindow(ctx, errCh)
  105. } else {
  106. go o.execProcessingWindow(ctx, errCh)
  107. }
  108. }
  109. func (o *WindowOperator) execProcessingWindow(ctx api.StreamContext, errCh chan<- error) {
  110. log := ctx.GetLogger()
  111. var (
  112. inputs []*xsql.Tuple
  113. c <-chan time.Time
  114. timeoutTicker *clock.Timer
  115. timeout <-chan time.Time
  116. )
  117. if o.ticker != nil {
  118. c = o.ticker.C
  119. }
  120. for {
  121. select {
  122. // process incoming item
  123. case item, opened := <-o.input:
  124. o.statManager.IncTotalRecordsIn()
  125. o.statManager.ProcessTimeStart()
  126. if !opened {
  127. o.statManager.IncTotalExceptions()
  128. break
  129. }
  130. switch d := item.(type) {
  131. case error:
  132. nodes.Broadcast(o.outputs, d, ctx)
  133. o.statManager.IncTotalExceptions()
  134. case *xsql.Tuple:
  135. log.Debugf("Event window receive tuple %s", d.Message)
  136. inputs = append(inputs, d)
  137. switch o.window.Type {
  138. case xsql.NOT_WINDOW:
  139. inputs, _ = o.scan(inputs, d.Timestamp, ctx)
  140. case xsql.SLIDING_WINDOW:
  141. inputs, _ = o.scan(inputs, d.Timestamp, ctx)
  142. case xsql.SESSION_WINDOW:
  143. if timeoutTicker != nil {
  144. timeoutTicker.Stop()
  145. timeoutTicker.Reset(time.Duration(o.window.Interval) * time.Millisecond)
  146. } else {
  147. timeoutTicker = common.GetTimer(o.window.Interval)
  148. timeout = timeoutTicker.C
  149. }
  150. }
  151. o.statManager.ProcessTimeEnd()
  152. o.statManager.SetBufferLength(int64(len(o.input)))
  153. default:
  154. nodes.Broadcast(o.outputs, fmt.Errorf("run Window error: expect xsql.Tuple type but got %[1]T(%[1]v)", d), ctx)
  155. o.statManager.IncTotalExceptions()
  156. }
  157. case now := <-c:
  158. if len(inputs) > 0 {
  159. o.statManager.ProcessTimeStart()
  160. n := common.TimeToUnixMilli(now)
  161. //For session window, check if the last scan time is newer than the inputs
  162. if o.window.Type == xsql.SESSION_WINDOW {
  163. //scan time for session window will record all triggers of the ticker but not the timeout
  164. lastTriggerTime := o.triggerTime
  165. o.triggerTime = n
  166. //Check if the current window has exceeded the max duration, if not continue expand
  167. if lastTriggerTime < inputs[0].Timestamp {
  168. break
  169. }
  170. }
  171. log.Debugf("triggered by ticker")
  172. inputs, _ = o.scan(inputs, n, ctx)
  173. o.statManager.ProcessTimeEnd()
  174. }
  175. case now := <-timeout:
  176. if len(inputs) > 0 {
  177. o.statManager.ProcessTimeStart()
  178. log.Debugf("triggered by timeout")
  179. inputs, _ = o.scan(inputs, common.TimeToUnixMilli(now), ctx)
  180. //expire all inputs, so that when timer scan there is no item
  181. inputs = make([]*xsql.Tuple, 0)
  182. o.statManager.ProcessTimeEnd()
  183. }
  184. // is cancelling
  185. case <-ctx.Done():
  186. log.Infoln("Cancelling window....")
  187. if o.ticker != nil {
  188. o.ticker.Stop()
  189. }
  190. return
  191. }
  192. }
  193. }
  194. func (o *WindowOperator) scan(inputs []*xsql.Tuple, triggerTime int64, ctx api.StreamContext) ([]*xsql.Tuple, bool) {
  195. log := ctx.GetLogger()
  196. log.Debugf("window %s triggered at %s(%d)", o.name, time.Unix(triggerTime/1000, triggerTime%1000), triggerTime)
  197. var delta int64
  198. if o.window.Type == xsql.HOPPING_WINDOW || o.window.Type == xsql.SLIDING_WINDOW {
  199. delta = o.calDelta(triggerTime, delta, log)
  200. }
  201. var results xsql.WindowTuplesSet = make([]xsql.WindowTuples, 0)
  202. i := 0
  203. //Sync table
  204. for _, tuple := range inputs {
  205. if o.window.Type == xsql.HOPPING_WINDOW || o.window.Type == xsql.SLIDING_WINDOW {
  206. diff := o.triggerTime - tuple.Timestamp
  207. if diff > int64(o.window.Length)+delta {
  208. log.Debugf("diff: %d, length: %d, delta: %d", diff, o.window.Length, delta)
  209. log.Debugf("tuple %s emitted at %d expired", tuple, tuple.Timestamp)
  210. //Expired tuple, remove it by not adding back to inputs
  211. continue
  212. }
  213. //Added back all inputs for non expired events
  214. inputs[i] = tuple
  215. i++
  216. } else if tuple.Timestamp > triggerTime {
  217. //Only added back early arrived events
  218. inputs[i] = tuple
  219. i++
  220. }
  221. if tuple.Timestamp <= triggerTime {
  222. results = results.AddTuple(tuple)
  223. }
  224. }
  225. triggered := false
  226. if len(results) > 0 {
  227. log.Debugf("window %s triggered for %d tuples", o.name, len(inputs))
  228. if o.isEventTime {
  229. results.Sort()
  230. }
  231. log.Debugf("Sent: %v", results)
  232. //blocking if one of the channel is full
  233. nodes.Broadcast(o.outputs, results, ctx)
  234. triggered = true
  235. o.statManager.IncTotalRecordsOut()
  236. log.Debugf("done scan")
  237. }
  238. return inputs[:i], triggered
  239. }
  240. func (o *WindowOperator) calDelta(triggerTime int64, delta int64, log api.Logger) int64 {
  241. lastTriggerTime := o.triggerTime
  242. o.triggerTime = triggerTime
  243. if lastTriggerTime <= 0 {
  244. delta = math.MaxInt16 //max int, all events for the initial window
  245. } else {
  246. if !o.isEventTime && o.window.Interval > 0 {
  247. delta = o.triggerTime - lastTriggerTime - int64(o.window.Interval)
  248. if delta > 100 {
  249. log.Warnf("Possible long computation in window; Previous eviction time: %d, current eviction time: %d", lastTriggerTime, o.triggerTime)
  250. }
  251. } else {
  252. delta = 0
  253. }
  254. }
  255. return delta
  256. }
  257. func (o *WindowOperator) GetMetrics() [][]interface{} {
  258. if o.statManager != nil {
  259. return [][]interface{}{
  260. o.statManager.GetMetrics(),
  261. }
  262. } else {
  263. return nil
  264. }
  265. }