Page tree
Skip to end of metadata
Go to start of metadata

The predicate window opens and closes the window regarding a start and optional an end condition.

It can simulate any other window.

The operator works as follows:

  • It first checks, if the maxWindowTime is reached. In this case all internal buffers for each group is cleared, where the first element is older than the given threashold.
  • After then, it checks, if closewindowafternoupdatesfor is set and closes all buffers where the last element has reached the buffer a time longer than the parameter.
  • The operator determines the group (partition) for the current input.
  • After that, if set, the clear condition is checked for the current group and the current input.
  • If the window for this group is already opened the next step is to check,
    • if the end condition is true. Then the operator creates an output. Typically, the whole window is written and the buffer is cleared. With the clear and advanceWhen condition, this behaviour can be changed.
    • if the end condition is false, the current element is added to the window and kept inside the operator.
  • If the window for this group is not opened, the start condition is checked. It the condition is true, the operator opens a new window and adds the current element to the window.

For the output there are different configurations:

  • samestarttime: Each element in the output will get the same starttime, i.e. the starttime from the first element
  • nesting: In Odysseus the output is typically a set of elements that are send one after the other. If this flag is set to true, the output of the window is a single list, with all elements from the window. This can be advantage, if the processing afterwards treats the elements together (e.g. in a MAP-Operation).  Samestarttime sets the time for each element in the list, the list get the union of all intervals inside the list.

So simulate some kind of sliding window, the following parameters are used:

  • AdvanceWhen: This condition checks, if the window should move, i.e. if elements in the current buffer (for the current group) should be removed. If this predicate evalutes to true, the next parameter is used to determine which number of elements the move of the window should be
  • AdvanceSize: This size tells the operator if cases of AdvanceWhen is true, how many elements should be removed from the start of the current window. If the value is below 0 or the current window has less elements that this value, the buffer is cleared.

Remark: Advance is only used, if an output is generated and will be used after the results are produced. To clear the buffer independent of an output, clear needs to be used.


  • start: The start condition for a predicate window. If the condition evaluates to true, the windows is opened until the end predicate evaluates to true (or if not given the start predicate evaluates to false). Note, that all elements that are not inside a window are send to ouput port 1
  • end: The end condition for a predicate window. The tuple for which this condition is evaluated to true is only part of the result, if keepEndingElement is set to true!
  • clear: If this parameter is set, the window will only be cleared, if the condition is true. By this, the same element can be part of multiple windows (sliding)
  • sameStartTime: For predicate windows: If set to true, all produced elements get the same start timestamp
  • size: The maximum size of the window. Can be either a single number or a pair of a number and a time unit. Possible values for the unit are one of TimeUnit like SECONDS, NANOSECODS etc. - default time is the base time of the stream (typically milliseconds)
  • keepEndingElement: Typically, the object that fulfils the end condition will not be part of the result. If setting this attribute to true, the element will be part

  • partition: Evaluate the predicates on partitioned defined by different values of this attribute (similar to group by in aggreation)

  • useElementOnlyForStartOrEnd: Typically, an object is only used to evaluate the start or the end condition. If this value is set to true, an element can be used for both and can be part of multiple windows.
  • keepTimeOrder: If set to false, the output could be out of order.
  • closeWindowWithHeartbeat: if true, the window is closed when a heartbeat is received. Take a look at the session window to see how it works.
  • closewindowafternoupdatesfor: A time parameter by which the window could be closed if some time no new element reaches the buffer. Mostly makes sense for partioned windows but works also with heartbeats.
  • closeWindowAfterNoUpdateTimePort: In cases, the window closes because of closeWindowAfterNoUpdateTime this port is used for the output. Default is 0, i.e. the default output port. This can be used to handle outputs of this kind differently.

Parameters for MaxWindowTime

  • maxWindowTime: The maximum possible age of a window. If reached, the current window is closed.
  • outputIfMaxWindowTime: A window can close by condition or when maxWindowTime is reached. Set to false to avoid writing in case of maxWindowTime (default is true)
  • maxWindowTimeOutputPort: A special output port can be defined to allow to write in cases where maxWindowTime is reached to this port. Default is 0, i.e. the default output port.

Remark: This is a blocking operator. The operator does not write elements before it sees new elements not belonging to the window anymore (similiar to ElementWindow)


In the following we provide some examples and the corresponding output.

As input, we assume the following simple input:

ID	Time	isLast
A    1    false
A    2    false
A    3    false
A    4    true
B    5    false
B    6    false
B    7    false
B    8    false
B    9    false
B    10    true
C    11    false
C    12    false
C    13    false
C    14    false
C    15    false
C    16    false


With some preprocessing

in = CSVFILESOURCE({SCHEMA = [['ID', 'String'],['pos','STARTTIMESTAMP'],['isLast','Boolean']], DELIMITER = '\t', SOURCE = 'source', FILENAME = '${PROJECTPATH}/input.csv'})

map = STATEMAP({EXPRESSIONS = [['isNull(__last_1.ID) OR (__last_1.ID != ID)','newElem']], KEEPINPUT = true}, in)

we will get:

A|1|false|true | META | 1|oo
A|2|false|false | META | 2|oo
A|3|false|false | META | 3|oo
A|4|true|false | META | 4|oo
B|5|false|true | META | 5|oo
B|6|false|false | META | 6|oo
B|7|false|false | META | 7|oo
B|8|false|false | META | 8|oo
B|9|false|false | META | 9|oo
B|10|true|false | META | 10|oo
C|11|false|true | META | 11|oo
C|12|false|false | META | 12|oo
C|13|false|false | META | 13|oo
C|14|false|false | META | 14|oo
C|15|false|false | META | 15|oo
C|16|false|false | META | 16|oo

Using only a start predicate

win = PREDICATEWINDOW({start = 'newElem', SAMESTARTTIME = true}, map)

will result in:

A|1|false|true | META | 1|2
B|5|false|true | META | 5|6
C|11|false|true | META | 11|12

Here the window is opened for every true evaluation of the start condition and is closed for every evaluation of !start. All elements between these elements are discarded. They do not open a new window.

Using a start and an end predicate

win = PREDICATEWINDOW({start = 'newElem', end = 'newElem', SAMESTARTTIME = true}, map)
A|1|false|true | META | 1|5
A|2|false|false | META | 1|5
A|3|false|false | META | 1|5
A|4|true|false | META | 1|5
B|5|false|true | META | 5|11
B|6|false|false | META | 5|11
B|7|false|false | META | 5|11
B|8|false|false | META | 5|11
B|9|false|false | META | 5|11
B|10|true|false | META | 5|11
C|11|false|true | META | 11|17
C|12|false|false | META | 11|17
C|13|false|false | META | 11|17
C|14|false|false | META | 11|17
C|15|false|false | META | 11|17
C|16|false|false | META | 11|17

Here each time a new window opens, the old window is closed, i.e. the same input element is responsible for starting and closing a window.

Using a start and an end predicate and keeping the ending element:

win = PREDICATEWINDOW({start = 'newElem', end = 'isLast', KEEPENDINGELEMENT = true, SAMESTARTTIME = true}, map)

will result in:

A|1|false|true | META | 1|4
A|2|false|false | META | 1|4
A|3|false|false | META | 1|4
A|4|true|false | META | 1|4
B|5|false|true | META | 5|10
B|6|false|false | META | 5|10
B|7|false|false | META | 5|10
B|8|false|false | META | 5|10
B|9|false|false | META | 5|10
B|10|true|false | META | 5|10
C|11|false|true | META | 11|17
C|12|false|false | META | 11|17
C|13|false|false | META | 11|17
C|14|false|false | META | 11|17
C|15|false|false | META | 11|17
C|16|false|false | META | 11|17

Remark the difference: This operator blocks only until the end predicate is reached. This works only, if samestarttime is set to true, else e.g. A|4|true|false | META | 1|4 would be A|4|true|false | META | 4|4, this has no validitiy and will not be produced.

Using predicate window to simulate standard windows

The predicate window can be used to simulate other windows. Remark: This is only for demonstration purposes as the element and time window are much faster!

All the following examples use this source:

timer = TIMER({
            period = 1000,
            source = 'timer'

ticker := MAP({
              expressions = [['counter()','tick'],['TimeInterval.START','TS']]

Element Window

#INCLUDE ${PROJECTPATH}/TickerSource.qry

/// Tumbling element window
		  /// Start window with every element
          start = "true",
          /// close window, when size of buffer is 5
          end = "size(__all)==5",
		  /// output as list
          nesting = true
#INCLUDE ${PROJECTPATH}/TickerSource.qry

/// Sliding element window
		  /// start window with every element
          start = "true",
          /// If set to false, end element is not part of result
          KEEPENDINGELEMENT = true,
		  /// end predicate is tested before element is added to window
		  /// thats why size(__all) must be one below window size!
          end = "size(__all)=2",
          /// Move window when size of window is 3
          ADVANCEWHEN = 'size(__all)=3',
          /// move by one position
          ADVANCESIZE = 1,
          /// output as list
          nesting = true

Time Window

#INCLUDE ${PROJECTPATH}/TickerSource.qry

/// Tumbling time window
        	start = "true",
        	end = "!isNull(__first.TS) && __first.TS + 10 < TS",
        	nesting = true
  • No labels