Window Types

By choosing its two parameters, the @(k, w) operator lets you build every one of the classic window types used in stream processing. Below is a comparison of the patterns on one shared source stream.

Source stream

The file data.txt - 12 consecutive integers:

$ seq 1 12 > data.txt

The source declaration - one record per second, one field:

DECLARE val INTEGER STREAM src, 1 TEXTFILE 'data.txt'

Tumbling window - non-overlapping windows

Hop equal to window size: k = w. Successive emitted windows do not overlap. Their phase follows from stamping by the window end, so an initial input prefix may precede the first complete window.

SELECT * STREAM tumbling FROM src@(4,4)

Output interval: 1s × 4 / 1 = 4s. The first two output records:

$ xqry --null -s tumbling
5 4 3 2
9 8 7 6

The first record has logical index 1 and spans source positions 1..4, which contain values 2..5. A positive width puts the newest value first. Sample 1 precedes the first complete window of this partition.

Use cases: aggregating samples over fixed time intervals (e.g. per-minute, per-hour).

Sliding window - overlapping windows

Hop smaller than window size: k < w. Every input element appears in several successive windows.

SELECT * STREAM sliding FROM src@(1,4)

Output interval: 1s × 1 / 1 = 1s. Output records:

$ xqry --null -s sliding
4 3 2 1
5 4 3 2
6 5 4 3
7 6 5 4
...

Use cases: moving averages, trend detection, FIR filters (as in the signal filter implementation).

Sampling - windows with gaps

Hop larger than window size: k > w. Some input elements are skipped.

SELECT * STREAM sampled FROM src@(3,1)

Output interval: 1s × 3 / 1 = 3s. Output records:

$ xqry --null -s sampled
1
4
7
10

Use cases: signal decimation, sample-rate reduction, diagnostics on every Nth measurement.

Mirrored window - reversed field order

A negative w value reverses the order of fields in the output record, while keeping the same window size.

SELECT * STREAM mirrored FROM src@(2,-2)

Output interval: 1s × 2 / 1 = 2s. Output records (fields in reversed order):

$ xqry --null -s mirrored
2 3
4 5
6 7
8 9
...

Compare this with src@(2,2), which would give 3 2, 5 4, 7 6 - the newest value first. A negative width preserves arrival order but does not change the phase of the partition into pairs. Phase alignment when reversing serialization is described in the serialization example.

Summary of patterns

Query execution plan

All four variants can be run at once by placing them in a single .rql file:

DECLARE val INTEGER STREAM src, 1 TEXTFILE 'data.txt'

SELECT * STREAM tumbling FROM src@(4,4)
SELECT * STREAM sliding  FROM src@(1,4)
SELECT * STREAM sampled  FROM src@(3,1)
SELECT * STREAM mirrored FROM src@(2,-2)

$ xretractor -c windows.rql -f -p -d > out.dot && dot -Tsvg out.dot -o out.svg

The query plan shows four independent branches originating from a shared src node. Each branch implements a different window type with no dependencies between them.

NOTE: The functionality described here is covered by the tests: agse1, agse2, agse3, Pattern6, described in the appendix Integration Tests.