timeSeriesPredictLinearToGrid

Aggregate function that takes time series data as pairs of timestamps and values and calculates a PromQL-like linear prediction with a specified prediction timestamp offset from this data on a regular time grid described by start timestamp, end timestamp and step. For each point on the grid the samples for calculating predict_linear are considered within the specified time window.

Parameters:

• start timestamp - specifies start of the grid
• end timestamp - specifies end of the grid
• grid step - specifies step of the grid in seconds
• staleness - specifies the maximum "staleness" in seconds of the considered samples
• predict_offset - specifies number of seconds of offset to add to prediction time

Arguments:

• timestamp - timestamp of the sample
• value - value of the time series corresponding to the timestamp

Return value: predict_linear values on the specified grid as an Array(Nullable(Float64)). The returned array contains one value for each time grid point. The value is NULL if there are not enough samples within the window to calculate the rate value for a particular grid point.

Example: The following query calculates predict_linear values on the grid [90, 105, 120, 135, 150, 165, 180, 195, 210] with a 60 second offset:

WITH
    -- NOTE: the gap between 140 and 190 is to show how values are filled for ts = 150, 165, 180 according to window paramater
    [110, 120, 130, 140, 190, 200, 210, 220, 230]::Array(DateTime) AS timestamps,
    [1, 1, 3, 4, 5, 5, 8, 12, 13]::Array(Float32) AS values, -- array of values corresponding to timestamps above
    90 AS start_ts,       -- start of timestamp grid
    90 + 120 AS end_ts,   -- end of timestamp grid
    15 AS step_seconds,   -- step of timestamp grid
    45 AS window_seconds, -- "staleness" window
    60 AS predict_offset  -- prediction time offset
SELECT timeSeriesPredictLinearToGrid(start_ts, end_ts, step_seconds, window_seconds, predict_offset)(timestamp, value)
FROM
(
    -- This subquery converts arrays of timestamps and values into rows of `timestamp`, `value`
    SELECT
        arrayJoin(arrayZip(timestamps, values)) AS ts_and_val,
        ts_and_val.1 AS timestamp,
        ts_and_val.2 AS value
);

Response:

   ┌─timeSeriesPredictLinearToGrid(start_ts, end_ts, step_seconds, window_seconds, predict_offset)(timestamp, value)─┐
1. │ [NULL,NULL,1,9.166667,11.6,16.916666,NULL,NULL,16.5]                                                            │
   └─────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘

Also it is possible to pass multiple samples of timestamps and values as Arrays of equal size. The same query with array arguments:

WITH
    [110, 120, 130, 140, 190, 200, 210, 220, 230]::Array(DateTime) AS timestamps,
    [1, 1, 3, 4, 5, 5, 8, 12, 13]::Array(Float32) AS values,
    90 AS start_ts,
    90 + 120 AS end_ts,
    15 AS step_seconds,
    45 AS window_seconds,
    60 AS predict_offset
SELECT timeSeriesPredictLinearToGrid(start_ts, end_ts, step_seconds, window_seconds, predict_offset)(timestamps, values);

Note

This function is experimental, enable it by setting allow_experimental_ts_to_grid_aggregate_function=true.