timeSeriesLastTwoSamples

聚合函数接收时间序列数据，作为时间戳和值的对，并最多保存最近的 2 个样本。

参数：

timestamp - 样本的时间戳
value - 与 timestamp 对应的时间序列值此外，还可以将多个时间戳和值作为相同大小的数组传递。

返回值：一个 Tuple(Array(DateTime), Array(Float64)) - 一对长度为 0 到 2 的数组。第一个数组包含采样时间序列的时间戳，第二个数组包含对应的时间序列值。

示例：此聚合函数旨在与物化视图和聚合表结合使用，该表存储具有网格对齐时间戳的重新采样时间序列数据。考虑以下原始数据示例表，以及用于存储重新采样数据的表：

-- Table for raw data
CREATE TABLE t_raw_timeseries
(
    metric_id UInt64,
    timestamp DateTime64(3, 'UTC') CODEC(DoubleDelta, ZSTD),
    value Float64 CODEC(DoubleDelta)
)
ENGINE = MergeTree()
ORDER BY (metric_id, timestamp);

-- Table with data re-sampled to bigger (15 sec) time steps
CREATE TABLE t_resampled_timeseries_15_sec
(
    metric_id UInt64,
    grid_timestamp DateTime('UTC') CODEC(DoubleDelta, ZSTD), -- Timestamp aligned to 15 sec
    samples AggregateFunction(timeSeriesLastTwoSamples, DateTime64(3, 'UTC'), Float64)
)
ENGINE = AggregatingMergeTree()
ORDER BY (metric_id, grid_timestamp);

-- MV for populating re-sampled table
CREATE MATERIALIZED VIEW mv_resampled_timeseries TO t_resampled_timeseries_15_sec
(
    metric_id UInt64,
    grid_timestamp DateTime('UTC') CODEC(DoubleDelta, ZSTD),
    samples AggregateFunction(timeSeriesLastTwoSamples, DateTime64(3, 'UTC'), Float64)
)
AS SELECT
    metric_id,
    ceil(toUnixTimestamp(timestamp + interval 999 millisecond) / 15, 0) * 15 AS grid_timestamp,   -- Round timestamp up to the next grid point
    initializeAggregation('timeSeriesLastTwoSamplesState', timestamp, value) AS samples
FROM t_raw_timeseries
ORDER BY metric_id, grid_timestamp;

插入一些测试数据并读取 '2024-12-12 12:00:12' 和 '2024-12-12 12:00:30' 之间的数据

-- Insert some data
INSERT INTO t_raw_timeseries(metric_id, timestamp, value) SELECT number%10 AS metric_id, '2024-12-12 12:00:00'::DateTime64(3, 'UTC') + interval ((number/10)%100)*900 millisecond as timestamp, number%3+number%29 AS value FROM numbers(1000);

-- Check raw data
SELECT *
FROM t_raw_timeseries
WHERE metric_id = 3 AND timestamp BETWEEN '2024-12-12 12:00:12' AND '2024-12-12 12:00:31'
ORDER BY metric_id, timestamp;

3    2024-12-12 12:00:12.870    29
3    2024-12-12 12:00:13.770    8
3    2024-12-12 12:00:14.670    19
3    2024-12-12 12:00:15.570    30
3    2024-12-12 12:00:16.470    9
3    2024-12-12 12:00:17.370    20
3    2024-12-12 12:00:18.270    2
3    2024-12-12 12:00:19.170    10
3    2024-12-12 12:00:20.070    21
3    2024-12-12 12:00:20.970    3
3    2024-12-12 12:00:21.870    11
3    2024-12-12 12:00:22.770    22
3    2024-12-12 12:00:23.670    4
3    2024-12-12 12:00:24.570    12
3    2024-12-12 12:00:25.470    23
3    2024-12-12 12:00:26.370    5
3    2024-12-12 12:00:27.270    13
3    2024-12-12 12:00:28.170    24
3    2024-12-12 12:00:29.069    6
3    2024-12-12 12:00:29.969    14
3    2024-12-12 12:00:30.869    25

查询时间戳 '2024-12-12 12:00:15' 和 '2024-12-12 12:00:30' 的最后 2 个样本：

-- Check re-sampled data
SELECT metric_id, grid_timestamp, (finalizeAggregation(samples).1 as timestamp, finalizeAggregation(samples).2 as value) 
FROM t_resampled_timeseries_15_sec
WHERE metric_id = 3 AND grid_timestamp BETWEEN '2024-12-12 12:00:15' AND '2024-12-12 12:00:30'
ORDER BY metric_id, grid_timestamp;

3    2024-12-12 12:00:15    (['2024-12-12 12:00:14.670','2024-12-12 12:00:13.770'],[19,8])
3    2024-12-12 12:00:30    (['2024-12-12 12:00:29.969','2024-12-12 12:00:29.069'],[14,6])

聚合表仅存储每个 15 秒对齐时间戳的最后 2 个值。这使得通过读取比原始表中存储的数据少得多来计算类似 PromQL 的 irate 和 idelta 成为可能。

-- Calculate idelta and irate from the raw data
WITH
    '2024-12-12 12:00:15'::DateTime64(3,'UTC') AS start_ts,       -- start of timestamp grid
    start_ts + INTERVAL 60 SECOND AS end_ts,   -- end of timestamp grid
    15 AS step_seconds,   -- step of timestamp grid
    45 AS window_seconds  -- "staleness" window
SELECT
    metric_id,
    timeSeriesInstantDeltaToGrid(start_ts, end_ts, step_seconds, window_seconds)(timestamp, value),
    timeSeriesInstantRateToGrid(start_ts, end_ts, step_seconds, window_seconds)(timestamp, value)
FROM t_raw_timeseries
WHERE metric_id = 3 AND timestamp BETWEEN start_ts - interval window_seconds seconds AND end_ts
GROUP BY metric_id;

3    [11,8,-18,8,11]    [12.222222222222221,8.88888888888889,1.1111111111111112,8.88888888888889,12.222222222222221]

-- Calculate idelta and irate from the re-sampled data
WITH
    '2024-12-12 12:00:15'::DateTime64(3,'UTC') AS start_ts,       -- start of timestamp grid
    start_ts + INTERVAL 60 SECOND AS end_ts,   -- end of timestamp grid
    15 AS step_seconds,   -- step of timestamp grid
    45 AS window_seconds  -- "staleness" window
SELECT
    metric_id,
    timeSeriesInstantDeltaToGrid(start_ts, end_ts, step_seconds, window_seconds)(timestamps, values),
    timeSeriesInstantRateToGrid(start_ts, end_ts, step_seconds, window_seconds)(timestamps, values)
FROM (
    SELECT
        metric_id,
        finalizeAggregation(samples).1 AS timestamps,
        finalizeAggregation(samples).2 AS values
    FROM t_resampled_timeseries_15_sec
    WHERE metric_id = 3 AND grid_timestamp BETWEEN start_ts - interval window_seconds seconds AND end_ts
)
GROUP BY metric_id;

3    [11,8,-18,8,11]    [12.222222222222221,8.88888888888889,1.1111111111111112,8.88888888888889,12.222222222222221]

备注

此功能仍处于实验阶段，请通过设置 allow_experimental_ts_to_grid_aggregate_function=true 来启用它。