使用time_bucket的慢时间刻度数据库查询

Question

我有一个简单的数据库(PostgreSQL 11(，里面装满了数百万个数据。我想获得每天的平均value。为此，我正在使用time_bucket()功能。

数据库架构

-- create database tables + indexes
CREATE TABLE IF NOT EXISTS machine (
    id   SMALLSERIAL PRIMARY KEY,
    name TEXT UNIQUE
);
CREATE TABLE IF NOT EXISTS reject_rate (
    time       TIMESTAMPTZ UNIQUE NOT NULL,
    machine_id SMALLINT REFERENCES machine(id) ON DELETE CASCADE,
    value      FLOAT NOT NULL,
    PRIMARY KEY(time, machine_id)
);
CREATE INDEX ON reject_rate (machine_id, value, time DESC);
-- hypertable
SELECT create_hypertable('reject_rate', 'time');
-- generate data with 54M rows
-- value column is generated randomly
-- this tooks minutes to finish but that's OK
INSERT INTO machine (name) VALUES ('machine1'), ('machine2');
INSERT INTO reject_rate (time, machine_id, value)
   SELECT to_timestamp(generate_series(1, 54e6)), 1, random();

我正在尝试做的查询是：

查询

SELECT
time_bucket('1 day', reject_rate.time) AS day,
AVG(value)
FROM reject_rate
GROUP BY day

结果 + 解释

即使我使用索引，查询的运行时间也非常慢。查询返回 626 行，需要 26.5 秒才能完成。已经创建了 90 个 TimescaleDB 块。下面是此查询的 EXPLAIN 语句：

"GroupAggregate  (cost=41.17..5095005.10 rows=54000000 width=16)"
"  Group Key: (time_bucket('1 day'::interval, _hyper_120_45_chunk."time"))"
"  ->  Result  (cost=41.17..4015005.10 rows=54000000 width=16)"
"        ->  Merge Append  (cost=41.17..3340005.10 rows=54000000 width=16)"
"              Sort Key: (time_bucket('1 day'::interval, _hyper_120_45_chunk."time"))"
"              ->  Index Scan using "45_86_reject_rate_time_key" on _hyper_120_45_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
"              ->  Index Scan using "50_96_reject_rate_time_key" on _hyper_120_50_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
"              ->  Index Scan using "55_106_reject_rate_time_key" on _hyper_120_55_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
"              ->  Index Scan using "60_116_reject_rate_time_key" on _hyper_120_60_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
"              ->  Index Scan using "65_126_reject_rate_time_key" on _hyper_120_65_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
"              ->  Index Scan using "70_136_reject_rate_time_key" on _hyper_120_70_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
"              ->  Index Scan using "75_146_reject_rate_time_key" on _hyper_120_75_chunk  (cost=0.42..14752.62 rows=604800 width=16)"
+ ~80 another rows of Index scan

问题

我是否正确创建了索引？我是否正确创建了数据库？还是对于如此多的行，TimescaleDB 很慢？

这可能就是time_bucket()慢的原因：https://github.com/timescale/timescaledb/issues/1229。建议的解决方案是使用连续聚合视图。这是如何在PostgreSQL中使用时间序列的推荐方法吗？

Answer 1

我很好奇内置的 Postgres 函数需要多长时间：

SELECT date_trunc('day', rr.time) as day,
       AVG(value)
FROM reject_rate rr
GROUP BY date_trunc('day', rr.time);