我的功能存在性能问题，我想了解原因。我不习惯阅读EXPLAIN结果，所以我要求你的建议。
我创建了一个根据搜索条件返回id的函数：
scientific.selectChannels(_measurelabel := '{O3,SO2,NO2}', _siteid:='{41R001}', _networkid := '{RTU}')
它返回一个数组，运行时间不到60毫秒：
'{120,122,125}'
当我执行此查询时，答案非常缓慢：
WITH
D AS (
SELECT
DISTINCT ON (ChannelId)
ChannelId
,TimeValue
,FloatValue
FROM
datastore.inline
WHERE
-- ANY('{120,122,125}')
channelId = ANY(scientific.selectChannels(_measurelabel := '{O3,SO2,NO2}', _siteid:='{41R001}', _networkid := '{RTU}'))
AND QualityCodeId = 6
AND GranulityId = 1
AND TimeValue >= date_trunc('day', now()::TIMESTAMP) AND TimeValue < date_trunc('day', now()::TIMESTAMP) + '1 day'::INTERVAL
ORDER BY
ChannelId
,TimeValue DESC
)
SELECT
D.*
,CM.ChannelUnitsId
,CM.MeasureLabel
,CM.SiteId
,CM.SiteLabel
FROM
D JOIN scientific.channelMetadata AS CM ON (D.ChannelId = CM.ChannelId)
查询刨床返回：
"Hash Join (cost=12320.23..12331.54 rows=3 width=332) (actual time=9150.405..9150.463 rows=3 loops=1)"
" Hash Cond: (s.channelid = d.channelid)"
" CTE d"
" -> Unique (cost=12191.27..12192.55 rows=1 width=20) (actual time=9146.634..9146.647 rows=3 loops=1)"
" -> Sort (cost=12191.27..12191.91 rows=256 width=20) (actual time=9146.633..9146.644 rows=108 loops=1)"
" Sort Key: inline.channelid, inline.timevalue DESC"
" Sort Method: quicksort Memory: 33kB"
" -> Bitmap Heap Scan on inline (cost=197.79..12181.03 rows=256 width=20) (actual time=6729.253..9146.572 rows=108 loops=1)"
" Recheck Cond: ((timevalue >= date_trunc('day'::text, (now())::timestamp without time zone)) AND (timevalue < (date_trunc('day'::text, (now())::timestamp without time zone) + '1 day'::interval)))"
" Filter: ((qualitycodeid = 6) AND (granulityid = 1) AND (channelid = ANY (scientific.selectchannels('{41R001}'::text[], '{}'::text[], '{O3,SO2,NO2}'::text[], NULL::text[], '{RTU}'::text[], NULL::text[], NULL::text[], false))))"
" Rows Removed by Filter: 6132"
" Heap Blocks: exact=104"
" -> Bitmap Index Scan on idx_inline_timevalue (cost=0.00..197.73 rows=5729 width=0) (actual time=1.016..1.016 rows=6240 loops=1)"
" Index Cond: ((timevalue >= date_trunc('day'::text, (now())::timestamp without time zone)) AND (timevalue < (date_trunc('day'::text, (now())::timestamp without time zone) + '1 day'::interval)))"
" -> Sort (cost=127.65..129.39 rows=694 width=450) (actual time=3.714..3.740 rows=694 loops=1)"
" Sort Key: s.siteid, s.networkid, s.measureid, s.jokerid"
" Sort Method: quicksort Memory: 118kB"
" CTE s"
" -> Hash Join (cost=14.44..81.02 rows=694 width=182) (actual time=0.117..1.219 rows=694 loops=1)"
" Hash Cond: ((c.unitsid)::text = (u.id)::text)"
" -> Hash Join (cost=12.79..66.35 rows=694 width=150) (actual time=0.098..0.866 rows=694 loops=1)"
" Hash Cond: ((c.networkid)::text = (n.id)::text)"
" -> Hash Join (cost=11.43..55.45 rows=694 width=137) (actual time=0.088..0.644 rows=694 loops=1)"
" Hash Cond: ((c.siteid)::text = (s_1.id)::text)"
" -> Hash Join (cost=5.15..39.63 rows=694 width=122) (actual time=0.052..0.406 rows=694 loops=1)"
" Hash Cond: ((c.measureid)::text = (m.id)::text)"
" -> Seq Scan on channel c (cost=0.00..24.94 rows=694 width=97) (actual time=0.001..0.164 rows=694 loops=1)"
" -> Hash (cost=3.40..3.40 rows=140 width=31) (actual time=0.047..0.047 rows=140 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 16kB"
" -> Seq Scan on measure m (cost=0.00..3.40 rows=140 width=31) (actual time=0.002..0.023 rows=140 loops=1)"
" -> Hash (cost=3.90..3.90 rows=190 width=22) (actual time=0.032..0.032 rows=101 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 14kB"
" -> Seq Scan on site s_1 (cost=0.00..3.90 rows=190 width=22) (actual time=0.002..0.014 rows=101 loops=1)"
" -> Hash (cost=1.16..1.16 rows=16 width=17) (actual time=0.007..0.007 rows=18 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 9kB"
" -> Seq Scan on network n (cost=0.00..1.16 rows=16 width=17) (actual time=0.001..0.002 rows=18 loops=1)"
" -> Hash (cost=1.29..1.29 rows=29 width=36) (actual time=0.013..0.013 rows=29 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 10kB"
" -> Seq Scan on units u (cost=0.00..1.29 rows=29 width=36) (actual time=0.003..0.009 rows=29 loops=1)"
" -> CTE Scan on s (cost=0.00..13.88 rows=694 width=450) (actual time=0.119..1.839 rows=694 loops=1)"
" -> Hash (cost=0.02..0.02 rows=1 width=20) (actual time=9146.650..9146.650 rows=3 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 9kB"
" -> CTE Scan on d (cost=0.00..0.02 rows=1 width=20) (actual time=9146.636..9146.649 rows=3 loops=1)"
"Planning time: 1.931 ms"
"Execution time: 9150.594 ms"
需要大约10秒才能获取3行。
如果我用硬编码的结果'{120,122,125}'替换函数调用，则需要不到100毫秒。查询刨床然后返回：
"Hash Join (cost=376.05..387.36 rows=3 width=332) (actual time=7.483..7.595 rows=3 loops=1)"
" Hash Cond: (s.channelid = d.channelid)"
" CTE d"
" -> Unique (cost=247.94..248.37 rows=1 width=20) (actual time=0.172..0.194 rows=3 loops=1)"
" -> Sort (cost=247.94..248.15 rows=87 width=20) (actual time=0.172..0.182 rows=108 loops=1)"
" Sort Key: inline.channelid, inline.timevalue DESC"
" Sort Method: quicksort Memory: 33kB"
" -> Index Scan using uq_inline on inline (cost=0.44..245.13 rows=87 width=20) (actual time=0.036..0.109 rows=108 loops=1)"
" Index Cond: ((qualitycodeid = 6) AND (channelid = ANY ('{120,122,125}'::integer[])) AND (granulityid = 1) AND (timevalue >= date_trunc('day'::text, (now())::timestamp without time zone)) AND (timevalue < (date_trunc('day'::text, (now( (...)"
" -> Sort (cost=127.65..129.39 rows=694 width=450) (actual time=7.215..7.269 rows=694 loops=1)"
" Sort Key: s.siteid, s.networkid, s.measureid, s.jokerid"
" Sort Method: quicksort Memory: 118kB"
" CTE s"
" -> Hash Join (cost=14.44..81.02 rows=694 width=182) (actual time=0.225..2.351 rows=694 loops=1)"
" Hash Cond: ((c.unitsid)::text = (u.id)::text)"
" -> Hash Join (cost=12.79..66.35 rows=694 width=150) (actual time=0.190..1.661 rows=694 loops=1)"
" Hash Cond: ((c.networkid)::text = (n.id)::text)"
" -> Hash Join (cost=11.43..55.45 rows=694 width=137) (actual time=0.173..1.279 rows=694 loops=1)"
" Hash Cond: ((c.siteid)::text = (s_1.id)::text)"
" -> Hash Join (cost=5.15..39.63 rows=694 width=122) (actual time=0.106..0.817 rows=694 loops=1)"
" Hash Cond: ((c.measureid)::text = (m.id)::text)"
" -> Seq Scan on channel c (cost=0.00..24.94 rows=694 width=97) (actual time=0.002..0.306 rows=694 loops=1)"
" -> Hash (cost=3.40..3.40 rows=140 width=31) (actual time=0.096..0.096 rows=140 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 16kB"
" -> Seq Scan on measure m (cost=0.00..3.40 rows=140 width=31) (actual time=0.005..0.040 rows=140 loops=1)"
" -> Hash (cost=3.90..3.90 rows=190 width=22) (actual time=0.063..0.063 rows=101 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 14kB"
" -> Seq Scan on site s_1 (cost=0.00..3.90 rows=190 width=22) (actual time=0.005..0.037 rows=101 loops=1)"
" -> Hash (cost=1.16..1.16 rows=16 width=17) (actual time=0.011..0.011 rows=18 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 9kB"
" -> Seq Scan on network n (cost=0.00..1.16 rows=16 width=17) (actual time=0.002..0.003 rows=18 loops=1)"
" -> Hash (cost=1.29..1.29 rows=29 width=36) (actual time=0.027..0.027 rows=29 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 10kB"
" -> Seq Scan on units u (cost=0.00..1.29 rows=29 width=36) (actual time=0.007..0.016 rows=29 loops=1)"
" -> CTE Scan on s (cost=0.00..13.88 rows=694 width=450) (actual time=0.228..3.590 rows=694 loops=1)"
" -> Hash (cost=0.02..0.02 rows=1 width=20) (actual time=0.202..0.202 rows=3 loops=1)"
" Buckets: 1024 Batches: 1 Memory Usage: 9kB"
" -> CTE Scan on d (cost=0.00..0.02 rows=1 width=20) (actual time=0.175..0.199 rows=3 loops=1)"
"Planning time: 1.676 ms"
"Execution time: 7.734 ms"
有人可以解释为什么我的功能挂起了这个过程吗？ 功能定义如下：
CREATE OR REPLACE FUNCTION scientific.selectChannels(
_siteid TEXT[] = '{%}',
_measureid TEXT[] = '{}',
_measurelabel TEXT[] = '{%}',
_jokerid TEXT[] = NULL,
_networkid TEXT[] = NULL,
_sitetype TEXT[] = NULL,
_tags TEXT[] = NULL,
_inactive BOOLEAN = FALSE
)
RETURNS INTEGER[] AS
$selectChannels$
SELECT
array(
SELECT
CM.ChannelId
FROM
scientific.ChannelMetadata AS CM
WHERE
(CM.SiteId ILIKE ANY(_siteid))
AND ((CM.MeasureId ILIKE ANY(_measureid)) OR (CM.MeasureLabel LIKE ANY(_measurelabel)))
AND ((CM.JokerId ILIKE ANY(_jokerid)) OR (_jokerid IS NULL))
AND (CM.NetworkId ILIKE ANY(_networkid) OR (_networkid IS NULL))
AND ((CM.SiteType ILIKE ANY(_sitetype)) OR (_sitetype IS NULL) OR (CM.SiteType IS NULL))
AND ((array_lowercase(_tags) && array_lowercase(CM.Tags)) OR (_tags IS NULL) OR (CM.Tags IS NULL))
AND (_inactive OR CM.ActiveFlag)
ORDER BY
-- Do not alter this ORDER BY clause (Python API relies on this invariant: channels are ordered by id):
CM.ChannelId
);
$selectChannels$
LANGUAGE SQL;
其中channelMetada是合并5个表的VIEW。
我可以在第一次运行中看到rows removed（我不知道如何解释这个消息），但它可能是答案的一部分。
我认为唯一的事情是，我的函数被调用了几次而不是被计算一次，成为我查询的瓶颈。
我不知道如何解决这个问题。
问题是：
更新
如果我提取数组并在查询中重新创建它，那么从远处开始表现会更好。 为什么我的功能签名会减慢流程？
[...]
channelId = ANY(array(SELECT scientific.selectChannels(_measurelabel := '{O3,SO2,NO2}', _siteid:='{41R001}', _networkid := '{RTU}')))
[...]
答案 0 :(得分：4)
我的函数被多次调用了吗？
是的。
为CTE返回的108行中的每一行调用它。实际上你很幸运，TimeValue列上的条件可以通过索引来解决，否则你的函数将被调用一次表inline中的6240行中的每一行，而不仅仅是108由where条件的其他部分选择。
由于您的函数不依赖于从表中检索到的任何值，因此您可以通过将结果放入另一个CTE来改进此功能：
WITH func_result (channels) AS (
select scientific.selectChannels(_measurelabel := '{O3,SO2,NO2}', _siteid:='{41R001}', _networkid := '{RTU}')
), D AS (
SELECT DISTINCT ON (ChannelId) ChannelId ,TimeValue, FloatValue
FROM datastore.inline
WHERE
channelId = ANY( (select channels from func_result)::int[] )
AND QualityCodeId = 6
AND GranulityId = 1
AND TimeValue >= date_trunc('day', now()::TIMESTAMP) AND TimeValue < date_trunc('day', now()::TIMESTAMP) + '1 day'::INTERVAL
ORDER BY ChannelId, TimeValue DESC
)
SELECT
D.*
,CM.ChannelUnitsId
,CM.MeasureLabel
,CM.SiteId
,CM.SiteLabel
FROM
D JOIN scientific.channelMetadata AS CM ON (D.ChannelId = CM.ChannelId)
而不是
FROM datastore.inline
WHERE channelId = ANY( (select channels from func_result)::int[] )
你也可以使用：
FROM datastore.inline
JOIN func_result f on channelid = any(f.channels)
WHERE QualityCodeId = 6
AND GranulityId = 1
AND TimeValue ...
这应该导致函数只被调用一次。
从理论上讲，你可以通过将函数定义为immutable来欺骗Postgres只调用一次函数。但这将是一个公然的谎言（并且是一个非常肮脏的技巧），因为该函数确实执行数据库查找，并且执行该操作的函数永远不应该被定义为immutable。