arangodb Facet计算/聚合缓慢吗？
[英]arangodb Facet calculation / aggregation slow?
问题描述
我不知道为什么以下方面计算如此缓慢：
FOR q IN LRQ
COLLECT profile = q.LongRunningQuery.Profile INTO profiles
RETURN { "Profile" : profile, "Count" : LENGTH(profiles)}
尽管大约只有5,000个文档在数据库中，并且结果只有30个不同的方面，但这大约需要30秒。
LongHunningQuery.Profile字段使用哈希索引和跳过列表索引进行索引。 （我也尝试过使用它们的不同组合）。
有没有人给我提示可能出什么问题？ 查询是否可能无法从索引中受益？ （5,000条记录的大小约为1 GB，因此我假设将不使用哈希索引，而可能使用全表扫描吗？）
有趣的是，以下替代方案仅持续2秒：
FOR q IN SKIPLIST(LRQ, { "LongRunningQuery.Profile": [ [ '>', '' ] ] })[*].LongRunningQuery.Profile
COLLECT profile = q INTO profiles
RETURN { "Profile" : profile, "Count" : LENGTH(profiles) }
但是对于如此少量的记录，它仍然需要2秒。 在这里看起来好像使用了跳过列表索引，但它可能不是完美的索引变体。
更新2014-11-27：
arangosh [_system]> stmt._query
FOR q IN LRQ COLLECT profile = q.LongRunningQuery.Profile INTO profiles RETURN {
"Profile" : profile, "Count" : LENGTH(profiles)}
arangosh [_system]> db.LRQ.ensureHashIndex("LongRunningQuery.Profile");
{
"id" : "LRQ/296017913256",
"type" : "hash",
"unique" : false,
"fields" : [
"LongRunningQuery.Profile"
],
"isNewlyCreated" : false,
"error" : false,
"code" : 200
}
查询耗时约32秒，返回31个简短结果。
执行计划：
{
"plan": {
"nodes": [
{
"type": "SingletonNode",
"dependencies": [],
"id": 1,
"estimatedCost": 1,
"estimatedNrItems": 1
},
{
"type": "EnumerateCollectionNode",
"dependencies": [
1
],
"id": 2,
"estimatedCost": 5311,
"estimatedNrItems": 5310,
"database": "_system",
"collection": "LRQ",
"outVariable": {
"id": 0,
"name": "q"
}
},
{
"type": "CalculationNode",
"dependencies": [
2
],
"id": 3,
"estimatedCost": 10621,
"estimatedNrItems": 5310,
"expression": {
"type": "attribute access",
"name": "Profile",
"subNodes": [
{
"type": "attribute access",
"name": "LongRunningQuery",
"subNodes": [
{
"type": "reference",
"name": "q",
"id": 0
}
]
}
]
},
"outVariable": {
"id": 3,
"name": "3"
},
"canThrow": false
},
{
"type": "SortNode",
"dependencies": [
3
],
"id": 4,
"estimatedCost": 56166.713176593075,
"estimatedNrItems": 5310,
"elements": [
{
"inVariable": {
"id": 3,
"name": "3"
},
"ascending": true
}
],
"stable": true
},
{
"type": "AggregateNode",
"dependencies": [
4
],
"id": 5,
"estimatedCost": 61476.713176593075,
"estimatedNrItems": 5310,
"aggregates": [
{
"outVariable": {
"id": 1,
"name": "profile"
},
"inVariable": {
"id": 3,
"name": "3"
}
}
],
"outVariable": {
"id": 2,
"name": "profiles"
}
},
{
"type": "CalculationNode",
"dependencies": [
5
],
"id": 6,
"estimatedCost": 66786.71317659307,
"estimatedNrItems": 5310,
"expression": {
"type": "array",
"subNodes": [
{
"type": "array element",
"name": "Profile",
"subNodes": [
{
"type": "reference",
"name": "profile",
"id": 1
}
]
},
{
"type": "array element",
"name": "Count",
"subNodes": [
{
"type": "function call",
"name": "LENGTH",
"subNodes": [
{
"type": "list",
"subNodes": [
{
"type": "reference",
"name": "profiles",
"id": 2
}
]
}
]
}
]
}
]
},
"outVariable": {
"id": 4,
"name": "4"
},
"canThrow": false
},
{
"type": "ReturnNode",
"dependencies": [
6
],
"id": 7,
"estimatedCost": 72096.71317659307,
"estimatedNrItems": 5310,
"inVariable": {
"id": 4,
"name": "4"
}
}
],
"rules": [],
"collections": [
{
"name": "LRQ",
"type": "read"
}
],
"variables": [
{
"id": 0,
"name": "q"
},
{
"id": 1,
"name": "profile"
},
{
"id": 4,
"name": "4"
},
{
"id": 2,
"name": "profiles"
},
{
"id": 3,
"name": "3"
}
],
"estimatedCost": 72096.71317659307,
"estimatedNrItems": 5310
},
"warnings": []
}
2014年12月5日更新：
以下是其他措施：理解，谢谢。 这是输出：
AQL_EXECUTE（'FOR q LRQ FILTER q.LongRunningQuery.Profile ==“ Admin” LIMIT 1 RETURN q.LongRunningQuery.Profile'，{}，{profile：true}）。profile-> {“初始化”：0 ，“解析”：0，“优化ast”：15.364980936050415，“实例化计划”：0，“优化计划”：0，“执行中”：0}
执行AQL_EXECUTE（'FOR q IN LRQ COLLECT profile = q.LongRunningQuery.Profile INTO profile RETURN {“ Profile”：profile，“ Count”：LENGTH（profiles）}'，{}，{profile：true}）。profile- -> {“正在初始化”：0，“正在解析”：0，“优化ast”：0，“实例化计划”：0，“优化计划”：0，“正在执行”：77.88313102722168}
2014年12月19日更新：
从2.3.2版本开始，查询arangosh [_system]> stmt2 = db._createStatement（'FOR q IN LRQ COLLECT profile = q.LongRunningQuery.Profile INTO profile RETURN {“” Profile：“，” Count“：LENGTH（个人资料）}'）
看起来像这样：
arangosh [_system]> stmt2.explain()
{
"plan" : {
"nodes" : [
{
"type" : "SingletonNode",
"dependencies" : [ ],
"id" : 1,
"estimatedCost" : 1,
"estimatedNrItems" : 1
},
{
"type" : "IndexRangeNode",
"dependencies" : [
1
],
"id" : 8,
"estimatedCost" : 5311,
"estimatedNrItems" : 5310,
"database" : "_system",
"collection" : "LRQ",
"outVariable" : {
"id" : 0,
"name" : "q"
},
"ranges" : [
[ ]
],
"index" : {
"type" : "skiplist",
"id" : "530975525379",
"unique" : false,
"fields" : [
"LongRunningQuery.Profile"
]
},
"reverse" : false
},
{
"type" : "CalculationNode",
"dependencies" : [
8
],
"id" : 3,
"estimatedCost" : 10621,
"estimatedNrItems" : 5310,
"expression" : {
"type" : "attribute access",
"name" : "Profile",
"subNodes" : [
{
"type" : "attribute access",
"name" : "LongRunningQuery",
"subNodes" : [
{
"type" : "reference",
"name" : "q",
"id" : 0
}
]
}
]
},
"outVariable" : {
"id" : 3,
"name" : "3"
},
"canThrow" : false
},
{
"type" : "AggregateNode",
"dependencies" : [
3
],
"id" : 5,
"estimatedCost" : 15931,
"estimatedNrItems" : 5310,
"aggregates" : [
{
"outVariable" : {
"id" : 1,
"name" : "profile"
},
"inVariable" : {
"id" : 3,
"name" : "3"
}
}
],
"outVariable" : {
"id" : 2,
"name" : "profiles"
}
},
{
"type" : "CalculationNode",
"dependencies" : [
5
],
"id" : 6,
"estimatedCost" : 21241,
"estimatedNrItems" : 5310,
"expression" : {
"type" : "array",
"subNodes" : [
{
"type" : "array element",
"name" : "Profile",
"subNodes" : [
{
"type" : "reference",
"name" : "profile",
"id" : 1
}
]
},
{
"type" : "array element",
"name" : "Count",
"subNodes" : [
{
"type" : "function call",
"name" : "LENGTH",
"subNodes" : [
{
"type" : "list",
"subNodes" : [
{
"type" : "reference",
"name" : "profiles",
"id" : 2
}
]
}
]
}
]
}
]
},
"outVariable" : {
"id" : 4,
"name" : "4"
},
"canThrow" : false
},
{
"type" : "ReturnNode",
"dependencies" : [
6
],
"id" : 7,
"estimatedCost" : 26551,
"estimatedNrItems" : 5310,
"inVariable" : {
"id" : 4,
"name" : "4"
}
}
],
"rules" : [
"use-index-for-sort"
],
"collections" : [
{
"name" : "LRQ",
"type" : "read"
}
],
"variables" : [
{
"id" : 0,
"name" : "q"
},
{
"id" : 1,
"name" : "profile"
},
{
"id" : 4,
"name" : "4"
},
{
"id" : 2,
"name" : "profiles"
},
{
"id" : 3,
"name" : "3"
}
],
"estimatedCost" : 26551,
"estimatedNrItems" : 5310
},
"warnings" : [ ],
"stats" : {
"rulesExecuted" : 25,
"rulesSkipped" : 0,
"plansCreated" : 1
}
}
2 个解决方案
解决方案1
1 2014-12-04 12:52:37
嗯，看看说明中有一个sortnode，而您的查询没有提供排序？ 收集可能会使优化器无法使用索引（然后您将拥有一个IndexRangeNode而不是EnumerateCollectionNode）
如果传递查询的options参数（db._query（）的第4个参数）{profile：true}，它将输出各阶段使用的时间； 您可以重新运行您的查询，然后向我们显示回复吗？
解决方案2
0 2014-12-19 10:44:33
COLLECT语句需要排序的输入。 因此，即使原始查询字符串不包含显式SORT语句，也会将SORT语句自动添加到执行计划中。
这就是为什么SortNode出现在计划中的原因。 如果sort属性上有一个跳过列表索引（在本例中为LongRunningQuery.Profile ），将优化LongRunningQuery.Profile 。 因此，在属性上添加一个skiplist索引可以加快速度，因为可以节省（昂贵）的排序步骤。
如果已经设置了这样的索引并运行查询，则它应该比仅存在哈希索引时要快。 实际上，原始查询应该已经忽略了哈希索引。
如果已经设置了跳过列表索引并解释了该查询，则还应该看到不再有SortNode。
从ArangoDB 2.4（目前处于开发阶段）开始，添加了仅用于计算方面的更有效的语法：
FOR q IN LRQ
COLLECT profile = q.LongRunningQuery.Profile WITH COUNT INTO numProfiles
RETURN { "Profile" : profile, "Count" : numProfiles }
这样可以进一步加快查询速度。
使用AQL在arangodb中进行聚合
声明:本站的技术帖子网页，遵循CC BY-SA 4.0协议，如果您需要转载，请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.