MongoDB復合索引詳解


摘要: 對於MongoDB的多鍵查詢,創建復合索引可以有效提高性能。


什么是復合索引?

復合索引,即Compound Index,指的是將多個鍵組合到一起創建索引,這樣可以加速匹配多個鍵的查詢。不妨通過一個簡單的示例理解復合索引。

students集合如下:

db.students.find().pretty()
{
"_id" : ObjectId("5aa7390ca5be7272a99b042a"),
"name" : "zhang",
"age" : "15"
}
{
"_id" : ObjectId("5aa7393ba5be7272a99b042b"),
"name" : "wang",
"age" : "15"
}
{
"_id" : ObjectId("5aa7393ba5be7272a99b042c"),
"name" : "zhang",
"age" : "14"
}

在name和age兩個鍵分別創建了索引(_id自帶索引):

db.students.getIndexes()
[
{
"v" : 1,
"key" : {
"name" : 1
},
"name" : "name_1",
"ns" : "test.students"
},
{
"v" : 1,
"key" : {
"age" : 1
},
"name" : "age_1",
"ns" : "test.students"
}
]

當進行多鍵查詢時,可以通過explian()分析執行情況(結果僅保留winningPlan):

db.students.find({name:"zhang",age:"14"}).explain()
"winningPlan":
{
"stage": "FETCH",
"filter":
{
"name":
{
"$eq": "zhang"
}
},
"inputStage":
{
"stage": "IXSCAN",
"keyPattern":
{
"age": 1
},
"indexName": "age_1",
"isMultiKey": false,
"isUnique": false,
"isSparse": false,
"isPartial": false,
"indexVersion": 1,
"direction": "forward",
"indexBounds":
{
"age": [
"[\"14\", \"14\"]"
]
}
}
}

由winningPlan可知,這個查詢依次分為IXSCANFETCH兩個階段。IXSCAN即索引掃描,使用的是age索引;FETCH即根據索引去查詢文檔,查詢的時候需要使用name進行過濾。

為name和age創建復合索引:

db.students.createIndex({name:1,age:1})

db.students.getIndexes()
[
{
"v" : 1,
"key" : {
"name" : 1,
"age" : 1
},
"name" : "name_1_age_1",
"ns" : "test.students"
}
]

有了復合索引之后,同一個查詢的執行方式就不同了:

db.students.find({name:"zhang",age:"14"}).explain()
"winningPlan":
{
"stage": "FETCH",
"inputStage":
{
"stage": "IXSCAN",
"keyPattern":
{
"name": 1,
"age": 1
},
"indexName": "name_1_age_1",
"isMultiKey": false,
"isUnique": false,
"isSparse": false,
"isPartial": false,
"indexVersion": 1,
"direction": "forward",
"indexBounds":
{
"name": [
"[\"zhang\", \"zhang\"]"
],
"age": [
"[\"14\", \"14\"]"
]
}
}
}

由winningPlan可知,這個查詢的順序沒有變化,依次分為IXSCANFETCH兩個階段。但是,IXSCAN使用的是name與age的復合索引;FETCH即根據索引去查詢文檔,不需要過濾。

這個示例的數據量太小,並不能看出什么問題。但是實際上,當數據量很大,IXSCAN返回的索引比較多時,FETCH時進行過濾將非常耗時。接下來將介紹一個真實的案例。

定位MongoDB性能問題

隨着接收的錯誤數據不斷增加,我們Fundebug已經累計處理3.5億錯誤事件,這給我們的服務不斷帶來性能方面的挑戰,尤其對於MongoDB集群來說。

對於生產數據庫,配置profile,可以記錄MongoDB的性能數據。執行以下命令,則所有超過1s的數據庫讀寫操作都會被記錄下來。

db.setProfilingLevel(1,1000)

查詢profile所記錄的數據,會發現events集合的某個查詢非常慢:

db.system.profile.find().pretty()
{
"op" : "command",
"ns" : "fundebug.events",
"command" : {
"count" : "events",
"query" : {
"createAt" : {
"$lt" : ISODate("2018-02-05T20:30:00.073Z")
},
"projectId" : ObjectId("58211791ea2640000c7a3fe6")
}
},
"keyUpdates" : 0,
"writeConflicts" : 0,
"numYield" : 1414,
"locks" : {
"Global" : {
"acquireCount" : {
"r" : NumberLong(2830)
}
},
"Database" : {
"acquireCount" : {
"r" : NumberLong(1415)
}
},
"Collection" : {
"acquireCount" : {
"r" : NumberLong(1415)
}
}
},
"responseLength" : 62,
"protocol" : "op_query",
"millis" : 28521,
"execStats" : {

},
"ts" : ISODate("2018-03-07T20:30:59.440Z"),
"client" : "192.168.59.226",
"allUsers" : [ ],
"user" : ""
}

events集合中有數億個文檔,因此count操作比較慢也不算太意外。根據profile數據,這個查詢耗時28.5s,時間長得有點離譜。另外,numYield高達1414,這應該就是操作如此之慢的直接原因。根據MongoDB文檔,numYield的含義是這樣的:

The number of times the operation yielded to allow other operations to complete. Typically, operations yield when they need access to data that MongoDB has not yet fully read into memory. This allows other operations that have data in memory to complete while MongoDB reads in data for the yielding operation.

這就意味着大量時間消耗在讀取硬盤上,且讀了非常多次。可以推測,應該是索引的問題導致的。

不妨使用explian()來分析一下這個查詢(僅保留executionStats):

db.events.explain("executionStats").count({"projectId" : ObjectId("58211791ea2640000c7a3fe6"),createAt:{"$lt" : ISODate("2018-02-05T20:30:00.073Z")}})
"executionStats":
{
"executionSuccess": true,
"nReturned": 20853,
"executionTimeMillis": 28055,
"totalKeysExamined": 28338,
"totalDocsExamined": 28338,
"executionStages":
{
"stage": "FETCH",
"filter":
{
"createAt":
{
"$lt": ISODate("2018-02-05T20:30:00.073Z")
}
},
"nReturned": 20853,
"executionTimeMillisEstimate": 27815,
"works": 28339,
"advanced": 20853,
"needTime": 7485,
"needYield": 0,
"saveState": 1387,
"restoreState": 1387,
"isEOF": 1,
"invalidates": 0,
"docsExamined": 28338,
"alreadyHasObj": 0,
"inputStage":
{
"stage": "IXSCAN",
"nReturned": 28338,
"executionTimeMillisEstimate": 30,
"works": 28339,
"advanced": 28338,
"needTime": 0,
"needYield": 0,
"saveState": 1387,
"restoreState": 1387,
"isEOF": 1,
"invalidates": 0,
"keyPattern":
{
"projectId": 1
},
"indexName": "projectId_1",
"isMultiKey": false,
"isUnique": false,
"isSparse": false,
"isPartial": false,
"indexVersion": 1,
"direction": "forward",
"indexBounds":
{
"projectId": [
"[ObjectId('58211791ea2640000c7a3fe6'), ObjectId('58211791ea2640000c7a3fe6')]"
]
},
"keysExamined": 28338,
"dupsTested": 0,
"dupsDropped": 0,
"seenInvalidated": 0
}
}
}

可知,events集合並沒有為projectId與createAt建立復合索引,因此IXSCAN階段采用的是projectId索引,其nReturned為28338; FETCH階段需要根據createAt進行過濾,其nReturned為20853,過濾掉了7485個文檔;另外,IXSCAN與FETCH階段的executionTimeMillisEstimate分別為30ms27815ms,因此基本上所有時間都消耗在了FETCH階段,這應該是讀取硬盤導致的。

創建復合索引

沒有為projectId和createAt創建復合索引是個尷尬的錯誤,趕緊補救一下:

db.events.createIndex({projectId:1,createTime:-1},{background: true})

在生產環境構建索引這種事最好是晚上做,這個命令一共花了大概7個小時吧!background設為true,指的是不要阻塞數據庫的其他操作,保證數據庫的可用性。但是,這個命令會一直占用着終端,這時不能使用CTRL + C,否則會終止索引構建過程。

復合索引創建成果之后,前文的查詢就快了很多(僅保留executionStats):

db.javascriptevents.explain("executionStats").count({"projectId" : ObjectId("58211791ea2640000c7a3fe6"),createAt:{"$lt" : ISODate("2018-02-05T20:30:00.073Z")}})
"executionStats":
{
"executionSuccess": true,
"nReturned": 0,
"executionTimeMillis": 47,
"totalKeysExamined": 20854,
"totalDocsExamined": 0,
"executionStages":
{
"stage": "COUNT",
"nReturned": 0,
"executionTimeMillisEstimate": 50,
"works": 20854,
"advanced": 0,
"needTime": 20853,
"needYield": 0,
"saveState": 162,
"restoreState": 162,
"isEOF": 1,
"invalidates": 0,
"nCounted": 20853,
"nSkipped": 0,
"inputStage":
{
"stage": "COUNT_SCAN",
"nReturned": 20853,
"executionTimeMillisEstimate": 50,
"works": 20854,
"advanced": 20853,
"needTime": 0,
"needYield": 0,
"saveState": 162,
"restoreState": 162,
"isEOF": 1,
"invalidates": 0,
"keysExamined": 20854,
"keyPattern":
{
"projectId": 1,
"createAt": -1
},
"indexName": "projectId_1_createTime_-1",
"isMultiKey": false,
"isUnique": false,
"isSparse": false,
"isPartial": false,
"indexVersion": 1
}
}
}

可知,count操作使用了projectId和createAt的復合索引,因此非常快,只花了46ms,性能提升了將近600倍!!!對比使用復合索引前后的結果,發現totalDocsExamined從28338降到了0,表示使用復合索引之后不再需要去查詢文檔,只需要掃描索引就好了,這樣就不需要去訪問磁盤了,自然快了很多。

參考


 

 


免責聲明!

本站轉載的文章為個人學習借鑒使用,本站對版權不負任何法律責任。如果侵犯了您的隱私權益,請聯系本站郵箱yoyou2525@163.com刪除。



 
粵ICP備18138465號   © 2018-2025 CODEPRJ.COM