MongoDB 性能与 PostgreSQL 的 550 万行/文档对比

Mik*_*e A 12 postgresql performance mongodb

有人可以帮我比较这些查询并解释为什么 PostgreSQL 查询的执行时间不到 2000 毫秒,而 MongoDB 聚合查询需要将近 9000 毫秒,有时甚至高达 130K 毫秒吗?

PostgreSQL 9.3.2 on x86_64-apple-darwin, compiled by i686-apple-darwin11-llvm-gcc-4.2 (GCC) 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.9.00), 64-bit
Run Code Online (Sandbox Code Playgroud)

PostgreSQL 查询

SELECT locomotive_id,
   SUM(date_trunc('second', datetime) - date_trunc('second', prevDatetime)) AS utilization_time

FROM bpkdmp 
WHERE datetime >= '2013-7-26 00:00:00.0000' 
AND   datetime <= '2013-7-26 23:59:59.9999'
GROUP BY locomotive_id
order by locomotive_id
Run Code Online (Sandbox Code Playgroud)

MongoDB 查询

db.bpkdmp.aggregate([
   {
      $match : {
          datetime : { $gte : new Date(2013,6,26, 0, 0, 0, 0), $lt : new Date(2013,6,26, 23, 59, 59, 9999) }
   }
   },
   {
      $project: {
         locomotive_id : "$locomotive_id",
         loco_time : { $subtract : ["$datetime", "$prevdatetime"] }, 
      }
   },
   {
      $group : {
         _id : "$locomotive_id",
         utilization_time : { $sum : "$loco_time" }
      }
   },
   {
      $sort : {_id : 1}
   }
])
Run Code Online (Sandbox Code Playgroud)

PostgreSQL 表和 MongoDB 集合都在 datetime : 1 和 locomotive_id : 1 上建立索引

这些查询正在带有 2TB 混合驱动器和 16GB 内存的 iMac 上进行测试。我在具有 8GB 内存和 256GB SSD 的 Windows 7 机器上收到了类似的结果。

谢谢!

** 更新:我在发布我的问题后发布了 EXPLAIN (BUFFERS, ANALYZE) 结果

"Sort  (cost=146036.84..146036.88 rows=19 width=24) (actual time=2182.443..2182.457 rows=152 loops=1)"
"  Sort Key: locomotive_id"
"  Sort Method: quicksort  Memory: 36kB"
"  Buffers: shared hit=13095"
"  ->  HashAggregate  (cost=146036.24..146036.43 rows=19 width=24) (actual time=2182.144..2182.360 rows=152 loops=1)"
"        Buffers: shared hit=13095"
"        ->  Bitmap Heap Scan on bpkdmp  (cost=12393.84..138736.97 rows=583942 width=24) (actual time=130.409..241.087 rows=559529 loops=1)"
"              Recheck Cond: ((datetime >= '2013-07-26 00:00:00'::timestamp without time zone) AND (datetime <= '2013-07-26 23:59:59.9999'::timestamp without time zone))"
"              Buffers: shared hit=13095"
"              ->  Bitmap Index Scan on bpkdmp_datetime_ix  (cost=0.00..12247.85 rows=583942 width=0) (actual time=127.707..127.707 rows=559529 loops=1)"
"                    Index Cond: ((datetime >= '2013-07-26 00:00:00'::timestamp without time zone) AND (datetime <= '2013-07-26 23:59:59.9999'::timestamp without time zone))"
"                    Buffers: shared hit=1531"
"Total runtime: 2182.620 ms"
Run Code Online (Sandbox Code Playgroud)

** 更新:Mongo 解释:

从MongoDB解释

{
"serverPipeline" : [
    {
        "query" : {
            "datetime" : {
                "$gte" : ISODate("2013-07-26T04:00:00Z"),
                "$lt" : ISODate("2013-07-27T04:00:08.999Z")
            }
        },
        "projection" : {
            "datetime" : 1,
            "locomotive_id" : 1,
            "prevdatetime" : 1,
            "_id" : 1
        },
        "cursor" : {
            "cursor" : "BtreeCursor datetime_1",
            "isMultiKey" : false,
            "n" : 559572,
            "nscannedObjects" : 559572,
            "nscanned" : 559572,
            "nscannedObjectsAllPlans" : 559572,
            "nscannedAllPlans" : 559572,
            "scanAndOrder" : false,
            "indexOnly" : false,
            "nYields" : 1,
            "nChunkSkips" : 0,
            "millis" : 988,
            "indexBounds" : {
                "datetime" : [
                    [
                        ISODate("2013-07-26T04:00:00Z"),
                        ISODate("2013-07-27T04:00:08.999Z")
                    ]
                ]
            },
            "allPlans" : [
                {
                    "cursor" : "BtreeCursor datetime_1",
                    "n" : 559572,
                    "nscannedObjects" : 559572,
                    "nscanned" : 559572,
                    "indexBounds" : {
                        "datetime" : [
                            [
                                ISODate("2013-07-26T04:00:00Z"),
                                ISODate("2013-07-27T04:00:08.999Z")
                            ]
                        ]
                    }
                }
            ],
            "oldPlan" : {
                "cursor" : "BtreeCursor datetime_1",
                "indexBounds" : {
                    "datetime" : [
                        [
                            ISODate("2013-07-26T04:00:00Z"),
                            ISODate("2013-07-27T04:00:08.999Z")
                        ]
                    ]
                }
            },
            "server" : "Michaels-iMac.local:27017"
        }
    },
    {
        "$project" : {
            "locomotive_id" : "$locomotive_id",
            "loco_time" : {
                "$subtract" : [
                    "$datetime",
                    "$prevdatetime"
                ]
            }
        }
    },
    {
        "$group" : {
            "_id" : "$locomotive_id",
            "utilization_time" : {
                "$sum" : "$loco_time"
            }
        }
    },
    {
        "$sort" : {
            "sortKey" : {
                "_id" : 1
            }
        }
    }
],
"ok" : 1
}
Run Code Online (Sandbox Code Playgroud)

Cra*_*ger 9

PostgreSQL 在这里所做的只是位图堆扫描bpkdmp_datetime_ix以查找可能包含匹配行的块,然后对这些块进行堆扫描以查找bpkdmp. 然后使用分组键的散列将行分组到散列桶中,对每个桶求和,并对结果进行排序。这是一个简单、基本的查询计划——如果你投入大量的work_mem时间,它可能会表现得更好,但也可能不会。

该查询中的任何地方也没有并行性;这一切都将发生在一个核心上。

我只能假设 MongoDB 使用的是效率较低的方法,或者没有从适当的索引中受益。您需要explain为 MongoDB 查询显示有用的注释,以便有可能;见cursor.explain