【笔记】最佳实践—混合负载HTAP的实践和优化

本文涉及的产品
阿里云百炼推荐规格 ADB PostgreSQL,4核16GB 100GB 1个月
云原生数据库 PolarDB 分布式版,标准版 2核8GB
云数据库 RDS MySQL,集群系列 2核4GB
推荐场景:
搭建个人博客
简介: 背景信息 本文主要提供数据库上云后OLTP+OLAP一体化架构的最佳实践,既HTAP。主要面对的业务应用范围: 混合型HTAP数据库需求:如ORACLE数据库改造上云,云上数据库方案选型; OLTP系统查询慢,存在分析型场景和瓶颈的客户; 读写分离需求。

PolarDB-X HTAP架构更多信息,请参见混合负载HTAP

HTAP集群

您购买的PolarDB-X主实例,主要面向在线通用业务场景。如果业务针对同一份数据有分析、专注离线拖数、跑批等场景,您可以在PolarDB-X主实例上购买多个只读实例116.png业务如果有在线HTAP混合流量或者读写分离的需求,推荐使用集群地址。PolarDB-X内部会基于智能路由或者读写权重将部分流量转发给只读实例;业务上只有离线数据分析需求时,推荐使用只读地址,只读地址会直接访问只读实例,只读地址的流量会采用MPP加速。关于连接地址信息,请参见集群地址和只读地址

路由

智能路由

PolarDB-X优化器会基于代价分析出查询物理扫描行数、CPU、内存、IO、网络等核心资源消耗量,将请求区分为TP与AP负载。当您在集群地址上开启了智能路由,会主动识别SQL的工作负载类型来做路由,比如将识别为AP负载的流量路由给只读实例。您可以通过explain cost指令查看SQL工作负载类型的识别情况。例如以下查询,该查询涉及到物理扫描行数rowcount很小,计算资源(CPU&Memory)也消耗比较少,所以这个查询被识别为TP负载。


mysql> explain cost  select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;                                                                                                                                                                                                                                                           |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| TopN(sort="cnt ASC", offset=?2, fetch=?3): rowcount = 1.0, cumulative cost = value = 2.8765038E7, cpu = 37.0, memory = 64.0, io = 3.0, net = 5.75, id = 163602178                                                                                                                 |
|   Filter(condition="cnt > ?1"): rowcount = 1.0, cumulative cost = value = 2.8765026E7, cpu = 26.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602177                                                                                                                            |
|     HashAgg(group="k", cnt="COUNT()"): rowcount = 1.0, cumulative cost = value = 2.8765025E7, cpu = 25.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602171                                                                                                                     |
|       BKAJoin(condition="k = id", type="inner"): rowcount = 1.0, cumulative cost = value = 2.8765012E7, cpu = 12.0, memory = 18.0, io = 3.0, net = 5.75, id = 163602169                                                                                                           |
|         Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 2.3755003E7, cpu = 3.0, memory = 0.0, io = 1.0, net = 4.75, id = 163602164                                                                                                                             |
|           LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)"): rowcount = 1.0, cumulative cost = value = 2.3755002E7, cpu = 2.0, memory = 0.0, io = 1.0, net = 4.75, id = 163601451         |
|         Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 5003.0, cpu = 3.0, memory = 0.0, io = 1.0, net = 0.0, id = 163602167                                                                                                                                   |
|           LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))"): rowcount = 1.0, cumulative cost = value = 5002.0, cpu = 2.0, memory = 0.0, io = 1.0, net = 0.0, id = 163601377 |                                                                                                                                                                                                                                                                    |
| WorkloadType: TP                                                                                                                                                                                                                                                                  |                                                                                                                                                                                                                                                            |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

工作负载的识别,对于HTAP路由至关重要。这里也允许您通过HINT WORKLOAD_TYPE指定工作负载。同样以上述查询为例,可以将查询负载强制指定为AP。


mysql> explain cost /*+TDDL:WORKLOAD_TYPE=AP*/ select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;                                                                                                                                                                                                                                                           |

+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| TopN(sort="cnt ASC", offset=?2, fetch=?3): rowcount = 1.0, cumulative cost = value = 2.8765038E7, cpu = 37.0, memory = 64.0, io = 3.0, net = 5.75, id = 163602178 |
| Filter(condition="cnt > ?1"): rowcount = 1.0, cumulative cost = value = 2.8765026E7, cpu = 26.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602177 |
| HashAgg(group="k", cnt="COUNT()"): rowcount = 1.0, cumulative cost = value = 2.8765025E7, cpu = 25.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602171 |
| BKAJoin(condition="k = id", type="inner"): rowcount = 1.0, cumulative cost = value = 2.8765012E7, cpu = 12.0, memory = 18.0, io = 3.0, net = 5.75, id = 163602169 |
| Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 2.3755003E7, cpu = 3.0, memory = 0.0, io = 1.0, net = 4.75, id = 163602164 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)"): rowcount = 1.0, cumulative cost = value = 2.3755002E7, cpu = 2.0, memory = 0.0, io = 1.0, net = 4.75, id = 163601451 |
| Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 5003.0, cpu = 3.0, memory = 0.0, io = 1.0, net = 0.0, id = 163602167 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))"): rowcount = 1.0, cumulative cost = value = 5002.0, cpu = 2.0, memory = 0.0, io = 1.0, net = 0.0, id = 163601377 | |
| WorkloadType: AP | |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

规则路由

除了基于代价的智能路由以外,我们也支持基于读写规则的路由。允许您在控制台参数管理上设置读写分离权重参数MASTER_READ_WEIGHT,默认值为100,可配置值区间[0, 100]。如果配置为Weight=60,意味着占60%的流量会继续在主实例执行,40%的剩余流量会路由到只读实例执行,如果只读实例有多个会进行自动分配。

智能路由和规则路由这两者关系是解耦的,具体关系请查看下表格。

智能路由规则 规则路由 (MASTER_READ_WEIGHT) 路由结果
开启 以代价的读写分离为主规则路由建议保持默认值为100
  • 事务和写操作流量,全部路由给主实例;
  • 识别为AP的查询流量,全部路由给只读实例;
  • 识别为TP的查询流量,按照(100-MASTER_READ_WEIGHT)路由给只读实例。
关闭 以规则的读写分离为主规则路由的可选范围:[0-100]
  • 事务和写操作流量,全部路由给主实例;
  • 识别为TP/AP的查询流量,一律按照(100-MASTER_READ_WEIGHT)路由给只读实例。

执行模式

目前PolarDB-X支持了三种执行模式:

  • 单机单线程(TP_LOCAL):查询过程中,是单线程计算,TP负载的查询涉及到的扫描行数比较少,往往会采样这种执行模式,比如基于主键的点查。
  • 单机并行(AP_LOCAL):查询过程中,会利用节点的多核资源做并行计算,您在没有购买只读实例的前提下,针对AP负载的查询,往往会采样这种执行模式,一般也称之为Parallel Query模式。
  • 多机并行(MPP):您若购买了只读实例,针对AP负载的查询,可以协调只读实例上多个节点的多核做分布式多机并行加速。

为了让您可以准确知道执行模式,我们在原有EXPLAIN指令上,扩展出了EXPLAIN PHYSICAL。例如以下查询,通过指令可以查看到当前查询采样的是MPP模式,此外还可以获取到每个执行片段的并发数。


mysql> explain physical select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 or
der by cnt limit 5, 10;
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| PLAN |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| ExecutorType: MPP |
| The Query's MaxConcurrentParallelism: 2 |
| Fragment 1 |
| Shuffle Output layout: [BIGINT, BIGINT] Output layout: [BIGINT, BIGINT] |
| Output partitioning: SINGLE [] Parallelism: 1 |
| TopN(sort="cnt ASC", offset=?2, fetch=?3) |
| Filter(condition="cnt > ?1") |
| HashAgg(group="k", cnt="COUNT()") |
| BKAJoin(condition="k = id", type="inner") |
| RemoteSource(sourceFragmentIds=[0], type=RecordType(INTEGER_UNSIGNED id, INTEGER_UNSIGNED k)) |
| Gather(concurrent=true) |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))") |
| Fragment 0 |
| Shuffle Output layout: [BIGINT, BIGINT] Output layout: [BIGINT, BIGINT] |
| Output partitioning: SINGLE [] Parallelism: 1 Splits: 16 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)") |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+

工作负载和执行模式有一定的耦合关系,AP工作负载会采用多机并行计算。同样的也允许您通过HINT EXECUTOR_MODE指定执行模式。假如主实例空闲资源很多,可以考虑强制设置为单机或者多机并行模式来加速。


mysql> explain physical /+TDDL:EXECUTOR_MODE=AP_LOCAL/select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;                                                                                                                                                     |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------+
| ExecutorMode: AP_LOCAL |
| Fragment 0 dependency: [] parallelism: 4 |
| BKAJoin(condition="k = id", type="inner") |
| Gather(concurrent=true) |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)") |
| Gather(concurrent=true) |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))") |
| Fragment 1 dependency: [] parallelism: 8 |
| LocalBuffer |
| RemoteSource(sourceFragmentIds=[0], type=RecordType(INTEGER_UNSIGNED id, INTEGER_UNSIGNED k, INTEGER_UNSIGNED k0)) |
| Fragment 2 dependency: [0, 1] parallelism: 8 |
| Filter(condition="cnt > ?1") |
| HashAgg(group="k", cnt="COUNT()") |
| RemoteSource(sourceFragmentIds=[1], type=RecordType(INTEGER_UNSIGNED id, INTEGER_UNSIGNED k, INTEGER_UNSIGNED k0)) |
| Fragment 3 dependency: [0, 1] parallelism: 1 |
| LocalBuffer |
| RemoteSource(sourceFragmentIds=[2], type=RecordType(INTEGER_UNSIGNED k, BIGINT cnt)) |
| Fragment 4 dependency: [2, 3] parallelism: 1 |
| TopN(sort="cnt ASC", offset=?2, fetch=?3) |
| RemoteSource(sourceFragmentIds=[3], type=RecordType(INTEGER_UNSIGNED k, BIGINT cnt)) |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------+

在多机并行MPP执行模式的并发度是根据物理扫描行数、实例规格和计算所涉及到表的分表数来计算出来的,整体的并行度要考虑高并发场景,所以并行度的计算会偏保守,您可以通过上述EXPLAIN PHYSICAL指令查看并行度。当然也同样支持HINT MPP_PARALLELISM强制指定并行度,


/+TDDL:EXECUTOR_MODE=MPP MPP_PARALLELISM=8/select a.k, count() cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;

调度策略

假设您购买了多个只读实例并加入到集群地址中,您通过集群地址的查询SQL路由到只读实例的流量,会被均匀调度到只读实例多个节点上执行,调度会考虑各个节点的资源负载,确保各个节点的负载压力差不多。比如PolarDB-X会将只读实例延迟作为调度参考指标,避免将流量调度到延迟较大的只读实例上执行。115.png

业务如果有在线HTAP混合流量或者读写分离的需求,推荐使用集群地址。PolarDB-X内部会基于智能路由或者读写权重将部分流量转发给只读实例;业务上只有离线数据分析需求时,推荐使用只读地址,只读地址会直接访问只读实例,只读地址的流量会采用MPP加速。关于连接地址信息,请参见集群地址和只读地址

路由

智能路由

PolarDB-X优化器会基于代价分析出查询物理扫描行数、CPU、内存、IO、网络等核心资源消耗量,将请求区分为TP与AP负载。当您在集群地址上开启了智能路由,会主动识别SQL的工作负载类型来做路由,比如将识别为AP负载的流量路由给只读实例。您可以通过explain cost指令查看SQL工作负载类型的识别情况。例如以下查询,该查询涉及到物理扫描行数rowcount很小,计算资源(CPU&Memory)也消耗比较少,所以这个查询被识别为TP负载。


mysql> explain cost  select a.k, count() cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;                                                                                                                                                                                                                                                           |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| TopN(sort="cnt ASC", offset=?2, fetch=?3): rowcount = 1.0, cumulative cost = value = 2.8765038E7, cpu = 37.0, memory = 64.0, io = 3.0, net = 5.75, id = 163602178 |
| Filter(condition="cnt > ?1"): rowcount = 1.0, cumulative cost = value = 2.8765026E7, cpu = 26.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602177 |
| HashAgg(group="k", cnt="COUNT()"): rowcount = 1.0, cumulative cost = value = 2.8765025E7, cpu = 25.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602171 |
| BKAJoin(condition="k = id", type="inner"): rowcount = 1.0, cumulative cost = value = 2.8765012E7, cpu = 12.0, memory = 18.0, io = 3.0, net = 5.75, id = 163602169 |
| Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 2.3755003E7, cpu = 3.0, memory = 0.0, io = 1.0, net = 4.75, id = 163602164 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)"): rowcount = 1.0, cumulative cost = value = 2.3755002E7, cpu = 2.0, memory = 0.0, io = 1.0, net = 4.75, id = 163601451 |
| Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 5003.0, cpu = 3.0, memory = 0.0, io = 1.0, net = 0.0, id = 163602167 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))"): rowcount = 1.0, cumulative cost = value = 5002.0, cpu = 2.0, memory = 0.0, io = 1.0, net = 0.0, id = 163601377 | |
| WorkloadType: TP | |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

工作负载的识别,对于HTAP路由至关重要。这里也允许您通过HINT WORKLOAD_TYPE指定工作负载。同样以上述查询为例,可以将查询负载强制指定为AP。


mysql> explain cost /+TDDL:WORKLOAD_TYPE=AP/ select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;                                                                                                                                                                                                                                                           |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| TopN(sort="cnt ASC", offset=?2, fetch=?3): rowcount = 1.0, cumulative cost = value = 2.8765038E7, cpu = 37.0, memory = 64.0, io = 3.0, net = 5.75, id = 163602178 |
| Filter(condition="cnt > ?1"): rowcount = 1.0, cumulative cost = value = 2.8765026E7, cpu = 26.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602177 |
| HashAgg(group="k", cnt="COUNT()"): rowcount = 1.0, cumulative cost = value = 2.8765025E7, cpu = 25.0, memory = 47.0, io = 3.0, net = 5.75, id = 163602171 |
| BKAJoin(condition="k = id", type="inner"): rowcount = 1.0, cumulative cost = value = 2.8765012E7, cpu = 12.0, memory = 18.0, io = 3.0, net = 5.75, id = 163602169 |
| Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 2.3755003E7, cpu = 3.0, memory = 0.0, io = 1.0, net = 4.75, id = 163602164 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)"): rowcount = 1.0, cumulative cost = value = 2.3755002E7, cpu = 2.0, memory = 0.0, io = 1.0, net = 4.75, id = 163601451 |
| Gather(concurrent=true): rowcount = 1.0, cumulative cost = value = 5003.0, cpu = 3.0, memory = 0.0, io = 1.0, net = 0.0, id = 163602167 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))"): rowcount = 1.0, cumulative cost = value = 5002.0, cpu = 2.0, memory = 0.0, io = 1.0, net = 0.0, id = 163601377 | |
| WorkloadType: AP | |
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

规则路由

除了基于代价的智能路由以外,我们也支持基于读写规则的路由。允许您在控制台参数管理上设置读写分离权重参数MASTER_READ_WEIGHT,默认值为100,可配置值区间[0, 100]。如果配置为Weight=60,意味着占60%的流量会继续在主实例执行,40%的剩余流量会路由到只读实例执行,如果只读实例有多个会进行自动分配。

智能路由和规则路由这两者关系是解耦的,具体关系请查看下表格。

智能路由规则 规则路由 (MASTER_READ_WEIGHT) 路由结果
开启 以代价的读写分离为主规则路由建议保持默认值为100
  • 事务和写操作流量,全部路由给主实例;
  • 识别为AP的查询流量,全部路由给只读实例;
  • 识别为TP的查询流量,按照(100-MASTER_READ_WEIGHT)路由给只读实例。
关闭 以规则的读写分离为主规则路由的可选范围:[0-100]
  • 事务和写操作流量,全部路由给主实例;
  • 识别为TP/AP的查询流量,一律按照(100-MASTER_READ_WEIGHT)路由给只读实例。

执行模式

目前PolarDB-X支持了三种执行模式:

  • 单机单线程(TP_LOCAL):查询过程中,是单线程计算,TP负载的查询涉及到的扫描行数比较少,往往会采样这种执行模式,比如基于主键的点查。
  • 单机并行(AP_LOCAL):查询过程中,会利用节点的多核资源做并行计算,您在没有购买只读实例的前提下,针对AP负载的查询,往往会采样这种执行模式,一般也称之为Parallel Query模式。
  • 多机并行(MPP):您若购买了只读实例,针对AP负载的查询,可以协调只读实例上多个节点的多核做分布式多机并行加速。

为了让您可以准确知道执行模式,我们在原有EXPLAIN指令上,扩展出了EXPLAIN PHYSICAL。例如以下查询,通过指令可以查看到当前查询采样的是MPP模式,此外还可以获取到每个执行片段的并发数。


mysql> explain physical select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 or
der by cnt limit 5, 10;
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| PLAN |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| ExecutorType: MPP |
| The Query's MaxConcurrentParallelism: 2 |
| Fragment 1 |
| Shuffle Output layout: [BIGINT, BIGINT] Output layout: [BIGINT, BIGINT] |
| Output partitioning: SINGLE [] Parallelism: 1 |
| TopN(sort="cnt ASC", offset=?2, fetch=?3) |
| Filter(condition="cnt > ?1") |
| HashAgg(group="k", cnt="COUNT()") |
| BKAJoin(condition="k = id", type="inner") |
| RemoteSource(sourceFragmentIds=[0], type=RecordType(INTEGER_UNSIGNED id, INTEGER_UNSIGNED k)) |
| Gather(concurrent=true) |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))") |
| Fragment 0 |
| Shuffle Output layout: [BIGINT, BIGINT] Output layout: [BIGINT, BIGINT] |
| Output partitioning: SINGLE [] Parallelism: 1 Splits: 16 |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)") |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+

工作负载和执行模式有一定的耦合关系,AP工作负载会采用多机并行计算。同样的也允许您通过HINT EXECUTOR_MODE指定执行模式。假如主实例空闲资源很多,可以考虑强制设置为单机或者多机并行模式来加速。


mysql> explain physical /+TDDL:EXECUTOR_MODE=AP_LOCAL/select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;                                                                                                                                                     |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------+
| ExecutorMode: AP_LOCAL |
| Fragment 0 dependency: [] parallelism: 4 |
| BKAJoin(condition="k = id", type="inner") |
| Gather(concurrent=true) |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `id`, `k` FROM `sbtest1` AS `sbtest1` WHERE (`id` > ?)") |
| Gather(concurrent=true) |
| LogicalView(tables="[000000-000003].sbtest1_[00-15]", shardCount=16, sql="SELECT `k` FROM `sbtest1` AS `sbtest1` WHERE ((`k` > ?) AND (`k` IN (...)))") |
| Fragment 1 dependency: [] parallelism: 8 |
| LocalBuffer |
| RemoteSource(sourceFragmentIds=[0], type=RecordType(INTEGER_UNSIGNED id, INTEGER_UNSIGNED k, INTEGER_UNSIGNED k0)) |
| Fragment 2 dependency: [0, 1] parallelism: 8 |
| Filter(condition="cnt > ?1") |
| HashAgg(group="k", cnt="COUNT()") |
| RemoteSource(sourceFragmentIds=[1], type=RecordType(INTEGER_UNSIGNED id, INTEGER_UNSIGNED k, INTEGER_UNSIGNED k0)) |
| Fragment 3 dependency: [0, 1] parallelism: 1 |
| LocalBuffer |
| RemoteSource(sourceFragmentIds=[2], type=RecordType(INTEGER_UNSIGNED k, BIGINT cnt)) |
| Fragment 4 dependency: [2, 3] parallelism: 1 |
| TopN(sort="cnt ASC", offset=?2, fetch=?3) |
| RemoteSource(sourceFragmentIds=[3], type=RecordType(INTEGER_UNSIGNED k, BIGINT cnt)) |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------+

在多机并行MPP执行模式的并发度是根据物理扫描行数、实例规格和计算所涉及到表的分表数来计算出来的,整体的并行度要考虑高并发场景,所以并行度的计算会偏保守,您可以通过上述EXPLAIN PHYSICAL指令查看并行度。当然也同样支持HINT MPP_PARALLELISM强制指定并行度,


/+TDDL:EXECUTOR_MODE=MPP MPP_PARALLELISM=8/select a.k, count(*) cnt from sbtest1 a, sbtest1 b where a.id = b.k and a.id > 1000 group by k having cnt > 1300 order by cnt limit 5, 10;

调度策略

假设您购买了多个只读实例并加入到集群地址中,您通过集群地址的查询SQL路由到只读实例的流量,会被均匀调度到只读实例多个节点上执行,调度会考虑各个节点的资源负载,确保各个节点的负载压力差不多。比如PolarDB-X会将只读实例延迟作为调度参考指标,避免将流量调度到延迟较大的只读实例上执行。

相关实践学习
快速体验PolarDB开源数据库
本实验环境已内置PostgreSQL数据库以及PolarDB开源数据库:PolarDB PostgreSQL版和PolarDB分布式版,支持一键拉起使用,方便各位开发者学习使用。
相关文章
|
3月前
|
存储 SQL 监控
OceanBase 的水平扩展与性能优化
【8月更文第31天】随着业务的增长,单一数据库服务器往往难以满足日益增长的数据存储和处理需求。OceanBase 作为一款分布式数据库解决方案,通过其独特的水平扩展能力,能够在不牺牲性能的前提下支持海量数据存储和高并发事务处理。本文将详细介绍 OceanBase 的水平扩展机制,并提供一些性能优化的建议。
274 0
|
3月前
|
存储 缓存 安全
MPP架构数据仓库使用问题之DADI相比其他方案,在资源使用上有什么优势
MPP架构数据仓库使用问题之DADI相比其他方案,在资源使用上有什么优势
|
4月前
|
存储 关系型数据库 数据库
优化数据库性能的关键技术与实践
数据库作为现代应用架构的核心组成部分,其性能优化直接关系到系统整体的稳定性和效率。本文探讨了提升数据库性能的关键技术和实际应用,涵盖了索引优化、查询优化、存储引擎选择以及硬件优化等方面,旨在帮助开发者有效提升应用程序的响应速度和吞吐量。
|
5月前
|
存储 NoSQL 关系型数据库
PolarDB产品使用问题之如何充分利用好产品的性能,提升并发处理能力
PolarDB产品使用合集涵盖了从创建与管理、数据管理、性能优化与诊断、安全与合规到生态与集成、运维与支持等全方位的功能和服务,旨在帮助企业轻松构建高可用、高性能且易于管理的数据库环境,满足不同业务场景的需求。用户可以通过阿里云控制台、API、SDK等方式便捷地使用这些功能,实现数据库的高效运维与持续优化。
|
4月前
|
固态存储 测试技术 数据库
最新测试揭秘:TDengine 线性扩展能力能否满足你的性能需求?
**摘要:** TDengine 的线性扩展能力通过CPU和磁盘测试得以验证。在CPU测试中,使用TDengine V3.3.0.0和taosBenchmark,随着CPU核数从4增至12,写入性能线性提升,每个CPU增加对应约50W条/秒的提升,保持CPU在瓶颈状态。磁盘IO测试中,使用低速机械盘,增加磁盘数量导致写入性能成比例增长,充分利用新增磁盘IO。测试结果表明,无论CPU还是磁盘扩展,TDengine都能有效利用资源,展示出出色的线性扩展能力。
63 0
|
6月前
|
存储 SQL 分布式计算
TiDB整体架构概览:构建高效分布式数据库的关键设计
【2月更文挑战第26天】本文旨在全面概述TiDB的整体架构,深入剖析其关键组件和功能,从而帮助读者理解TiDB如何构建高效、稳定的分布式数据库。我们将探讨TiDB的计算层、存储层以及其他核心组件,并解释这些组件是如何协同工作以实现卓越的性能和扩展性的。通过本文,读者将能够深入了解TiDB的整体架构,为后续的学习和实践奠定坚实基础。
|
6月前
|
存储 缓存 监控
如何在云原生可观测工具中获得更好的性能
如何在云原生可观测工具中获得更好的性能
|
SQL 缓存 关系型数据库
PolarDB-X 混沌测试实践:如何衡量数据库索引选择能力
随着PolarDB分布式版的不断演进,功能不断完善,新的特性不断增多,整体架构扩大的同时带来了测试链路长,出现问题前难发现,出现问题后难排查等等问题。原有的测试框架已经难以支撑实际场景的复杂模拟测试。因此,我们实现了一个基于业务场景面向优化器索引选择的混沌查询实验室,本文之后简称为CEST(complex environment simulation test)。
|
数据采集 算法 编译器
倚天710规模化应用 - 性能优化 -自动反馈优化分析与实践
编译器优化分成静态优化与动态优化,静态优化指传统编译器gcc/llvm时,增加的优化等级,如O1,O2,O3,Ofast,此时,编译器会依据编译优化等级增加一些优化算法,如函数inline、循环展开以及分支静态预测等等。一般情况下,优化等级越高,编译器做的优化越多,性能会更会好。在阿里生产环境中,单纯依赖于静态优化,并不能达到程序运行流畅目的,通过分析CPU硬件取指令、执行指令,往往会出现一些分支预测失败导致iCacheMiss率高的场景,限制了程序的性能进一步提升。基于此,业务引入了动态反馈优化工具,依据生产环境的实际运行数据,反哺指导编译器对程序代码进一步调整编译优化策略,提高分支预准确率
|
6月前
|
前端开发 JavaScript API
前端 JS 经典:Proxy 和 DefineProperty
前端 JS 经典:Proxy 和 DefineProperty
90 0
下一篇
无影云桌面