混合负载HTAP的实践和优化
背景信息
本文主要提供数据库 OLTP + OLAP 一体化架构的最佳实践，即 HTAP。主要面对的业务应用范围：
混合型HTAP数据库需求：如ORACLE数据库改造迁移，新数据库方案选型；
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HTAP。
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，PolarDB-X 也支持基于读写规则的路由。可通过设置读写分离权重参数 MASTER_READ_WEIGHT 实现。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）：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
假设有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。
反馈机制
基于统计信息智能识别负载存在一定的误差。在负载识别出错的情况下，PolarDB-X能基于查询最终真实扫描的物理行数和执行时间，重新修正工作负载，这个过程我们称之为负载的自适应反馈。根据执行计划管理一文，PolarDB-X的执行计划会被管理起来，同样其工作负载也会被管理起来，用户可以通过以下语句，查看执行计划管理中各个计划的工作负载。
baseline [Select Statemtnt]
比如当一条TP负载的简单查询，执行时间和真实扫描行数都大于阈值，这个查询就会被修正为AP工作负载，写入到执行计划管理状态中，同理反之。除了上述提到了自适应反馈能力以外，我们也可以通过以下语句，手动修正计划管理中的工作负载。
baseline fix sql /*+TDDL:WORKLOAD_TYPE=AP*/ [Select Statemtnt]
执行计划管理中的计划负载被合理修正后，同类型的查询会从执行计划管理中获取正确的工作负载标记。
一致性读
如果业务使用的集群地址，对数据一致性读并没有强烈诉求，例如业务可以忽略当前只读实例复制延迟带来的数据影响，我们也支持在参数管理页面上配置ENABLE_CONSISTENT_REPLICA_READ，或者通过HINT对某个查询关闭一致性读能力。
/*+TDDL:ENABLE_CONSISTENT_REPLICA_READ=false*/ [Select Statemtnt]