1 Data structures

1.1 Slice Table

@startuml
class SliceTable {
+ NodeTag type
+ int localSlice
+ int numSlices
+ ExecSlice slices
+ bool hasMotions
+ int instrument_options
+ uint32 ic_instance_id
}

note right of SliceTable::localSlice
Index of the slice to execute
end note

note right of SliceTable::slices
Array of slices, indexed by SliceIndex
end note


note right of SliceTable::hasMotions
Are there any Motion nodes anywhere in the plan?
end note


class ExecSlice {
+ int sliceIndex
+ int rootIndex
+ int parentIndex
+ int planNumSegments
+ List children
+ GangType gangType
+ List segments
+ struct primaryGang
+ List primaryProcesses
+ Bitmapset processesMap
}

note right of ExecSlice::primaryProcesses
A list of CDBProcess nodes corresponding to the worker
processes allocated to implement this plan slice.
end note

note right of ExecSlice::processesMap
A bitmap to identify which QE should execute this slice
end note

SliceTable o-- ExecSlice

class Gang {
+ GangType type
+ int size
+ struct db_descriptors
+ bool allocated
}

note right of Gang::db_descriptors
Array of QEs/segDBs that make up this gang.
Sorted by segment index.
end note


ExecSlice *-- Gang

class CdbProcess {
+ NodeTag type
+ char listenerAddr
+ int listenerPort
+ int pid
+ int contentid
+ int dbid
}

ExecSlice o-- CdbProcess



class SegmentDatabaseDescriptor {
+ struct segment_database_info
+ int segindex
+ int conn
+ int motionListener
+ int backendPid
+ char whoami
+ int isWriter
+ int identifier
}

Gang o-- SegmentDatabaseDescriptor



class CdbComponentDatabases {
+ CdbComponentDatabaseInfo segment_db_info
+ int total_segment_dbs
+ CdbComponentDatabaseInfo entry_db_info
+ int total_entry_dbs
+ int total_segments
+ int fts_version
+ int expand_version
+ int numActiveQEs
+ int numIdleQEs
+ int qeCounter
+ List freeCounterList
}

note right of CdbComponentDatabaseInfo::segment_db_info
array of  SegmentDatabaseInfo for segment databases
end note

note right of CdbComponentDatabaseInfo::entry_db_info
array of  SegmentDatabaseInfo for entry databases
end note


class CdbComponentDatabaseInfo {
+ struct config
+ CdbComponentDatabases cdbs
+ int hostSegs
+ List freelist
+ int numIdleQEs
+ int numActiveQEs
}

note right of CdbComponentDatabaseInfo::cdbs
point to owners
end note

CdbComponentDatabases o-- CdbComponentDatabaseInfo



class GpSegConfigEntry {
+ int dbid
+ int segindex
+ char role
+ char preferred_role
+ char mode
+ char status
+ int port
+ char hostname
+ char address
+ char datadir
+ char hostip
+ char hostaddrs
}

CdbComponentDatabaseInfo o-- GpSegConfigEntry

SegmentDatabaseDescriptor o-- CdbComponentDatabaseInfo

@enduml

2

1.1.1 SliceTable

SliceTable : 由 Slice 组成的链表， Slice 组织成三类：

  • root slices: Slice 0

  • motion slices 1 ~ n 为 motion slices, 每个 slice 的根为 sending motion

  • 其余为 initPlans

1.1.2 ExecSlice

  • MPP 中，计划树 (PlanTree) 被切分成多个单独的执行单元 （又称 Slice ）
  • 一个 Slice 在进程组 （process gang） 的一个 worker 上执行

2 PostgresMain

2.1 Call graph (QE):

Main()
  PostmasterMain()
    ServerLoop()
      BackendStartup()
        BackendRun()
          PostgresMain()
            InitPostgres()
              cdb_setup()
                ensureInterconnectAddress()
                InitMotionLayerIPC()
                  InitMotionTCP()
                      setupTCPListeningSocket()
            sendQEDetails()

setupTCPListeningSocket() 会由操作系统分配端口，并返回上层。并在 InitMotionlayerIPC() 中存储在全局变量 Gp_listener_port 中，并随后在函数 sendQEDetails(void) 中将端口信息 “qe_listener_port” 发送给 client 。

2.2 QD

standard_ExecutorStart()
  CdbDispatchPlan()
    cdbdisp_dispatchX()
      AssignGangs()
        AssignWriterGangFirst()
          AllocateGang()
            cdbgang_createGang()
              cdbgang_createGang_async()
                cdbconn_doConnectComplete()
                  cdbconn_get_motion_listener_port()
          setupCdbProcessList()

函数 AssignGangs() 在 QD 上执行， 将 Executor 工厂分配的 gangs 分配给 slice table 中的 slices. 从而构建全局的 slice table 。该过程分成了两步：

3 Receiver

4 Sender