Greenplum数据库数据分片策略Hash分布——执行器行为

Greenplum数据库Hash分布执行器部分主要涉及Motion、Result和SplictUpdate节点。以使用CdbHash *makeCdbHash(int numsegs, int natts, Oid *hashfuncs)创建一个 CdbHash 结构体、cdbhashinit()执行初始化操作，仅仅是初始化hash初始值、cdbhash()函数会调用hashDatum()针对不同类型做不同的预处理，最后将处理后的列值添加到哈希计算中、cdbhashreduce() 映射哈希值到某个 segment为脉络学习以下执行器对Hash分布的处理。

Motion

只有当Motion类型为MOTIONTYPE_HASH且执行发送任务(MOTIONSTATE_SEND)的后端才可能涉及Hash分布处理（​​motionstate->mstype == MOTIONSTATE_SEND && node->motionType == MOTIONTYPE_HASH​​）。也就是说后端进程需要将处理的数据直接发送给其他后端进程，且这个接收后端可以通过分布键数据进行计算哈希值、映射segment后定位到。其执行堆栈为ExecInitNode --> ExecInitMotion --> makeCdbHash。

涉及hash的motion执行流程堆栈如下ExecMotion --> execMotionSender --> doSendTuple --> eval（nodeMotion.c) --> cdbhashinit和cdbhash。调用doSendTuple发送tuple的情况下，当motion类型为MOTIONTYPE_HASH，就需要计算出该segment的index，然后设置到targetRoute变量中。计算的函数就是eval（232, 232, 232); background: rgb(249, 249, 249);">

uint32 eval（ExprContext *econtext, List *hashkeys, CdbHash * h) {
  ListCell   *hk; unsigned int target_seg;
  ResetExprContext(econtext);
  MemoryContext oldContext = MemoryContextSwitchTo(econtext->ecxt_per_tuple_memory); // 切换到ecxt_per_tuple_memory
  
  if (list_length(hashkeys) > 0){ /* If we have 1 or more distribution keys for this relation, hash them. However, If this happens to be a relation with an empty policy (partitioning policy with a NULL distribution key list) then we have no hash key value to feed in, so use cdbhashrandomseg() to pick a segment at random. */
    int      i = 0;
    cdbhashinit(h);
    foreach(hk, hashkeys){
      ExprState  *keyexpr = (ExprState *) lfirst(hk);
      Datum    keyval; bool    isNull;    
      keyval = ExecEvalExpr(keyexpr, econtext, &isNull); /* Get the attribute value of the tuple */      
      cdbhash(h, i + 1, keyval, isNull); /* Compute the hash function */
      i++;
    }
    target_seg = cdbhashreduce(h);
  }else {
    target_seg = cdbhashrandomseg(h->numsegs);
  }
  MemoryContextSwitchTo(oldContext);
  return target_seg;
}