Re: using custom scan nodes to prototype parallel sequential scan - Mailing list pgsql-hackers

>       On 12 November 2014 07:54, David Rowley <dgrowleyml@gmail.com>
> Take int4_avg_accum() for example it does:
>
>
> transdata->count++;
> transdata->sum += newval;
>
> The merge function would need to perform something like:
>

> transdata->count += transdata2merge.count; sum += transdata2merge.sum;
>
> Then the final function could be called on the merged aggregate state.
>
More simplify, we can describe parallel aware aggregate function.
Please assume AVG(X) function that takes nrows and sum of X. Its transition
function performs as like you described above, then final function works as
usual.

The job of parallel seq scan needs to do is:
  1. replace AVG(X) by AVG(nrows, sum(X)
  2. generate count(*) on the partial relation being grouped.
  3. generate sum(X) on the partial relation being grouped.

It looks like the following query:
  SELECT AVG(nrows, sum_X) FROM (
    SELECT count(*) nrows, sum(X) sum_X FROM tbl WHERE blkno between 0 and 999 GROUP BY cat
    UNION
    SELECT count(*) nrows, sum(X) sum_X FROM tbl WHERE blkno between 1000 and 1999 GROUP BY cat
    UNION
    SELECT count(*) nrows, sum(X) sum_X FROM tbl WHERE blkno between 2000 and 2999 GROUP BY cat
    UNION
    SELECT count(*) nrows, sum(X) sum_X FROM tbl WHERE blkno between 3000 and 3999 GROUP BY cat
  );