Re: Enabling parallelism for queries coming from SQL or other PL functions

On Wed, Mar 15, 2017 at 10:52 AM, Dilip Kumar <dilipbalaut(at)gmail(dot)com> wrote:
> I have reviewed the patch, I have some questions.
>
> @@ -3031,7 +3031,7 @@ exec_stmt_return_query(PLpgSQL_execstate *estate,
> Assert(stmt->dynquery != NULL);
> portal = exec_dynquery_with_params(estate, stmt->dynquery,
> stmt->params, NULL,
> - CURSOR_OPT_PARALLEL_OK);
> + 0);
>
> After this patch, I have noticed that In exec_stmt_return_query, we
> allow parallel query if it's a static query
> but not for the dynamic query. Any specific reason for different
> handling for these 2 cases?
>
The reason for such behaviour is given upthread, in
exec_stmt_return_query, the query may be executed by either
exec_run_select which then uses SPI_execute_plan_with_paramlist(),
which calls _SPI_execute_plan(), which calls _SPI_pquery(), hence if
parallelOK is set then passing CURSOR_OPT_PARALLEL_OK is safe.
Otherwise, exec_stmt_return_query executes with a call to
SPI_cursor_fetch() with a fetch count of 50, and that calls
_SPI_cursor_operation() which calls PortalRunFetch(), hence, passing
CURSOR_OPT_PARALLEL_OK would not be safe in this case hence not
passed.
>
> @ -314,7 +314,7 @@ SPI_execute(const char *src, bool read_only, long tcount)
>
> memset(&plan, 0, sizeof(_SPI_plan));
> plan.magic = _SPI_PLAN_MAGIC;
> - plan.cursor_options = 0;
> + plan.cursor_options = CURSOR_OPT_PARALLEL_OK;
>
> In SPI_Execute, we are setting cursor_options to
> CURSOR_OPT_PARALLEL_OK. I have analysed call to this function from PL
> and seems fine to me. But, I have a question have you analyzed all the
> calls to this functions?
> e.g. build_tuplestore_recursively, get_crosstab_tuplestore.
>
The thing with SPI_execute is that it calls pg_plan_query through
_SPI_execute_plan, there if the query is parallel unsafe then
parallelism will be restricted by planner itself otherwise it is
enabled if CURSOR_OPT_PARALLEL_OK was set. So, yes I evaluated all the
calls for this functions and here is the analysis of why it should be
safe passing CURSOR_OPT_PARALLEL_OK, in some of the suspicious looking
calls(other calls are directly related to spy functions)
In crosstab: ret = SPI_execute(sql, true, 0);
In load_categories_hash: ret = SPI_execute(cats_sql, true, 0);
In get_crosstab_tuplestore: ret = SPI_execute(sql, true, 0);
In build_tuplestore_recursively: ret = SPI_execute(sql.data, true, 0);
In query_to_oid_list: SPI_execute(query, true, 0);

In all of these calls, read_only flag is passed to be true, hence
enabling parallelism will not cause any anomalous behaviour.

In refresh_by_match_merge: if (SPI_execute(querybuf.data, false, 1) !=
SPI_OK_SELECT)
In this case, since read_only is set to false, hence, in SPI_execute
when planner will recalled for such a case it will disable
parallelism, hence, no issues.

if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
IsUnderPostmaster &&
dynamic_shared_memory_type != DSM_IMPL_NONE &&
parse->commandType == CMD_SELECT &&
!parse->hasModifyingCTE &&
max_parallel_workers_per_gather > 0 &&
!IsParallelWorker() &&
!IsolationIsSerializable())

>
> On Fri, Mar 10, 2017 at 5:38 PM, Rafia Sabih
> <rafia(dot)sabih(at)enterprisedb(dot)com> wrote:
>> I wanted to clarify a few things here, I noticed that call of ExecutorRun in
>> postquel_getnext() uses !es->lazyEval as execute_once, this is confusing, as
>> it is true even in cases when a simple query like "select count(*) from t"
>> is used in a sql function. Hence, restricting parallelism for cases when it
>> shouldn't. It seems to me that es->lazyEval is not set properly or it should
>> not be true for simple select statements. I found that in the definition of
>> execution_state
>> bool lazyEval; /* true if should fetch one row at a time
>
> Hmm, It seems that es->lazyEval is not set properly, Ideally, it
> should be set true only if it's lazy evaluation but in this case, it's
> used for identifying the tupcount as well. IMHO, it should be fixed.
>
> Any other opinion on this?
>
Hmmm... I tried investigating into this but it looks like there isn't
much scope for this. LazyEvalOk is set for SELECT commands in
init_execution_state as per,
/*
* Mark the last canSetTag query as delivering the function result; then,
* if it is a plain SELECT, mark it for lazy evaluation. If it's not a
* SELECT we must always run it to completion.
...
if (lasttages && fcache->junkFilter)
{
lasttages->setsResult = true;
if (lazyEvalOK &&
lasttages->stmt->commandType == CMD_SELECT &&
!lasttages->stmt->hasModifyingCTE)
fcache->lazyEval = lasttages->lazyEval = true;
}
and then in postquel_getnext we set execute_once = !es->lazyEval [1],
which also makes sense since,
/* Run regular commands to completion unless lazyEval */

Hence, this situation looks like it is restricting parallelism for
some cases where it might not cause any issues, but clearly modifying
lazyEval is not a safe option. Additionally, I think we do not have
enough information to ensure that a select query will not cause
multiple ExecutorRun calls for these user-defined queries inside SQL
functions. Moreover, lazyEval is a kind of may be thing, meaning that
it might call ExecutorRun multiple times when set but not ensured that
it will as in the case of select count(*) from t queries.
Please let me know your views on this.

[1] https://www.postgresql.org/message-id/CA+TgmobXEhvHbJtWDuPZM9bVSLiTj-kShxQJ2uM5GPDze9fRYA@mail.gmail.com
--
Regards,
Rafia Sabih
EnterpriseDB: http://www.enterprisedb.com/