Re: [PATCH v3] pg_progress() SQL function to monitor progression of long running SQL queries/utilities

2017-07-26 15:27 GMT+02:00 Robert Haas <robertmhaas(at)gmail(dot)com>:

> On Wed, Jun 21, 2017 at 10:01 AM, Remi Colinet <remi(dot)colinet(at)gmail(dot)com>
> wrote:
> > test=# SELECT pid, ppid, bid, concat(repeat(' ', 3 * indent),name),
> value,
> > unit FROM pg_progress(0,0);
> > pid | ppid | bid | concat | value | unit
> > -------+------+-----+------------------+------------------+---------
> > 14106 | 0 | 4 | status | query running |
> > 14106 | 0 | 4 | relationship | progression |
> > 14106 | 0 | 4 | node name | Sort |
> > 14106 | 0 | 4 | sort status | on tapes writing |
> > 14106 | 0 | 4 | completion | 0 | percent
> > 14106 | 0 | 4 | relationship | Outer |
> > 14106 | 0 | 4 | node name | Seq Scan |
> > 14106 | 0 | 4 | scan on | t_10m |
> > 14106 | 0 | 4 | fetched | 25049 | block
> > 14106 | 0 | 4 | total | 83334 | block
> > 14106 | 0 | 4 | completion | 30 | percent
> > (11 rows)
> >
> > test=#
>
> Somehow I imagined that the output would look more like what EXPLAIN
> produces.
>

I had initially used the same output as for the ANALYZE command:

test=# PROGRESS 14611;
PLAN
PROGRESS
-----------------------------------------------------------------------------------------
Gather Merge
-> Sort=> dumping tuples to tapes
rows r/w merge 0/0 rows r/w effective 0/1464520 0%
Sort Key: md5
-> Parallel Seq Scan on t_10m => rows 1464520/4166700 35% blks
36011/83334 43%
(5 rows)

test=#

But this restricts the use to "human consumers". Using a table output with
name/value pairs, allows the use by utilities for instance, without
parsing. This is less handy for administrators, but far better for 3rd
party utilities. One solution is otherwise to create a PL/SQL command on
top of pg_progress() SQL function to produce an output similar to the one
of the ANALYZE command.

> > If the one shared memory page is not enough for the whole progress
> report,
> > the progress report transfert between the 2 backends is done with a
> series
> > of request/response. Before setting the latch, the monitored backend
> write
> > the size of the data dumped in shared memory and set a status to indicate
> > that more data is to be sent through the shared memory page. The
> monitoring
> > backends get the result and sends an other signal, and then wait for the
> > latch again. The monitored backend does not collect a new progress report
> > but continues to dump the already collected report. And the exchange
> goes on
> > until the full progress report has been dumped.
>
> This is basically what shm_mq does. We probably don't want to
> reinvent that code, as it has taken a surprising amount of debugging
> to get it fully working.
>

Yes, I had once considered this solution but then moved away as I was
unsure of the exact need for the transfert of the progress report between
the monitored and the monitoring backends.
I'am going to switch to shm_mq.

Thx & Rgds

>
> --
> Robert Haas
> EnterpriseDB: http://www.enterprisedb.com
> The Enterprise PostgreSQL Company
>