Обсуждение: proposal: contrib module - generic command scheduler

On Tue, May 12, 2015 at 10:25 AM, Pavel Stehule <pavel.stehule@gmail.com> wrote:
> Generic simple scheduler to contrib
> ===================================
> Job schedulers are important and sometimes very complex part of any
> software. PostgreSQL miss it. I propose new contrib module, that can be used
> simply for some tasks, and that can be used as base for other more richer
> schedulers. I prefer minimalist design - but strong enough for enhancing
> when it is necessary. Some complex logic can be implemented in PL better
> than in C. Motto: Simply to learn, simply to use, simply to customize.
>
> Motivation
> ----------
> Possibility to simplify administration of repeated tasks. Possibility to
> write complex schedulers in PL/pgSQL or other PL.
>
> Design
> ------
> Any scheduled command will be executed in independent worker. The number
> workers for one command can be limited. Total number of workers will be
> limited. Any command will be executed under specified user with known
> timeout in current database. Next step can be implementation global
> scheduler - but we have not a environment for running server side global
> scripts, so I don't think about it in this moment.
>
> This scheduler does not guarantee number of executions. Without available
> workers the execution will be suspended, after crash the execution can be
> repeated. But it can be solved in upper layer if it is necessary. It is not
> designed as realtime system. Scheduled task will be executed immediately
> when related worker will be free, but the execution window is limited to
> next start.
>
> This design don't try to solve mechanism for repeating tasks when tasks hash
> a crash. This can be solved better in PL on custom layer when it is
> necessary.
>
> Scheduled time is stored to type scheduled_time:
>
> create type scheduled_time as (second int[], minute int[], hour int[], dow
> int[], month int[]);
>
>  (,"{1,10,20,30,40,50}",,,) .. run every 10 minutes.
>  (,"{5}",,,) .. run once per hour
>
> The core is table pg_scheduled_commands
>
> Oid:           1
> name:        xxxx
> user:          pavel
> stime:        (,"{5}",,,)
> max_workers: 1
> timeout:          10s
> command: SELECT plpgsql_entry(scheduled_time(), scheduled_command_oid())
>
>
> set timeout to 0 ~ unlimited, -1 default statement_timeout
> set max_workers to 0 ~ disable tasks
>
> API
> ---
> pg_create_scheduled_command(name,
>                          stime,
>                          command,
>                          user default current_user,
>                          max_workers default 1,
>                          timeout default -1);
>
> pg_drop_scheduled_command(oid)
> pg_drop_scheduled_command(name);
>
> pg_update_scheduled_command(oid | name, ...
>
> Usage:
> ------
> pg_create_scheduled_command('delete obsolete data', '(,,"{1}",,)', $$DELETE
> FROM data WHERE inserted < current_timestamp - interval '1month'$$);
> pg_update_scheduled_command('delete obsolete data', max_workers => 2,
> timeout :=> '1h');
> pg_drop_scheduled_command('delete obsolete data');
>
> select * from pg_scheduled_commands;
>
>
> Comments, notices?

It's not integrated with the server (though it is integrated with
pgAdmin), but pgAgent provides scheduling services for PostgreSQL
already, offering multi-schedule, multi-step job execution.

http://www.pgadmin.org/docs/1.20/pgagent.html

--
Dave Page
Blog: http://pgsnake.blogspot.com
Twitter: @pgsnake

EnterpriseDB UK: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Tue, May 12, 2015 at 09:25:50AM +0200, Pavel Stehule wrote:
> create type scheduled_time as (second int[], minute int[], hour int[], dow
> int[], month int[]);
>  (,"{1,10,20,30,40,50}",,,) .. run every 10 minutes.
>  (,"{5}",,,) .. run once per hour
> Comments, notices?

First, please note that I'm definitely not a hacker, just a user.

One comment that I'd like to make, is that since we're at planning
phase, I think it would be great to add capability to limit number of
executions of given command.
This would allow running things like "at" in unix - run once, at given
time, and that's it.

Best regards,

depesz

--
The best thing about modern society is how easy it is to avoid contact with it.
                                                             http://depesz.com/

2015-05-12 11:27 GMT+02:00 hubert depesz lubaczewski <depesz@depesz.com>:

On Tue, May 12, 2015 at 09:25:50AM +0200, Pavel Stehule wrote:
> create type scheduled_time as (second int[], minute int[], hour int[], dow
> int[], month int[]);
>  (,"{1,10,20,30,40,50}",,,) .. run every 10 minutes.
>  (,"{5}",,,) .. run once per hour
> Comments, notices?

First, please note that I'm definitely not a hacker, just a user.

One comment that I'd like to make, is that since we're at planning
phase, I think it would be great to add capability to limit number of
executions of given command.
This would allow running things like "at" in unix - run once, at given
time, and that's it.

I would not to store state on this level - so "at" should be implemented on higher level. There is very high number of possible strategies, what can be done with failed tasks - and I would not to open this topic. I believe with proposed scheduler, anybody can simply implement what need in PLpgSQL with dynamic SQL. But on second hand "run once" can be implemented with proposed API too.

On 13 May 2015 at 00:31, Pavel Stehule <pavel.stehule@gmail.com> wrote:

2015-05-12 11:27 GMT+02:00 hubert depesz lubaczewski <depesz@depesz.com>:

On Tue, May 12, 2015 at 09:25:50AM +0200, Pavel Stehule wrote:
> create type scheduled_time as (second int[], minute int[], hour int[], dow
> int[], month int[]);
>  (,"{1,10,20,30,40,50}",,,) .. run every 10 minutes.
>  (,"{5}",,,) .. run once per hour
> Comments, notices?

First, please note that I'm definitely not a hacker, just a user.

One comment that I'd like to make, is that since we're at planning
phase, I think it would be great to add capability to limit number of
executions of given command.
This would allow running things like "at" in unix - run once, at given
time, and that's it.

I would not to store state on this level - so "at" should be implemented on higher level. There is very high number of possible strategies, what can be done with failed tasks - and I would not to open this topic. I believe with proposed scheduler, anybody can simply implement what need in PLpgSQL with dynamic SQL. But on second hand "run once" can be implemented with proposed API too.

That seems reasonable in a v1, so long as there's room to easily extend it without pain to add "at"-like one-shot commands, at-startup commands, etc.

I'd prefer to see a scheduling interface that's a close match for cron's or that leaves room for it - so things like "*/5" for every five minutes, ranges like "Mon-Fri", etc. If there's a way to express similar capabilities more cleanly using PostgreSQL's types and conventions that makes sense, but I'm not sure a composite type of arrays fits that.

How do you plan to manage the bgworkers?

In BDR, where we have a similar need to have workers across multiple databases, and where each database contains a list of workers to launch, we have:

* A single static "supervisor" bgworker. In 9.5 this will connect with InvalidOid as the target database so it can only access shared catalogs. In 9.4 this isn't possible in the bgworker API so we have to connect to a dummy database.

* A dynamic background worker for each database in which BDR is enabled, which is launched from the supervisor. We check which DBs are BDR-enabled by (ab)using database security labels and checking pg_shseclabel from the supervisor worker so we only launch bgworkers on BDR-enabled DBs.

* A dynamic background worker for each peer node, launched by the per-database worker based on the contents of that database's bdr.bdr_connections table.

What I suspect you're going to want is:

* A static worker launched by your extension when it starts, which launches per-db workers for each DB in which the scheduler is enabled. You could use a GUC listing scheduler-enabled DBs in postgresql.conf and have an on-reload hook to update it, you don't need to do the security label hack.

* A DB scheduler worker, which looks up the scheduled tasks list, finds the next scheduled event, and sleeps on a long latch timeout until then, resetting it when interrupted. When it reaches the scheduled event it would launch a one-shot BGW_NO_RESTART worker to run the desired PL/PgSQL procedure over the SPI.

* A task runner worker, which gets launched by the db scheduler to actually run a task using the SPI.

Does that match your current thinking?

--

 Craig Ringer                   http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services

On 5/12/15 11:32 PM, Pavel Stehule wrote:

        I would not to store state on this level - so "at" should be
        implemented on higher level. There is very high number of
        possible strategies, what can be done with failed tasks - and I
        would not to open this topic. I believe with proposed scheduler,
        anybody can simply implement what need in PLpgSQL with dynamic
        SQL. But on second hand "run once" can be implemented with
        proposed API too.


    That seems reasonable in a v1, so long as there's room to easily
    extend it without pain to add "at"-like one-shot commands,
    at-startup commands, etc.

Yeah, being able to run things after certain system events would be nice.

    I'd prefer to see a scheduling interface that's a close match for
    cron's or that leaves room for it - so things like "*/5" for every
    five minutes, ranges like "Mon-Fri", etc. If there's a way to
    express similar capabilities more cleanly using PostgreSQL's types
    and conventions that makes sense, but I'm not sure a composite type
    of arrays fits that.

It seems unfortunate to go with cron's limited syntax when we have such fully capable timestamp and interval capabilities already in the database. :/

Is there anything worth stealing from pgAgent?

I though about it too - but the parser for this cron time will be longer
than all other code probably. I see a possibility to write constructors
that simplify creating a value of this type. Some like

make_scheduled_time(secs => '*/5', dows => 'Mon-Fri') or
make_scheduled_time(at =>'2015-014-05 10:00:0'::timestamp);

Wouldn't that be just as bad as writing the parser in the first place?

1. don't hold a states, results of commands

...
> 3. When command fails, it writes info to log only
Unfortunate, but understandable in a first pass.

4. When command runs too long (over specified timeout), it is killed.

I think that needs to be optional.

5. When command waits to free worker, write to log
6. When command was not be executed due missing workers (and max_workers
 > 0), write to log

Also unfortunate. We already don't provide enough monitoring capability and this just makes that worse.

Perhaps it would be better to put something into PGXN first; this doesn't really feel like it's baked enough for contrib yet. (And I say that as someone who's really wanted this ability in the past...)

--
Jim Nasby, Data Architect, Blue Treble Consulting
Data in Trouble? Get it in Treble! http://BlueTreble.com

On 5/13/15 1:32 AM, Pavel Stehule wrote:

        5. When command waits to free worker, write to log
        6. When command was not be executed due missing workers (and
        max_workers
          > 0), write to log


    Also unfortunate. We already don't provide enough monitoring
    capability and this just makes that worse.


theoretically it can be supported some pg_stat_ view - but I would not
to implement a some history table for commands. Again it is task for
higher layers.

I don't think we want to log statements, but we should be able to log when a job has run and whether it succeeded or not. (log in a table, not just a logfile).

This isn't something that can be done at higher layers either; only the scheduler will know if the job failed to even start, or whether it tried to run the job.

--
Jim Nasby, Data Architect, Blue Treble Consulting
Data in Trouble? Get it in Treble! http://BlueTreble.com

On 5/14/15 1:36 AM, Pavel Stehule wrote:

    I don't think we want to log statements, but we should be able to
    log when a job has run and whether it succeeded or not. (log in a
    table, not just a logfile).

    This isn't something that can be done at higher layers either; only
    the scheduler will know if the job failed to even start, or whether
    it tried to run the job.


I don't agree - generic scheduler can run your procedure, and there you
can log start, you can run other commands and you can log result (now
there is no problem to catch any production nonfatal exception).

And what happens when the job fails to even start? You get no logging.

Personally I afraid about responsibility to maintain this log table -
when and by who it should be cleaned, who can see results, ... This is
job for top end scheduler.

Only if the top-end scheduler has callbacks for everytime the bottom-end scheduler tries to start a job. Otherwise, the top has no clue what the bottom has actually attempted.

To be clear, I don't think these need to be done in a first pass. I am concerned about not painting ourselves into a corner though.

--
Jim Nasby, Data Architect, Blue Treble Consulting
Data in Trouble? Get it in Treble! http://BlueTreble.com

Pavel Stehule wrote:

Hi,

> Job schedulers are important and sometimes very complex part of any
> software. PostgreSQL miss it. I propose new contrib module, that can be
> used simply for some tasks, and that can be used as base for other more
> richer schedulers. I prefer minimalist design - but strong enough for
> enhancing when it is necessary. Some complex logic can be implemented in PL
> better than in C. Motto: Simply to learn, simply to use, simply to
> customize.

Have you made any progress on this?

--
Álvaro Herrera                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services