On Mon, Dec 12, 2011 at 4:28 AM, Takatoshi MATSUO <[email protected]>wrote:

> Hello Serge
>
> 2011/12/12 Serge Dubrouski <[email protected]>:
> >
> >
> > On Thu, Dec 8, 2011 at 10:04 PM, Takatoshi MATSUO <[email protected]>
> > wrote:
> >>
> >> Hello Serge
> >>
> >> 2011/12/8 Serge Dubrouski <[email protected]>:
> >> >
> >> >
> >> > On Mon, Dec 5, 2011 at 9:15 PM, Takatoshi MATSUO <
> [email protected]>
> >> > wrote:
> >> >>
> >> >> Hello Serge
> >> >>
> >> >> Serge Dubrouski <[email protected]>:
> >> >> > Hello -
> >> >> >
> >> >> > Takatoshi MATSUO did a tremendous job on implementing support for
> >> >> > streaming
> >> >> > replication feature in pgsql RA. Also it looks like PostgeSQL 9.1
> has
> >> >> > all
> >> >> > necessary interfaces to successfully implement  Pacemaker's M/S
> >> >> > concept.
> >> >> > So
> >> >> > I think it's time to start discussion on how to merge Takatoshi's
> >> >> > work
> >> >> > into
> >> >> > pgsql RA baseline. Here is the link to Takatoshi's GitHUB if
> somebody
> >> >> > wants
> >> >> > to test his RA:
> >> >> >
> >> >> > https://github.com/t-matsuo/
> >> >> >
> >> >> > So far I tested it for backward compatibility in a standard
> >> >> > non-replication
> >> >> > mode  and also tested M/S model and found no real issues. Though it
> >> >> > definitely requires some more polishing and testing.
> >> >> >
> >> >> > Takatoshi, here are some changes that I want to discuss with you:
> >> >> >
> >> >> > 1. Is it possible to add a check for PostgreSQL version and fail
> with
> >> >> > OCF_ERR_INSTALLED when one tries to start replication on version
> less
> >> >> > than
> >> >> > 9.1? A simple cat on PG_VERSION with some analysis would probably
> do.
> >> >>
> >> >> I'll add a check.
> >>
> >> I added a check.
> >>
> >>
> https://github.com/t-matsuo/resource-agents/commit/3ab7cfdcce118043cd149b348740e50e7a946eb3
> >>
> >> >>
> >> >> > 2. I think that following lines should be moved from pgsql_start to
> >> >> > pgsql_validate_all
> >> >> >
> >> >> >  535     # Check whether tmpdir is readable by pgdba user
> >> >> >  536     if ! runasowner "test -r $OCF_RESKEY_tmpdir"; then
> >> >> >  537         ocf_log err "Directory $OCF_RESKEY_tmpdir is not
> >> >> > readable
> >> >> > by
> >> >> > $OCF_RESKEY_pgdba"
> >> >> >  538         return $OCF_ERR_PERM
> >> >> >  539     fi
> >> >>
> >> >> Thanks. I think so too.
> >> >> I'll fix it.
> >> >>
> >>
> >> I fixed it and I deleted a check for tmpdir existence
> >> because the checking for permittion fills the role.
> >>
> >>
> https://github.com/t-matsuo/resource-agents/commit/82d4939486bcca429e2deb804d7faf756099bb59
> >>
> >>
> >> > On a second thought I'm not sure why we need that parameter and
> >> > directory at
> >> > all. Why not to create rep_mode.conf, PGSQL.lock and xlog.note in
> >> > $OCF_RESKEY_pgdata ? What problems it can create?
> >> >
> >> > One more advantage to do
> >> > it in $OCF_RESKEY_pgdata is an ability to handle more than PostgreSQL
> >> > instance on the same server without a need for additional temp
> >> > directories.
> >>
> >> When backup is needed, customers may backup these files and restore it.
> >> It may cause problems.
> >> Specially PGSQL.lock causes an error on start.
> >>
> >> I think that they should be treated as a separate thing.
> >> because they are independent from PostgreSQL's data.
> >
> >
> > Ok. Then may be it should default to something like
> > ${OCF_RESKEY_pgdata}/temp and RA should create it and set the right
> > ownership and permissions if it doesn't exist? And again may be in this
> case
> > we don't need that parameter?
> >
>
> I agree that RA creates it and sets the right ownership and permissions.
> But considering backup using tar command, I think it's not better to
> create it
> under $OCF_RESKEY_pgdata.
>
> According to "Filesystem Hierarchy Standard"
> it's better to use /var/lib/somewhere.
>
> http://www.pathname.com/fhs/pub/fhs-2.3.html#VARLIBVARIABLESTATEINFORMATION
> ,
>
> Then I designed using /var/lig/[RA Name] (=/var/lib/pgsql).
> What do you think?
>

Ahh, I see now where confusion comes from. /var/lib/pgsql is actually taken
:-) It's used as default home directory in RedHat (at least) for postgres
user in PostgreSQL 8.XX. When they'll start start shipping version 9
they'll probably use it as well. OCF_RESKEY_pg_data  defaults to
/var/lib/pgsql/data.

Then, /var/lib is usually used for non-temporary data, For temporary files
it's probably better to use /var/run or /var/tmp. If you still want to use
/var/lib/pgsql you probably need to use /var/lib/pgsql/temp or so.


> >>
> >>
> >> Incidentally I considered to handle more than PostgreSQL instance.
> >> Initially I added port number to these filenames, but I deleted it
> >> to simplify filenames.
> >>
> >>
> https://github.com/t-matsuo/resource-agents/commit/b16faf2d797200048dc0fc07a45b6751cf5be190
> >>
> >>
> >> > Also I think it would be good if RA was able to take care of adding
> >> > "include
> >> > $WHATEVER_DIR/rep_mode.conf" in postgresql.conf. It will make the RA
> >> > self
> >> > sustainable. In a current situation admin has add that directive
> >> > manually.
> >> > RA though can something like this in a start function for replication
> >> > mode:
> >> >
> >> > if ! grep -i "include $WHATEVER_DIR/rep_mode.conf" $OCF_RESKEY_config
> >> > then
> >> >      echo "include $WHATEVER_DIR/rep_mode.conf" >> $OCF_RESKEY_config
> >> > fi
> >>
> >> Sounds good.
> >>
> >> > Don't know if it makes sense to remove it on stop.
> >>
> >> I think it doesn't make sense to remove it,
> >> because rep_mode.conf becomes empty on stop.
> >
> >
> > The only problem here if admin changes temp_dir parameter but doesn't
> delete
> > records from postgres.conf. We could end up with several include records
> > then or with records pointing to the non existing files.
>
> Should we consider this situation?
>
> If same parameters exist, PostgreSQL uses last parameter.
> So I will implement it to add several include if temp_dir is changed.
>

The problem happens when file gets deleted but reference to it still exists
in postgres.conf. They DB fails to start.


>
>
> >>
> >> >> > 3. I don't really like this part of pgsql_real_monitor:
> >> >> >
> >> >> >  775     if ! is_replication; then
> >> >> >  776         OCF_RESKEY_monitor_sql=`escape_string >
> >> >> > "$OCF_RESKEY_monitor_sql"`
> >> >> >  777         runasowner -q $loglevel "$OCF_RESKEY_psql
> $psql_options
> >> >> > -c
> >> >> > > '$OCF_RESKEY_monitor_sql'"
> >> >> >  778         rc=$?
> >> >> >  779     else
> >> >> >  780         output=`su $OCF_RESKEY_pgdba -c "cd
> $OCF_RESKEY_pgdata;
> >> >> > >
> >> >> > $OCF_RESKEY_psql $psql_options -Atc \"${CHECK_MS_SQL}\""`
> >> >> >  781         rc=$?
> >> >> >  782     fi
> >> >> >
> >> >> > I think that functional monitor (the one that uses monitor_sql)
> >> >> > should
> >> >> > run
> >> >> > always independently of DB mode since its primary role is to check
> >> >> > data
> >> >> > and
> >> >> > fail if it's not correct or corrupted. In replication mode there
> >> >> > should
> >> >> > be
> >> >> > additional monitoring. Other way it misleads customer on a usage of
> >> >> > monitor_sql.
> >> >>
> >> >> All right.
> >> >>
> >> >> Does it need to execute "select now();" if monitor_sql parameter is
> >> >> empty?
> >> >> I think it's unnecessary.
> >> >
> >> >
> >> > For a case of an empty parameter you are right and running "select
> >> > now()"
> >> > probably unnecessary, Non-empty monitro_sql shall be executed in my
> >> > opinion.
> >> >
> >> >>
> >> >>
> >> >> > 4. You already populate several attributes with crm_attribute. Does
> >> >> > it
> >> >> > make
> >> >> > sense to populate a name of a Master node in promote function and
> use
> >> >> > it
> >> >> > later on instead of running crm_mon on each monitor command?
> >> >> >
> >> >>
> >> >> Do you mean that you don't want to use crm_mon in monitor?
> >> >
> >> >
> >> > I prefer to have as few dependencies  on external programs as
> possible.
> >>
> >> Agree.
> >>
> >> > Using crm_attribute to communicate the name of a master node would be
> >> > consistent with the rest of the script since you already use it for
> >> > communicating state of the nodes. If you prefer using crm_mon than you
> >> > have
> >> > to add a check that that binary exists on the server. In 99.99% of the
> >> > case
> >> > it will but still the check is necessary I think.
> >>
> >> I use crm_mon to check for node existence and node online too.
> >> It is impossible to get it from attribute, so I prefer to using crm_mon.
> >>
> >> Do I have to add a check for crm_master, crm_attribute, crm_failcount
> too?
> >>
> >
> > That's a different  interesting topic. What is the reason for doing this?
> >
> > my_fail_count=`$CRM_FAILCOUNT -r $OCF_RESOURCE_INSTANCE -N $HOSTNAME -G
> -Q |
> > sed "s/INFINITY/1000000/g"`
> > if [ "$my_fail_count" != "0" ]; then
> >       ocf_log info "I don't start PostgreSQL on post-demote because of my
> > fail-count=$my_fail_count."
> >       return $OCF_SUCCESS
> > fi
> >
> > Why not to leave pacemaker to do its job on managing failcounts?
>
> Because I want to make fail-over faster.
> At any rate Pacemaker stops it?
>

I'm probably wrong here. Looks like you use this trick in case if slave's
data is newer than new master's data. In this case it makes sense to
disable such slave.


>
> Regards,
> Takatoshi MATSUO
> _______________________________________________________
> Linux-HA-Dev: [email protected]
> http://lists.linux-ha.org/mailman/listinfo/linux-ha-dev
> Home Page: http://linux-ha.org/
>



-- 
Serge Dubrouski.
_______________________________________________________
Linux-HA-Dev: [email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha-dev
Home Page: http://linux-ha.org/

Reply via email to