[pgpool-general: 154] Re: Healthcheck timeout not always respected

Sun Jan 15 07:44:21 JST 2012

Hello Tatsuo,

Thank you for the patch and effort, but unfortunately this change won't
work for us. We need to set disallow failover to prevent failover on child
reported connection errors (it's ok if few clients lose their connection or
can not connect), and still have pgpool perform failover but only on failed
health check (if configured, after max retries threshold has been reached).

Maybe it would be best to add an extra value for backend_flag -
ALLOW_TO_FAILOVER_ON_HEALTH_CHECK or DISALLOW_TO_FAILOVER_ON_CHILD_ERROR.
It should behave same as DISALLOW_TO_FAILOVER is set, with only difference
in behaviour when health check (if set, max retries) has failed - unlike
DISALLOW_TO_FAILOVER, this new flag should allow failover in this case only.

Without this change health check (especially health check retries) doesn't
make much sense - child error is more likely to occur on (temporary)
backend failure then health check and will or will not cause failover to
occur depending on backend flag, without giving health check retries a
chance to determine if failure was temporary or not, risking split brain
situation with two masters just because of temporary network link hiccup.

Our main problem remains though with the health check timeout not being
respected in these special conditions we have. Maybe Nenad can help you
more to reproduce the issue on your environment.

Kind regards,
Stevo.

2012/1/13 Tatsuo Ishii <ishii at postgresql.org>

> Thanks for pointing it out.
> Yes, checking DISALLOW_TO_FAILOVER before retrying is wrong.
> However, after retry count over, we should check DISALLOW_TO_FAILOVER I
> think.
> Attached is the patch attempt to fix it. Please try.
> --
> Tatsuo Ishii
> SRA OSS, Inc. Japan
> English: http://www.sraoss.co.jp/index_en.php
> Japanese: http://www.sraoss.co.jp
>
> > pgpool is being used in raw mode - just for (health check based) failover
> > part, so applications are not required to restart when standby gets
> > promoted to new master. Here is pgpool.conf file and a very small patch
> > we're using applied to pgpool 3.1.1 release.
> >
> > We have to have DISALLOW_TO_FAILOVER set for the backend since any child
> > process that detects condition that master/backend0 is not available, if
> > DISALLOW_TO_FAILOVER was not set, will degenerate backend without giving
> > health check a chance to retry. We need health check with retries because
> > condition that backend0 is not available could be temporary (network
> > glitches to the remote site where master is, or deliberate failover of
> > master postgres service from one node to the other on remote site - in
> both
> > cases remote means remote to the pgpool that is going to perform health
> > checks and ultimately the failover) and we don't want standby to be
> > promoted as easily to a new master, to prevent temporary network
> conditions
> > which could occur frequently to frequently cause split brain with two
> > masters.
> >
> > But then, with DISALLOW_TO_FAILOVER set, without the patch health check
> > will not retry and will thus give only one chance to backend (if health
> > check ever occurs before child process failure to connect to the
> backend),
> > rendering retry settings effectively to be ignored. That's where this
> patch
> > comes into action - enables health check retries while child processes
> are
> > prevented to degenerate backend.
> >
> > I don't think, but I could be wrong, that this patch influences the
> > behavior we're seeing with unwanted health check attempt delays. Also,
> > knowing this, maybe pgpool could be patched or some other support be
> built
> > into it to cover this use case.
> >
> > Regards,
> > Stevo.
> >
> >
> > 2012/1/12 Tatsuo Ishii <ishii at postgresql.org>
> >
> >> I have accepted the moderation request. Your post should be sent
> shortly.
> >> Also I have raised the post size limit to 1MB.
> >> I will look into this...
> >> --
> >> Tatsuo Ishii
> >> SRA OSS, Inc. Japan
> >> English: http://www.sraoss.co.jp/index_en.php
> >> Japanese: http://www.sraoss.co.jp
> >>
> >> > Here is the log file and strace output file (this time in an archive,
> >> > didn't know about 200KB constraint on post size which requires
> moderator
> >> > approval). Timings configured are 30sec health check interval, 5sec
> >> > timeout, and 2 retries with 10sec retry delay.
> >> >
> >> > It takes a lot more than 5sec from started health check to sleeping
> 10sec
> >> > for first retry.
> >> >
> >> > Seen in code (main.x, health_check() function), within (retry) attempt
> >> > there is inner retry (first with postgres database then with
> template1)
> >> and
> >> > that part doesn't seem to be interrupted by alarm.
> >> >
> >> > Regards,
> >> > Stevo.
> >> >
> >> > 2012/1/12 Stevo Slavić <sslavic at gmail.com>
> >> >
> >> >> Here is the log file and strace output file. Timings configured are
> >> 30sec
> >> >> health check interval, 5sec timeout, and 2 retries with 10sec retry
> >> delay.
> >> >>
> >> >> It takes a lot more than 5sec from started health check to sleeping
> >> 10sec
> >> >> for first retry.
> >> >>
> >> >> Seen in code (main.x, health_check() function), within (retry)
> attempt
> >> >> there is inner retry (first with postgres database then with
> template1)
> >> and
> >> >> that part doesn't seem to be interrupted by alarm.
> >> >>
> >> >> Regards,
> >> >> Stevo.
> >> >>
> >> >>
> >> >> 2012/1/11 Tatsuo Ishii <ishii at postgresql.org>
> >> >>
> >> >>> Ok, I will do it. In the mean time you could use "strace -tt -p PID"
> >> >>> to see which system call is blocked.
> >> >>> --
> >> >>> Tatsuo Ishii
> >> >>> SRA OSS, Inc. Japan
> >> >>> English: http://www.sraoss.co.jp/index_en.php
> >> >>> Japanese: http://www.sraoss.co.jp
> >> >>>
> >> >>> > OK, got the info - key point is that ip forwarding is disabled for
> >> >>> security
> >> >>> > reasons. Rules in iptables are not important, iptables can be
> >> stopped,
> >> >>> or
> >> >>> > previously added rules removed.
> >> >>> >
> >> >>> > Here are the steps to reproduce (kudos to my colleague Nenad
> >> Bulatovic
> >> >>> for
> >> >>> > providing this):
> >> >>> >
> >> >>> > 1.) make sure that ip forwarding is off:
> >> >>> >     echo 0 > /proc/sys/net/ipv4/ip_forward
> >> >>> > 2.) create IP alias on some interface (and have postgres listen on
> >> it):
> >> >>> >     ip addr add x.x.x.x/yy dev ethz
> >> >>> > 3.) set backend_hostname0 to aforementioned IP
> >> >>> > 4.) start pgpool and monitor health checks
> >> >>> > 5.) remove IP alias:
> >> >>> >     ip addr del x.x.x.x/yy dev ethz
> >> >>> >
> >> >>> >
> >> >>> > Here is the interesting part in pgpool log after this:
> >> >>> > 2012-01-11 17:38:04 DEBUG: pid 24358: starting health checking
> >> >>> > 2012-01-11 17:38:04 DEBUG: pid 24358: health_check: 0 th DB node
> >> >>> status: 2
> >> >>> > 2012-01-11 17:38:04 DEBUG: pid 24358: health_check: 1 th DB node
> >> >>> status: 1
> >> >>> > 2012-01-11 17:38:34 DEBUG: pid 24358: starting health checking
> >> >>> > 2012-01-11 17:38:34 DEBUG: pid 24358: health_check: 0 th DB node
> >> >>> status: 2
> >> >>> > 2012-01-11 17:41:43 DEBUG: pid 24358: health_check: 0 th DB node
> >> >>> status: 2
> >> >>> > 2012-01-11 17:41:46 ERROR: pid 24358: health check failed. 0 th
> host
> >> >>> > 192.168.2.27 at port 5432 is down
> >> >>> > 2012-01-11 17:41:46 LOG:   pid 24358: health check retry sleep
> time:
> >> 10
> >> >>> > second(s)
> >> >>> >
> >> >>> > That pgpool was configured with health check interval of 30sec,
> 5sec
> >> >>> > timeout, and 10sec retry delay with 2 max retries.
> >> >>> >
> >> >>> > Making use of libpq instead for connecting to db in health checks
> IMO
> >> >>> > should resolve it, but you'll best determine which call exactly
> gets
> >> >>> > blocked waiting. Btw, psql with PGCONNECT_TIMEOUT env var
> configured
> >> >>> > respects that env var timeout.
> >> >>> >
> >> >>> > Regards,
> >> >>> > Stevo.
> >> >>> >
> >> >>> > On Wed, Jan 11, 2012 at 11:15 AM, Stevo Slavić <sslavic at gmail.com
> >
> >> >>> wrote:
> >> >>> >
> >> >>> >> Tatsuo,
> >> >>> >>
> >> >>> >> Did you restart iptables after adding rule?
> >> >>> >>
> >> >>> >> Regards,
> >> >>> >> Stevo.
> >> >>> >>
> >> >>> >>
> >> >>> >> On Wed, Jan 11, 2012 at 11:12 AM, Stevo Slavić <
> sslavic at gmail.com>
> >> >>> wrote:
> >> >>> >>
> >> >>> >>> Looking into this to verify if these are all necessary changes
> to
> >> have
> >> >>> >>> port unreachable message silently rejected (suspecting some
> kernel
> >> >>> >>> parameter tuning is needed).
> >> >>> >>>
> >> >>> >>> Just to clarify it's not a problem that host is being detected
> by
> >> >>> pgpool
> >> >>> >>> to be down, but the timing when that happens. On environment
> where
> >> >>> issue is
> >> >>> >>> reproduced pgpool as part of health check attempt tries to
> connect
> >> to
> >> >>> >>> backend and hangs for tcp timeout instead of being interrupted
> by
> >> >>> timeout
> >> >>> >>> alarm. Can you verify/confirm please the health check retry
> timings
> >> >>> are not
> >> >>> >>> delayed?
> >> >>> >>>
> >> >>> >>> Regards,
> >> >>> >>> Stevo.
> >> >>> >>>
> >> >>> >>>
> >> >>> >>> On Wed, Jan 11, 2012 at 10:50 AM, Tatsuo Ishii <
> >> ishii at postgresql.org
> >> >>> >wrote:
> >> >>> >>>
> >> >>> >>>> Ok, I did:
> >> >>> >>>>
> >> >>> >>>> # iptables -A FORWARD -j REJECT --reject-with
> >> icmp-port-unreachable
> >> >>> >>>>
> >> >>> >>>> on the host where pgpoo is running. And pull network cable from
> >> >>> >>>> backend0 host network interface. Pgpool detected the host being
> >> down
> >> >>> >>>> as expected...
> >> >>> >>>> --
> >> >>> >>>> Tatsuo Ishii
> >> >>> >>>> SRA OSS, Inc. Japan
> >> >>> >>>> English: http://www.sraoss.co.jp/index_en.php
> >> >>> >>>> Japanese: http://www.sraoss.co.jp
> >> >>> >>>>
> >> >>> >>>> > Backend is not destination of this message, pgpool host is,
> and
> >> we
> >> >>> >>>> don't
> >> >>> >>>> > want it to ever get it. With command I've sent you rule will
> be
> >> >>> >>>> created for
> >> >>> >>>> > any source and destination.
> >> >>> >>>> >
> >> >>> >>>> > Regards,
> >> >>> >>>> > Stevo.
> >> >>> >>>> >
> >> >>> >>>> > On Wed, Jan 11, 2012 at 10:38 AM, Tatsuo Ishii <
> >> >>> ishii at postgresql.org>
> >> >>> >>>> wrote:
> >> >>> >>>> >
> >> >>> >>>> >> I did following:
> >> >>> >>>> >>
> >> >>> >>>> >> Do following on the host where pgpool is running on:
> >> >>> >>>> >>
> >> >>> >>>> >> # iptables -A FORWARD -j REJECT --reject-with
> >> >>> icmp-port-unreachable -d
> >> >>> >>>> >> 133.137.177.124
> >> >>> >>>> >> (133.137.177.124 is the host where backend is running on)
> >> >>> >>>> >>
> >> >>> >>>> >> Pull network cable from backend0 host network interface.
> Pgpool
> >> >>> >>>> >> detected the host being down as expected. Am I missing
> >> something?
> >> >>> >>>> >> --
> >> >>> >>>> >> Tatsuo Ishii
> >> >>> >>>> >> SRA OSS, Inc. Japan
> >> >>> >>>> >> English: http://www.sraoss.co.jp/index_en.php
> >> >>> >>>> >> Japanese: http://www.sraoss.co.jp
> >> >>> >>>> >>
> >> >>> >>>> >> > Hello Tatsuo,
> >> >>> >>>> >> >
> >> >>> >>>> >> > With backend0 on one host just configure following rule on
> >> other
> >> >>> >>>> host
> >> >>> >>>> >> where
> >> >>> >>>> >> > pgpool is:
> >> >>> >>>> >> >
> >> >>> >>>> >> > iptables -A FORWARD -j REJECT --reject-with
> >> >>> icmp-port-unreachable
> >> >>> >>>> >> >
> >> >>> >>>> >> > and then have pgpool startup with health checking and
> >> retrying
> >> >>> >>>> >> configured,
> >> >>> >>>> >> > and then pull network cable from backend0 host network
> >> >>> interface.
> >> >>> >>>> >> >
> >> >>> >>>> >> > Regards,
> >> >>> >>>> >> > Stevo.
> >> >>> >>>> >> >
> >> >>> >>>> >> > On Wed, Jan 11, 2012 at 6:27 AM, Tatsuo Ishii <
> >> >>> ishii at postgresql.org
> >> >>> >>>> >
> >> >>> >>>> >> wrote:
> >> >>> >>>> >> >
> >> >>> >>>> >> >> I want to try to test the situation you descrived:
> >> >>> >>>> >> >>
> >> >>> >>>> >> >> >> > When system is configured for security reasons not
> to
> >> >>> return
> >> >>> >>>> >> >> destination
> >> >>> >>>> >> >> >> > host unreachable messages, even though
> >> >>> health_check_timeout is
> >> >>> >>>> >> >>
> >> >>> >>>> >> >> But I don't know how to do it. I pulled out the network
> >> cable
> >> >>> and
> >> >>> >>>> >> >> pgpool detected it as expected. Also I configured the
> server
> >> >>> which
> >> >>> >>>> >> >> PostgreSQL is running on to disable the 5432 port. In
> this
> >> case
> >> >>> >>>> >> >> connect(2) returned EHOSTUNREACH (No route to host) so
> >> pgpool
> >> >>> >>>> detected
> >> >>> >>>> >> >> the error as expected.
> >> >>> >>>> >> >>
> >> >>> >>>> >> >> Could you please instruct me?
> >> >>> >>>> >> >> --
> >> >>> >>>> >> >> Tatsuo Ishii
> >> >>> >>>> >> >> SRA OSS, Inc. Japan
> >> >>> >>>> >> >> English: http://www.sraoss.co.jp/index_en.php
> >> >>> >>>> >> >> Japanese: http://www.sraoss.co.jp
> >> >>> >>>> >> >>
> >> >>> >>>> >> >> > Hello Tatsuo,
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> > Thank you for replying!
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> > I'm not sure what exactly is blocking, just by pgpool
> code
> >> >>> >>>> analysis I
> >> >>> >>>> >> >> > suspect it is the part where a connection is made to
> the
> >> db
> >> >>> and
> >> >>> >>>> it
> >> >>> >>>> >> >> doesn't
> >> >>> >>>> >> >> > seem to get interrupted by alarm. Tested thoroughly
> health
> >> >>> check
> >> >>> >>>> >> >> behaviour,
> >> >>> >>>> >> >> > it works really well when host/ip is there and just
> >> >>> >>>> backend/postgres
> >> >>> >>>> >> is
> >> >>> >>>> >> >> > down, but not when backend host/ip is down. I could
> see in
> >> >>> log
> >> >>> >>>> that
> >> >>> >>>> >> >> initial
> >> >>> >>>> >> >> > health check and each retry got delayed when host/ip is
> >> not
> >> >>> >>>> reachable,
> >> >>> >>>> >> >> > while when just backend is not listening (is down) on
> the
> >> >>> >>>> reachable
> >> >>> >>>> >> >> host/ip
> >> >>> >>>> >> >> > then initial health check and all retries are exact to
> the
> >> >>> >>>> settings in
> >> >>> >>>> >> >> > pgpool.conf.
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> > PGCONNECT_TIMEOUT is listed as one of the libpq
> >> environment
> >> >>> >>>> variables
> >> >>> >>>> >> in
> >> >>> >>>> >> >> > the docs (see
> >> >>> >>>> >> >>
> http://www.postgresql.org/docs/9.1/static/libpq-envars.html)
> >> >>> >>>> >> >> > There is equivalent parameter in libpq
> PGconnectdbParams (
> >> >>> see
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >>
> >> >>> >>>> >>
> >> >>> >>>>
> >> >>>
> >>
> http://www.postgresql.org/docs/9.1/static/libpq-connect.html#LIBPQ-CONNECT-CONNECT-TIMEOUT
> >> >>> >>>> >> >> )
> >> >>> >>>> >> >> > At the beginning of that same page there are some
> >> important
> >> >>> >>>> infos on
> >> >>> >>>> >> >> using
> >> >>> >>>> >> >> > these functions.
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> > psql respects PGCONNECT_TIMEOUT.
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> > Regards,
> >> >>> >>>> >> >> > Stevo.
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> > On Wed, Jan 11, 2012 at 12:13 AM, Tatsuo Ishii <
> >> >>> >>>> ishii at postgresql.org>
> >> >>> >>>> >> >> wrote:
> >> >>> >>>> >> >> >
> >> >>> >>>> >> >> >> > Hello pgpool community,
> >> >>> >>>> >> >> >> >
> >> >>> >>>> >> >> >> > When system is configured for security reasons not
> to
> >> >>> return
> >> >>> >>>> >> >> destination
> >> >>> >>>> >> >> >> > host unreachable messages, even though
> >> >>> health_check_timeout is
> >> >>> >>>> >> >> >> configured,
> >> >>> >>>> >> >> >> > socket call will block and alarm will not get raised
> >> >>> until TCP
> >> >>> >>>> >> timeout
> >> >>> >>>> >> >> >> > occurs.
> >> >>> >>>> >> >> >>
> >> >>> >>>> >> >> >> Interesting. So are you saying that read(2) cannot be
> >> >>> >>>> interrupted by
> >> >>> >>>> >> >> >> alarm signal if the system is configured not to return
> >> >>> >>>> destination
> >> >>> >>>> >> >> >> host unreachable message? Could you please guide me
> >> where I
> >> >>> can
> >> >>> >>>> get
> >> >>> >>>> >> >> >> such that info? (I'm not a network expert).
> >> >>> >>>> >> >> >>
> >> >>> >>>> >> >> >> > Not a C programmer, found some info that select call
> >> >>> could be
> >> >>> >>>> >> replace
> >> >>> >>>> >> >> >> with
> >> >>> >>>> >> >> >> > select/pselect calls. Maybe it would be best if
> >> >>> >>>> PGCONNECT_TIMEOUT
> >> >>> >>>> >> >> value
> >> >>> >>>> >> >> >> > could be used here for connection timeout. pgpool
> has
> >> >>> libpq as
> >> >>> >>>> >> >> >> dependency,
> >> >>> >>>> >> >> >> > why isn't it using libpq for the healthcheck db
> connect
> >> >>> >>>> calls, then
> >> >>> >>>> >> >> >> > PGCONNECT_TIMEOUT would be applied?
> >> >>> >>>> >> >> >>
> >> >>> >>>> >> >> >> I don't think libpq uses select/pselect for
> establishing
> >> >>> >>>> connection,
> >> >>> >>>> >> >> >> but using libpq instead of homebrew code seems to be
> an
> >> >>> idea.
> >> >>> >>>> Let me
> >> >>> >>>> >> >> >> think about it.
> >> >>> >>>> >> >> >>
> >> >>> >>>> >> >> >> One question. Are you sure that libpq can deal with
> the
> >> case
> >> >>> >>>> (not to
> >> >>> >>>> >> >> >> return destination host unreachable messages) by using
> >> >>> >>>> >> >> >> PGCONNECT_TIMEOUT?
> >> >>> >>>> >> >> >> --
> >> >>> >>>> >> >> >> Tatsuo Ishii
> >> >>> >>>> >> >> >> SRA OSS, Inc. Japan
> >> >>> >>>> >> >> >> English: http://www.sraoss.co.jp/index_en.php
> >> >>> >>>> >> >> >> Japanese: http://www.sraoss.co.jp
> >> >>> >>>> >> >> >>
> >> >>> >>>> >> >>
> >> >>> >>>> >>
> >> >>> >>>>
> >> >>> >>>
> >> >>> >>>
> >> >>> >>
> >> >>>
> >> >>
> >> >>
> >>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.sraoss.jp/pipermail/pgpool-general/attachments/20120114/f6bc0bc3/attachment-0001.html>