Message-ID: <49F01D79.8020805@enterprisedb.com>
Date: Thu, 23 Apr 2009 10:49:13 +0300
From: Heikki Linnakangas <heikki.linnakangas@enterprisedb.com>
Organization: EnterpriseDB
User-Agent: Mozilla-Thunderbird 2.0.0.19 (X11/20090103)
MIME-Version: 1.0
To: Fujii Masao <masao.fujii@gmail.com>
CC: Simon Riggs <simon@2ndquadrant.com>,
	Guillaume Smet <guillaume.smet@gmail.com>,
	Tom Lane <tgl@sss.pgh.pa.us>, Peter Eisentraut <peter_e@gmx.net>,
	Kevin Grittner <Kevin.Grittner@wicourts.gov>,
	pgsql-hackers <pgsql-hackers@postgresql.org>
Subject: Re: New trigger option of pg_standby
References: <3f0b79eb0903242329j12865d55s348f5c873a956e71@mail.gmail.com>	
	<1d4e0c10904071456m735e3f67q123b87243d2f6f2a@mail.gmail.com>	
	<3f0b79eb0904071817u5c32254ei61f7e68e54669d9e@mail.gmail.com>	
	<49DDEE72.6030803@enterprisedb.com>	
	<3f0b79eb0904092047w355e0b3ar527a9285b161ef91@mail.gmail.com>	
	<1239381103.16396.30.camel@ebony.2ndQuadrant>	
	<3f0b79eb0904122252t1f4c5860v919d199f399ccf72@mail.gmail.com>	
	<3f0b79eb0904132241p18084914g792cb8e5ab3eebb3@mail.gmail.com>	
	<3f0b79eb0904150101pb1c59a3ubb422867e8e6a944@mail.gmail.com>	
	<49EE1E3F.9050307@enterprisedb.com>
	<3f0b79eb0904212209i7fbd8b68k71db1fef09c62751@mail.gmail.com>
In-Reply-To: <3f0b79eb0904212209i7fbd8b68k71db1fef09c62751@mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Fujii Masao wrote:
> On Wed, Apr 22, 2009 at 4:27 AM, Heikki Linnakangas
> <heikki.linnakangas@enterprisedb.com> wrote:
>> Fujii Masao wrote:
>>> On Tue, Apr 14, 2009 at 2:41 PM, Fujii Masao <masao.fujii@gmail.com>
>>> wrote:
>>>> I'd like to propose another simple idea; pg_standby deletes the
>>>> trigger file *whenever* the nextWALfile is a timeline history file.
>>>> A timeline history file is restored at the end of recovery, so it's
>>>> guaranteed that the trigger file is deleted whether nextWALfile
>>>> exists or not.
>>>>
>>>> A timeline history file is restored also at the beginning of
>>>> recovery, so the accidentally remaining trigger file is deleted
>>>> in early warm-standby as a side-effect of this idea.
>>> Here is the revised patch as above.
>>>
>>> If you notice something, please feel free to comment.
>> Ok, looking at this in more detail now. A couple of small things:
>>
>> We mustn't remove the trigger file immediately even in fast mode. As noted
>> elsewhere in this thread, we have the same bug in fast mode where pg_standby
>> gets stuck if you copy WAL files directly into pg_xlog.
> 
> Yes, there is the same problem also in fast mode. But, in fast
> mode, the trigger file has to be deleted immediately if it's found.
> Otherwise, recovery may fail as follows.
> 
> 1. pg_standby finds the trigger file for fast mode, and returns
>     non-zero without deleting the trigger file.
> 2. the startup process tries to read the WAL file from pg_xlog,
>     but it's not found.
> 3. the startup process tries to restore the last applied WAL file
>     using pg_standby.
> 4. (Again) pg_standby finds the trigger file for fast mode, and
>     returns non-zero without deleting the trigger file.
> 5. the startup process tries to read the last applied WAL file,
>     but it's not found.
>     (though the last applied file was of course restored before,
>      the restored one cannot be read here)
> 6. recovery fails because the last applied WAL file cannot be
>     read.
> 
> On the other hand, if pg_standby returns 0 also in fast mode
> when the requested file and trigger file exist, ISTM that there
> is not much difference between fast and smart mode; also in
> fast mode, all the available WAL files would be applied.

Hmm, pg_standby could truncate the trigger file, so that it acts like a 
smart trigger in the subsequent pg_standby invocations. Assuming the 
postgres user has write access to it; it probably does because it can 
delete it, but conceivably it has only read access on the file but write 
access on the directory it's in.

This is getting complicated, though. I guess it would be enough to 
document that you mustn't copy any extra files into pg_xlog if you use a 
fast trigger.

-- 
   Heikki Linnakangas
   EnterpriseDB   http://www.enterprisedb.com