Message-ID: <4D2126A4.9050203@2ndquadrant.com>
Date: Mon, 03 Jan 2011 02:30:12 +0100
From: Hannu Krosing <hannu@2ndquadrant.com>
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 6.1; en-US;
	rv:1.9.2.13) Gecko/20101207 Thunderbird/3.1.7
MIME-Version: 1.0
To: pgsql-hackers@postgresql.org
Subject: Re: Sync Rep Design
References: <1293729972.1892.23804.camel@ebony>
	<4D1CC47E.7050207@kaltenbrunner.cc>	<1293735853.1892.24886.camel@ebony>
	<AANLkTi=SvN2KQ1Tzh-tRKcXD0i2FyWFBbbR=kXXdepqN@mail.gmail.com>
	<4D1D8B2F.5070001@2ndquadrant.com>	<4D1DAB14.5000800@enterprisedb.com>
	<1293796137.1892.37313.camel@ebony>	<4D1DCF5A.7070808@enterprisedb.com>
	<4D1E4897.1050400@2ndquadrant.com>	<1293882875.1892.56143.camel@ebony>
	<AANLkTi=yYKvRCvfTprbj=jQDVMrDKUMcate=vCaQobSi@mail.gmail.com>
In-Reply-To: <AANLkTi=yYKvRCvfTprbj=jQDVMrDKUMcate=vCaQobSi@mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

On 2.1.2011 5:36, Robert Haas wrote:
> On Sat, Jan 1, 2011 at 6:54 AM, Simon Riggs<simon@2ndquadrant.com>  wrote:
>> Yes, working out the math is a good idea. Things are much clearer if we
>> do that.
>>
>> Let's assume we have 98% availability on any single server.
>>
>> 1. Having one primary and 2 standbys, either of which can acknowledge,
>> and we never lock up if both standbys fail, then we will have 99.9992%
>> server availability. (So PostgreSQL hits "5 Nines", with data
>> guarantees). ("Maximised availability")
> I don't agree with this math.  If the master and one standby fail
> simultaneously, the other standby is useless, because it may or may
> not be caught up with the master.  You know that the last transaction
> acknowledged as committed by the master is on at least one of the two
> standbys, but you don't know which one, and so you can't safely
> promote the surviving standby.
> (If you are working in an environment where promoting the surviving
> standby when it's possibly not caught up is OK, then you don't need
> sync rep in the first place: you can just run async rep and get much
> better performance.)
> So the availability is 98% (you are up when the master is up) + 98%^2
> * 2% (you are up when both slaves are up and the master is down) =
> 99.92%.  If you had only a single standby, then you could be certain
> that any commit acknowledged by the master was on that standby.  Thus
> your availability would be 98% (up when master is up) + 98% * 2% (you
> are up when the master is down and the slave is up) = 99.96%.
>
OTOH, in the case where you need _all_ the slaves to confirm any failing 
slave brings
the master down, so adding a slave brings down availability by extra 2%

The solution to achieving good durability AND availability is requiring 
N past the
post instead of 1 past the post.

In this case you can get to 99.9992% availability with master + 3 sync 
slaves, 2 of which have ACK.

---------------------------------------
Hannu Krosing
Performance and Infinite Scalability Consultant
http://www.2ndQuadrant.com/books/