Re: [HACKERS] Quorum commit for multiple synchronous replication. - Mailing list pgsql-hackers

On Tue, Apr 18, 2017 at 6:40 PM, Kyotaro HORIGUCHI
<horiguchi.kyotaro@lab.ntt.co.jp> wrote:
> At Tue, 18 Apr 2017 14:58:50 +0900, Masahiko Sawada <sawada.mshk@gmail.com> wrote in
<CAD21AoBqSjUGx0LCDrjEDLB-yx2EvgLMdT8Nz4ZR_xpxrbMU+Q@mail.gmail.com>
>> On Tue, Apr 18, 2017 at 3:04 AM, Fujii Masao <masao.fujii@gmail.com> wrote:
>> > On Wed, Apr 12, 2017 at 2:36 AM, Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>> >> On Thu, Apr 6, 2017 at 4:17 PM, Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>> >>> On Thu, Apr 6, 2017 at 10:51 AM, Noah Misch <noah@leadboat.com> wrote:
>> >>>> On Thu, Apr 06, 2017 at 12:48:56AM +0900, Fujii Masao wrote:
>> >>>>> On Wed, Apr 5, 2017 at 3:45 PM, Noah Misch <noah@leadboat.com> wrote:
>> >>>>> > On Mon, Dec 19, 2016 at 09:49:58PM +0900, Fujii Masao wrote:
>> >>>>> >> Regarding this feature, there are some loose ends. We should work on
>> >>>>> >> and complete them until the release.
>> >>>>> >>
>> >>>>> >> (1)
>> >>>>> >> Which synchronous replication method, priority or quorum, should be
>> >>>>> >> chosen when neither FIRST nor ANY is specified in s_s_names? Right now,
>> >>>>> >> a priority-based sync replication is chosen for keeping backward
>> >>>>> >> compatibility. However some hackers argued to change this decision
>> >>>>> >> so that a quorum commit is chosen because they think that most users
>> >>>>> >> prefer to a quorum.
>> >>>>> >>
>> >>>>> >> (2)
>> >>>>> >> There will be still many source comments and documentations that
>> >>>>> >> we need to update, for example, in high-availability.sgml. We need to
>> >>>>> >> check and update them throughly.
>> >>>>> >>
>> >>>>> >> (3)
>> >>>>> >> The priority value is assigned to each standby listed in s_s_names
>> >>>>> >> even in quorum commit though those priority values are not used at all.
>> >>>>> >> Users can see those priority values in pg_stat_replication.
>> >>>>> >> Isn't this confusing? If yes, it might be better to always assign 1 as
>> >>>>> >> the priority, for example.
>> >>>>> >
>> >>>>> > [Action required within three days.  This is a generic notification.]
>> >>>>> >
>> >>>>> > The above-described topic is currently a PostgreSQL 10 open item.  Fujii,
>> >>>>> > since you committed the patch believed to have created it, you own this open
>> >>>>> > item.  If some other commit is more relevant or if this does not belong as a
>> >>>>> > v10 open item, please let us know.  Otherwise, please observe the policy on
>> >>>>> > open item ownership[1] and send a status update within three calendar days of
>> >>>>> > this message.  Include a date for your subsequent status update.  Testers may
>> >>>>> > discover new open items at any time, and I want to plan to get them all fixed
>> >>>>> > well in advance of shipping v10.  Consequently, I will appreciate your efforts
>> >>>>> > toward speedy resolution.  Thanks.
>> >>>>> >
>> >>>>> > [1] https://www.postgresql.org/message-id/20170404140717.GA2675809%40tornado.leadboat.com
>> >>>>>
>> >>>>> Thanks for the notice!
>> >>>>>
>> >>>>> Regarding the item (2), Sawada-san told me that he will work on it after
>> >>>>> this CommitFest finishes. So we would receive the patch for the item from
>> >>>>> him next week. If there will be no patch even after the end of next week
>> >>>>> (i.e., April 14th), I will. Let's wait for Sawada-san's action at first.
>> >>>>
>> >>>> Sounds reasonable; I will look for your update on 14Apr or earlier.
>> >>>>
>> >>>>> The items (1) and (3) are not bugs. So I don't think that they need to be
>> >>>>> resolved before the beta release. After the feature freeze, many users
>> >>>>> will try and play with many new features including quorum-based syncrep.
>> >>>>> Then if many of them complain about (1) and (3), we can change the code
>> >>>>> at that timing. So we need more time that users can try the feature.
>> >>>>
>> >>>> I've moved (1) to a new section for things to revisit during beta.  If someone
>> >>>> feels strongly that the current behavior is Wrong and must change, speak up as
>> >>>> soon as you reach that conclusion.  Absent such arguments, the behavior won't
>> >>>> change.
>> >>>>
>> >>>>> BTW, IMO (3) should be fixed so that pg_stat_replication reports NULL
>> >>>>> as the priority if quorum-based sync rep is chosen. It's less confusing.
>> >>>>
>> >>>> Since you do want (3) to change, please own it like any other open item,
>> >>>> including the mandatory status updates.
>> >>>
>> >>> I agree to report NULL as the priority. I'll send a patch for this as well.
>> >>>
>> >>> Regards,
>> >>>
>> >>
>> >> Attached two draft patches. The one makes pg_stat_replication.sync
>> >> priority report NULL if in quorum-based sync replication. To prevent
>> >> extra change I don't change so far the code of setting standby
>> >> priority. The another one improves the comment and documentation. If
>> >> there is more thing what we need to mention in documentation please
>> >> give me feedback.
>> >
>> > Attached is the modified version of the doc improvement patch.
>> > Barring any objection, I will commit this version.
>>
>> Thank you for updating the patch.
>>
>> >
>> > +    In term of performance there is difference between two synchronous
>> > +    replication method. Generally quorum-based synchronous replication
>> > +    tends to be higher performance than priority-based synchronous
>> > +    replication. Because in quorum-based synchronous replication, the
>> > +    transaction can resume as soon as received the specified number of
>> > +    acknowledgement from synchronous standby servers without distinction
>> > +    of standby servers. On the other hand in priority-based synchronous
>> > +    replication, the standby server that the primary server must wait for
>> > +    is fixed until a synchronous standby fails. Therefore, if a server on
>> > +    low-performance machine a has high priority and is chosen as a
>> > +    synchronous standby server it can reduce performance for database
>> > +    applications.
>> >
>> > This description looks misleading. A quorum-based sync rep is basically
>> > more efficient when there are multiple standbys in s_s_names and you want
>> > to replicate the transactions to some of them synchronously. I think that
>> > this assumption should be documented explicitly. So I modified this
>> > description. Please see the modified version in the attached patch.
>>
>> You're right. The modified version looks good to me, thanks.
>
> It looks better to me, too. But (even I'm not sure, of course)
> the sentences seem to need improvement.
>
> |    <para>
> |     Quorum-based synchronous replication is basically more
> |     efficient than priority-based one when you specify multiple
> |     standbys in <varname>synchronous_standby_names</> and want
> |     to synchronously replicate transactions to two or more of
> |     them. In the priority-based case, the replication master
> |     must wait for a reply from the slowest standby in the
> |     required number of standbys in priority order, which may
> |     slower than the rest.

I supposed that Fujii-san pointed out that quorum-based sync
replication could be more efficient when we want to replicate the
transaction to "part of" standbys listed in s_s_names. So I guess it's
not good idea to mention "two or more of them" which also can mean the
all of standbys.

> On the other hand, quorum-based
> |     synchronous replication may reduce the latency because it
> |     allows transactions to wait only for replies from a
> |     required number of fastest standbys in all the listed
> |     standbys, i.e., such slow standby doesn't block
> |     transactions.
> |  </para>
>
> I'm not sure that this is actually an improvement..
>



--
Masahiko Sawada
NIPPON TELEGRAPH AND TELEPHONE CORPORATION
NTT Open Source Software Center