Replication on the backend

gustavotonini@gmail.com (Gustavo Tonini) writes:

What about replication or data distribution inside the backend. This
is a valid issue?

I'm not sure what your question is...
--
(reverse (concatenate 'string "gro.gultn" "@" "enworbbc"))
http://www.ntlug.org/~cbbrowne/x.html
"Love is like a snowmobile flying over the frozen tundra that suddenly
flips, pinning you underneath. At night, the ice weasels come."
-- Matt Groening

replication (master/slave, multi-master, etc) implemented inside
postgres...I would like to know what has been make in this area.

Gustavo.

P.S. Sorry for my bad English.

2005/12/5, Chris Browne <cbbrowne@acm.org>:

Gustavo Tonini wrote:

replication (master/slave, multi-master, etc) implemented inside
postgres...I would like to know what has been make in this area.

http://www.commandprompt.com/ - Master/Slave

Joshua D. Drake

replication (master/slave, multi-master, etc) implemented inside
postgres...I would like to know what has been make in this area.

It's not in the backend, check out things like Slony (www.slony.info)
and various other commercial solutions.

Chris

On 12/5/2005 8:18 PM, Gustavo Tonini wrote:

replication (master/slave, multi-master, etc) implemented inside
postgres...I would like to know what has been make in this area.

We do not plan to implement replication inside the backend. Replication
needs are so diverse that pluggable replication support makes a lot more
sense. To me it even makes more sense than keeping transaction support
outside of the database itself and add it via pluggable storage add-on.

Jan

Gustavo.

P.S. Sorry for my bad English.

2005/12/5, Chris Browne <cbbrowne@acm.org>:

gustavotonini@gmail.com (Gustavo Tonini) writes:

What about replication or data distribution inside the backend. This
is a valid issue?

I'm not sure what your question is...
--
(reverse (concatenate 'string "gro.gultn" "@" "enworbbc"))
http://www.ntlug.org/~cbbrowne/x.html
"Love is like a snowmobile flying over the frozen tundra that suddenly
flips, pinning you underneath. At night, the ice weasels come."
-- Matt Groening

---------------------------(end of broadcast)---------------------------
TIP 1: if posting/reading through Usenet, please send an appropriate
subscribe-nomail command to majordomo@postgresql.org so that your
message can get through to the mailing list cleanly

--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #

But, wouldn't the performance be better? And wouldn't asynchronous messages
be better processed?

Thanks for replies,
Gustavo.

2005/12/6, Jan Wieck <JanWieck@yahoo.com>:

On Tue, 2005-12-06 at 10:03 -0200, Gustavo Tonini wrote:

But, wouldn't the performance be better? And wouldn't asynchronous
messages be better processed?

At least for synchronous multi-master replication, the performance
bottelneck is going to be the interconnect between the nodes -
integration of the replication logic into the backend most probably
doesn't affect performance that much.

I'd rather like to ask Jan what different needs for replication he
discovered so far. And how he came to the conclusion, that it's not
possible to provide a general solution.

My point for integration into the backend is flexibility: obviously the
replication code can influence the database much more from within the
backend than from the outside. For example running one complex query on
several nodes. I know, this a very advanced feature - currently it's not
even possible to run one query on multiple backends (i.e. processors of
a multi-core system) - but I like to plan ahead instead of throwing away
code later. For such advanced features you simply have to dig around in
the backend code one day. Of course you can always add hooks, but IMHO
that only complicates matters.

Is there some discussion going on about such topics somewhere? What's up
with slony-2? The wiki on slony2.org still doesn't provide a lot of
technical information (and obviously got spammed BTW).

Regards

Markus

On 12/6/2005 8:10 AM, Markus Schiltknecht wrote:

On Tue, 2005-12-06 at 10:03 -0200, Gustavo Tonini wrote:

But, wouldn't the performance be better? And wouldn't asynchronous
messages be better processed?

At least for synchronous multi-master replication, the performance
bottelneck is going to be the interconnect between the nodes -
integration of the replication logic into the backend most probably
doesn't affect performance that much.

That is exactly right. Thus far, processor, memory and disk speeds have
allways advanced on a higher pace than network speeds. Thus, the few
percent of performance gain we'd get from moving things into the backend
will be irrelevant tomorrow with 4x-core and 16x-core CPU's.

I'd rather like to ask Jan what different needs for replication he
discovered so far. And how he came to the conclusion, that it's not
possible to provide a general solution.

- Asynchronous master to multi-slave. We have a few of those with
Mommoth-Replicator and Slony-I being the top players. Slony-I does
need some cleanup and/or reimplementation after we have a general
pluggable replication API in place.

- Synchronous multimaster. There are certain attempts out there, like
Postgres-R, pgcluster, Slony-II. Some more advanced, some less. But
certainly nothing I would send into the ring against Oracle-Grid.

- Asynchronous multimaster with conflict resolution. I have not seen
any reasonable attempt on this one yet. Plus, it divides again into
two camps. One is the idea to have one central system with thousands
of satellites (salesman on the street), the other being two or more
central systems doing load balancing (although this competes with
sync-mm).

My point for integration into the backend is flexibility: obviously the
replication code can influence the database much more from within the

We need a general API. It should be possible to define on a per-database
level which shared replication module to load on connect. The init
function of that replication module then installs all the required
callbacks at strategic points (like heap_update(), at_commit() ...) and
the rest is hidden in the module.

Is there some discussion going on about such topics somewhere? What's up
with slony-2? The wiki on slony2.org still doesn't provide a lot of
technical information (and obviously got spammed BTW).

Slony-II has been slow lately in the Eastern timezone.

Jan

--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #

Hello Jan,

On Tue, 2005-12-06 at 10:10 -0500, Jan Wieck wrote:

We need a general API. It should be possible to define on a per-database
level which shared replication module to load on connect. The init
function of that replication module then installs all the required
callbacks at strategic points (like heap_update(), at_commit() ...) and
the rest is hidden in the module.

thank you for your list of replication types. Those still have some
things in common. Thus your approach of providing hooks for different
modules might make sense. Only I fear that I would need way to many
hooks for what I want ;)

Slony-II has been slow lately in the Eastern timezone.

What is that supposed to mean? Who sits in the eastern timezone?

Regards

Markus

gustavotonini@gmail.com (Gustavo Tonini) writes:

But,� wouldn't the performance be better? And wouldn't asynchronous
messages be better processed?

Why do you think performance would be materially affected by this?

The MAJOR performance bottleneck is normally the slow network
connection between servers.

When looked at in the perspective of that bottleneck, pretty much
everything else is just noise. (Sometimes pretty loud noise, but
still noise :-).)
--
let name="cbbrowne" and tld="cbbrowne.com" in name ^ "@" ^ tld;;
http://cbbrowne.com/info/spreadsheets.html
"When the grammar checker identifies an error, it suggests a
correction and can even makes some changes for you."
-- Microsoft Word for Windows 2.0 User's Guide, p.35:

IMO this is not true. You can get affordable 10GBit network adapters, so you can have plenty of bandwith in a db server pool (if they are located in the same area). Even 1GBit Ethernet greatly helps here, and would make it possible to balance read-intensive (and not write intensive) applications. We using linux bonding interface with 2 gbit NICs, and 200 MBytes/sec throughput is something you need to have a quite some harddisks to reach that. Latency is not bad too.

Regards,
Mario weilguni

-----Original Message-----
From: pgsql-hackers-owner@postgresql.org [mailto:pgsql-hackers-owner@postgresql.org] On Behalf Of Chris Browne
Sent: Tuesday, December 06, 2005 4:43 PM
To: pgsql-hackers@postgresql.org
Subject: Re: [HACKERS] Replication on the backend

gustavotonini@gmail.com (Gustavo Tonini) writes:

But,  wouldn't the performance be better? And wouldn't asynchronous
messages be better processed?

Why do you think performance would be materially affected by this?

The MAJOR performance bottleneck is normally the slow network
connection between servers.

When looked at in the perspective of that bottleneck, pretty much
everything else is just noise. (Sometimes pretty loud noise, but
still noise :-).)
--
let name="cbbrowne" and tld="cbbrowne.com" in name ^ "@" ^ tld;;
http://cbbrowne.com/info/spreadsheets.html
"When the grammar checker identifies an error, it suggests a
correction and can even makes some changes for you."
-- Microsoft Word for Windows 2.0 User's Guide, p.35:

---------------------------(end of broadcast)---------------------------
TIP 4: Have you searched our list archives?

http://archives.postgresql.org

Postgres-R, pgcluster, Slony-II. Some more advanced, some less. But
certainly nothing I would send into the ring against Oracle-Grid.

Assuming that you mean Oracle Real Application Cluster (the Grid is more,
right?) I wonder if this technology technically still counts as replication.
AFAIK they do not replicate data but share a common data pool among different
servers. You still have communication overhead but you write a tuple only
once for all servers involved. Takes away a lot of overhead on a system
that's heavily written too.

Michael
--
Michael Meskes
Email: Michael at Fam-Meskes dot De, Michael at Meskes dot (De|Com|Net|Org)
ICQ: 179140304, AIM/Yahoo: michaelmeskes, Jabber: meskes@jabber.org
Go SF 49ers! Go Rhein Fire! Use Debian GNU/Linux! Use PostgreSQL

Just like MySql!

On Dec 5, 2005, at 10:35 PM, Jan Wieck wrote:

- Asynchronous master to multi-slave. We have a few of those with
Mommoth-Replicator and Slony-I being the top players. Slony-I does
need some cleanup and/or reimplementation after we have a general
pluggable replication API in place.

Is this API actually have people working on it or just something on
the todo list?

I don't see anything in the TODO list. I'm very interesting in work that. If
is possible...

Gustavo.

I would classify it as a clustered database system (Oracle 10g that is).
Clustered meaning more than one node in the cluster.

ALy.

On Tue, 6 Dec 2005, Michael Meskes wrote:

Postgres-R, pgcluster, Slony-II. Some more advanced, some less. But
certainly nothing I would send into the ring against Oracle-Grid.

Assuming that you mean Oracle Real Application Cluster (the Grid is more,
right?) I wonder if this technology technically still counts as replication.
AFAIK they do not replicate data but share a common data pool among different
servers. You still have communication overhead but you write a tuple only
once for all servers involved. Takes away a lot of overhead on a system
that's heavily written too.

Michael

--
Aly S.P Dharshi
aly.dharshi@telus.net

"A good speech is like a good dress
that's short enough to be interesting
and long enough to cover the subject"

On 12/6/2005 11:23 AM, Mario Weilguni wrote:

IMO this is not true. You can get affordable 10GBit network adapters, so you can have plenty of bandwith in a db server pool (if they are located in the same area). Even 1GBit Ethernet greatly helps here, and would make it possible to balance read-intensive (and not write intensive) applications. We using linux bonding interface with 2 gbit NICs, and 200 MBytes/sec throughput is something you need to have a quite some harddisks to reach that. Latency is not bad too.

It's not so much the bandwidth but more the roundtrips that limit your
maximum transaction throughput. Remember, whatever the priority, you
can't increase the speed of light.

Jan

Regards,
Mario weilguni

-----Original Message-----
From: pgsql-hackers-owner@postgresql.org [mailto:pgsql-hackers-owner@postgresql.org] On Behalf Of Chris Browne
Sent: Tuesday, December 06, 2005 4:43 PM
To: pgsql-hackers@postgresql.org
Subject: Re: [HACKERS] Replication on the backend

gustavotonini@gmail.com (Gustavo Tonini) writes:

But, wouldn't the performance be better? And wouldn't asynchronous
messages be better processed?

Why do you think performance would be materially affected by this?

The MAJOR performance bottleneck is normally the slow network
connection between servers.

When looked at in the perspective of that bottleneck, pretty much
everything else is just noise. (Sometimes pretty loud noise, but
still noise :-).)

--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #

On 12/6/05, Jan Wieck <JanWieck@yahoo.com> wrote:

IMO this is not true. You can get affordable 10GBit network adapters, so you can have plenty of bandwith in a db server pool (if they are located in the same area). Even 1GBit Ethernet greatly helps here, and would make it possible to balance read-intensive (and not write intensive) applications. We using linux bonding interface with 2 gbit NICs, and 200 MBytes/sec throughput is something you need to have a quite some harddisks to reach that. Latency is not bad too.

It's not so much the bandwidth but more the roundtrips that limit your
maximum transaction throughput. Remember, whatever the priority, you
can't increase the speed of light.

Eh, why would light limited delay be any slower than a disk on FC the
same distance away? :)

In any case, performance of PG on iscsi is just fine. You can't blame
the network... Doing multimaster replication is hard because the
locking primitives that are fine on a simple multiprocessor system
(with a VERY high bandwidth very low latency interconnect between
processors) just don't work across a network, so you're left finding
other methods and making them work...

But again, multimaster isn't hard because there of some inherently
slow property of networks.

On Tue, 2005-12-06 at 23:19 -0500, Jan Wieck wrote:

It's not so much the bandwidth but more the roundtrips that limit your
maximum transaction throughput.

I completely agree that the latency is counting, not the bandwith.

Does anybody have latency / roundtrip measurements for current hardware?
I'm interested in:
1Gb Ethernet,
10 Gb Ethernet,
InfiniBand,
probably even p2p usb2 or firewire links?

At least Quadrics claims(1) to have measured only 1.38 microseconds.
Assuming real world condition would give you 5 microseconds, on a 3 GHz
processor that's 15'000 CPY cycles. Which is IMHO not that much any
more. Or am I wrong (mental arithmetic never was my favourite subject)?

Regards
Markus

[1]: http://www.quadrics.com/quadrics/QuadricsHome.nsf/NewsByDate/98FFE60F799AC95180256FEA002A6D9D
http://www.quadrics.com/quadrics/QuadricsHome.nsf/NewsByDate/98FFE60F799AC95180256FEA002A6D9D

On Dec 6, 2005, at 11:42 PM, Markus Schiltknecht wrote:

Does anybody have latency / roundtrip measurements for current
hardware?
I'm interested in:
1Gb Ethernet,
10 Gb Ethernet,
InfiniBand,
probably even p2p usb2 or firewire links?

In another secret life, I know a bit about supercomputing fabrics.
The latency metrics have to be thoroughly qualified.

First, most of the RTT latency numbers for network fabrics are for 0
byte packet sizes, which really does not apply to anyone shuffling
real data around. For small packets, high-performance fabrics (HTX
Infiniband, Quadrics, etc) have approximately an order of magnitude
less latency than vanilla Ethernet, though the performance specifics
depend greatly on the actual usage. For large packet sizes, the
differences in latency become far less obvious. However, for "real"
packets a performant fabric will still look very good compared to
disk systems. Switched fiber fabrics have enough relatively
inexpensive throughput now to saturate most disk systems and CPU I/O
busses; only platforms like HyperTransport can really keep up. It is
worth pointing out that the latency of high-end network fabrics is
similar to large NUMA fabrics, which exposes some of the limits of
SMP scalability. As a point of reference, an organization that knows
what they are doing should have no problem getting 500 microsecond
RTT on a vanilla metropolitan area GigE fiber network -- a few
network operators actually do deliver this on a regional scale. For
a local cluster, a competent design can best this by orders of
magnitude.

There are a number of silicon limitations, but a system that connects
the fabric directly to HyperTransport can drive several GB/s with
very respectable microsecond latencies if the rest of the system is
up to it. There are Opteron system boards now that will drive
Infiniband directly from HyperTransport. I know Arima/Rioworks makes
some (great server boards generally), and several other companies are
either making them or have announced them in the pipeline. These
Opteron boards get pretty damn close to Big Iron SMP fabric
performance in a cheap package. Given how many companies have
announced plans to produce Opteron server boards with Infiniband
fabrics directly integrated into HyperTransport, I would say that
this is the future of server boards.

And if postgres could actually use an infiniband fabric for
clustering a single database instance across Opteron servers, that
would be very impressive...

J. Andrew Rogers

On Wed, 2005-12-07 at 01:04 -0800, J. Andrew Rogers wrote:

Opteron boards get pretty damn close to Big Iron SMP fabric
performance in a cheap package. Given how many companies have
announced plans to produce Opteron server boards with Infiniband
fabrics directly integrated into HyperTransport, I would say that
this is the future of server boards.

InfiniBand on-board? Wow, seems very interesting. Thank you for your
hints and numbers, very helpfull!

And if postgres could actually use an infiniband fabric for
clustering a single database instance across Opteron servers, that
would be very impressive...

full ACK

Regards

Markus

On Dec 6, 2005, at 9:09 PM, Gregory Maxwell wrote:

Eh, why would light limited delay be any slower than a disk on FC the
same distance away? :)

In any case, performance of PG on iscsi is just fine. You can't blame
the network... Doing multimaster replication is hard because the
locking primitives that are fine on a simple multiprocessor system
(with a VERY high bandwidth very low latency interconnect between
processors) just don't work across a network, so you're left finding
other methods and making them work...

Speed of light latency shows up pretty damn often in real networks,
even relatively local ones. The number of people that wonder why a
transcontinental SLA of 10ms is not possible is astonishing. The
silicon fabrics are sufficiently fast that most well-designed
networks are limited by how fast one can push photons through a
fiber, which is significantly slower than photons through a vacuum.
Silicon switch fabrics add latency measured in nanoseconds, which is
effectively zero for many networks that leave the system board.

Compared to single system simple SMP, a local cluster built on a
first-rate fabric will have about an order of magnitude higher
latency but very similar bandwidth. On the other hand, at those
latencies you can increase the number of addressable processors with
that kind of bandwidth by an order of magnitude, so it is a bit of a
trade. However, latency matters a lot such that one would have to be
a lot smarter about partitioning synchronization across that fabric
even though one would lose nothing in the bandwidth department.

But again, multimaster isn't hard because there of some inherently
slow property of networks.

Eh? As far as I know, the difficulty of multi-master is almost
entirely a product of the latency of real networks such that they are
too slow for scalable distributed locks. SMP is little more than a
distributed lock manager implemented in silicon. Therefore, multi-
master is hard in practice because we cannot drive networks fast
enough. That said, current state-of-the-art network fabrics are
within an order of magnitude of SMP fabrics such that they could be
real contenders, particularly once you get north of 8-16 processors.

The really sweet potential is in Opteron system boards with
Infiniband directly attached to HyperTransport. At that level of
bandwidth and latency, both per node and per switch fabric, the
architecture possibilities start to become intriguing.

J. Andrew Rogers

Andrew,

And if postgres could actually use an infiniband fabric for
clustering a single database instance across Opteron servers, that
would be very impressive...

That's what we do with Bizgres MPP. We've implemented an interconnect to do the data shuffling underneath the optimizer/executor and we currently use TCP/IP, though we could haul out SDP over Infiniband should we need it.

However, our optimizer effectively minimizes traffic over the interconnect now and that works well for all of the plans we've run so far. It would be nice to characterize the improvements we could get from moving to 3x infiniband.

Regarding a direct Hypertransport to Infiniband bridge, have you looked at Pathscale? http://www.pathscale.com/ I know the fellow behind the scenes who designed it, and I think it's probably well thought out. We were gunning for less than 1us RTT through the adapter and switch once, and I bet they are close.

- Luke

On Tue, Dec 06, 2005 at 12:35:43AM -0500, Jan Wieck wrote:

We do not plan to implement replication inside the backend. Replication
needs are so diverse that pluggable replication support makes a lot more
sense. To me it even makes more sense than keeping transaction support
outside of the database itself and add it via pluggable storage add-on.

And, as I say every single time this comes up, Oracle's and IBM's and
MS's and everybody else's replication systems are _also_ add ons. If
you don't believe me, look at the license costs. You can get a
system without it enabled, which means (by definition) it's a modular
extension.

A

--
Andrew Sullivan | ajs@crankycanuck.ca
In the future this spectacle of the middle classes shocking the avant-
garde will probably become the textbook definition of Postmodernism.
--Brad Holland

Are you sure that no way to implement a generic aproach on the backend? What
does specification say? Does Oracle 10g have a core implementation of
replication (cluster)?

Gustavo.

2005/12/7, Andrew Sullivan <ajs@crankycanuck.ca>:

What is the point of these questions? If you have a concrete, practical
proposal to make, please do so. Otherwise, you have already got the
answer from the people who are actually working on replication and
understand it far beyond abstract considerations. If you think there is
a good reason to do replication directly in the backend code rather than
as an addon, possibly using an agreed API, then you need to provide hard
evidence, not mere assertion or conjecture.

cheers

andrew

Gustavo Tonini wrote:

On 12/8/2005 1:28 PM, Gustavo Tonini wrote:

Are you sure that no way to implement a generic aproach on the backend? What

You mean "generic" as in a replication system that can do asynchronous
master-slave, asynchronous multimaster with conflict resolution based on
timestamps, system priority or user defined resolution stubs, can do
synchronous predicate locking but also does support thousands of
asynchronous, partial replica (salesman on the road), and last but not
least can run as a synchronous cluster in a LAN segment. All the above
can of course be mixed together ... like a central cluster of 8 load
balanced, fault tolerant systems, with async multimaster replica in all
external branch servers, partial multimaster replica for the road
warriers and some slave leaf nodes for reporting.

If you can present a prototype for the above, I am sure that we will
change our opinion and finally settle for one, builtin replication system.

does specification say? Does Oracle 10g have a core implementation of
replication (cluster)?

What specification?

Jan

--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #

Are you sure that no way to implement a generic aproach on the
backend? What does specification say?

What specification are you talking about?

Does Oracle 10g have a core implementation of replication (cluster)?

Since replication is sold as a separate product from Oracle 10g,
obviously not.

We *know* (particularly those of us that have had involvement in
actually implementing replication systems used in production
environments) that "user space" implementations of replication can
function satisfactorily. We've implemented it.

You might be able to convince us that implementing replication inside
the backend is preferable, but you'll have to have exceedingly strong
evidence in order to overcome the already known factor that doing it
outside the backend works quite well.
--
let name="cbbrowne" and tld="gmail.com" in name ^ "@" ^ tld;;
http://linuxdatabases.info/info/slony.html
"Now, suddenly, I _am_ the expanding Russian Frontier" -- Commander
Ivanova
"But with very nice borders..." -- Dr Franklin

Hello,

On Fri, 2005-12-09 at 08:47 -0500, Christopher Browne wrote:

We *know* (particularly those of us that have had involvement in
actually implementing replication systems used in production
environments) that "user space" implementations of replication can
function satisfactorily. We've implemented it.

While this might be true, allow me a sidenote: AFAIK the very first,
functional prototype we know of was Postgres-R for PostgreSQL 6.4.2 (1).
So the very same holds true for a replication solution integrated into
the backend: we know such an implementation can function satisfactorily.

As we mostly agree, the performance bottelneck is _not_ the CPU, but the
nodes interconnects (the network). Regarding communication between the
backends and the replication solution, performance isn't that much of an
issue, because the inter-node communication will allways be slower than
inter-process communication.

A different problem is how to distribute PostgreSQL with different
upcomming replication solutions. It seems to me that most people's main
concern is not being able to get a prebuilt PostgreSQL with
_just_one_replication_solution_that_works_(tm) For most users it really
doesn't matter _how_ exactly the solution technically got integrated.

This problem gets solved with hooks and preloading a library: you could
simply provide _one_ PostgreSQL package which provides hooks for
replication solutions. Those could then provide a package with their
library. This of course is only doable if the number of hooks is kept
low.

Regards

Markus

[1]: pgreplication project on gborg: http://gborg.postgresql.org/project/pgreplication/projdisplay.php
http://gborg.postgresql.org/project/pgreplication/projdisplay.php