Обсуждение: Slony-I makes progress
After some substantial progress on the Slony-I engine development, I'd
like to give a little status report and invite everyone who wants to
participate in this project to join the mailing list and the development
team.
The project homepage is here:
http://gborg.postgresql.org/project/slony1/projdisplay.php
====================================================================
The current development status:
Slony-I configures on RedHat Linux and FreeBSD 4.9 against a PostgreSQL
7.4. The engine is capable of hot install, hot subscribe with catch up
and cascaded slaves. The config utilities are either not existent yet or
ugly, incomplete shell scripts. The replication log trigger is the
pototypes version in PL/Tcl.
What this odd collection is capable of doing is creating a master, slave
and cascaded slave replication system with a pgbench database as master.
The kick is, that the master database is originally just a plain, stock
PG 7.4.1 with zero changes, and the pgbench application is running non
stop through the whole process.
====================================================================
Next steps:
The log trigger function must be reimplemented in C. I will do this
during the next couple of days.
Implement the functionality to change the data provider of a slave. With
that a slave can be added as a cascaded slave, copy the data, catch up
and then switch over to replicate against the master, or an existing
slave can become a cascaded one to reduce the load on the master.
Implement the failover capability to replace a failed master with a slave.
Add backup and replication logging (sort of PITR based on replication
information).
Add sequence replication.
Add script execution support for schema changes.
Both, the provider change and the failover need a much more complex
configuration than the current shell scripts can setup. The problem is
that the configuration and administration tools are not designed yet. So
here is a huge field for others to step up and take a key role in this
project.
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #
On Wed, Mar 03, 2004 at 04:57:28PM -0500, Jan Wieck wrote: > After some substantial progress on the Slony-I engine development, I'd > like to give a little status report and invite everyone who wants to > participate in this project to join the mailing list and the development > team. Jan, thank you so much for your hard work on this project. > Both, the provider change and the failover need a much more complex > configuration than the current shell scripts can setup. The problem is > that the configuration and administration tools are not designed yet. So > here is a huge field for others to step up and take a key role in this > project. So what are you looking for here? When I last built slony, things mostly worked, but a few niggling details were broken. I was going to submit a few patches, but when I talked to you, it seemed like you weren't quite ready for patches. Is the tree stable enough that I could do some work on it and expect it to remain relatively consistent over a few hours or even a day or two? Also, to get this out of the way (as it presently plagues erserver), do you have a particular language in mind? I'd like to avoid the dogmatic jihad by not submitting a perl tool if the eventual goal is to be end-to-end C (or java or tcl or whatever). Alex -- alex@posixnap.net Alex J. Avriette, Unix Systems Gladiator "Something seems to have happened to the life support system , Dave."
Followup-To: Slony1-general ML Alex J. Avriette wrote: > On Wed, Mar 03, 2004 at 04:57:28PM -0500, Jan Wieck wrote: >> After some substantial progress on the Slony-I engine development, I'd >> like to give a little status report and invite everyone who wants to >> participate in this project to join the mailing list and the development >> team. > > Jan, thank you so much for your hard work on this project. > >> Both, the provider change and the failover need a much more complex >> configuration than the current shell scripts can setup. The problem is >> that the configuration and administration tools are not designed yet. So >> here is a huge field for others to step up and take a key role in this >> project. > > So what are you looking for here? When I last built slony, things > mostly worked, but a few niggling details were broken. I was going to > submit a few patches, but when I talked to you, it seemed like you > weren't quite ready for patches. Is the tree stable enough that I could > do some work on it and expect it to remain relatively consistent over a > few hours or even a day or two? What I am looking for is a super-comfortable GUI application that makes planning and configuring a master-cascaded-multislave replication system doable by everyone who can identify a clickable button. Honestly, I personally can live with a sh+sed+m4 tool collection. But I guess only few would agree to that. So it's basically up to you and everyone else around here what the outcome of this is. What is required to fit into the data-center is a batch utility that can be called in a script and that causes a currently failing cluster to change the configuration (change the origin of data sets, change providers, drop nodes ... that kind of stuff). The same utility would ideally be able to setup new nodes etc. so that it can be used as an interims solution until the GUI wizzard is ready for prime time. The current CVS replicates fine and the test_?_pgbench scripts in the src/ducttape directory do it all at once. I have changed a couple of things in the autoconf stuff. The whole thing is now expected to be compiled and installed by the postgres user with --prefix pointing to the postgres home directory (the same as the --prefix for the PG installation from sources was). The problem here is, that if we ever want to create a single C function from a GUI tool on a remote box, its shared library better be in the PostgreSQL lib directory so it can be ... AS '$libdir/objfile' no matter where that is and what extension shared objects on that architecture have. > > Also, to get this out of the way (as it presently plagues erserver), do > you have a particular language in mind? I'd like to avoid the dogmatic > jihad by not submitting a perl tool if the eventual goal is to be > end-to-end C (or java or tcl or whatever). For the production batch commandline utility I think it is C. Other than that ... I said the field is open. Jan -- #======================================================================# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #================================================== JanWieck@Yahoo.com #
Jan, > What I am looking for is a super-comfortable GUI application that makes > planning and configuring a master-cascaded-multislave replication system > doable by everyone who can identify a clickable button. I personally don't think that a GUI tool should be the province of the Slony project. Seriously. I think that Slony should focus on a command-line api and catalogs, and allow the existing GUI projects to build a slony-supporting interface. But I'll join the project and we can have this discussion there. -- Josh Berkus Aglio Database Solutions San Francisco
Josh Berkus wrote: > > I personally don't think that a GUI tool should be the province of the Slony > project. Seriously. I think that Slony should focus on a command-line api > and catalogs, and allow the existing GUI projects to build a slony-supporting > interface. Why a command line api? I believe it would make sense to be able to configure and control all nodes of the entire system from psql connected to any of the nodes. That would also facilitate the existing GUI projects in adding a Slony-manager. Jochem -- I don't get it immigrants don't work and steal our jobs - Loesje
On Fri, Mar 05, 2004 at 12:47:23AM +0100, Jochem van Dieten wrote: > >I personally don't think that a GUI tool should be the province of the > >Slony project. Seriously. I think that Slony should focus on a I very much agree with this, but this is Jan's baby, so I didn't say anything. I have personally never used a GUI with a postgres database (well, okay, I used one for a bit to troubleshoot a problem my boss was having with a pg node once), and I don't really plan to. I guess I was unaware this is a common usage pattern. > >command-line api and catalogs, and allow the existing GUI projects to > >build a slony-supporting interface. > > Why a command line api? I believe it would make sense to be able > to configure and control all nodes of the entire system from psql > connected to any of the nodes. That would also facilitate the > existing GUI projects in adding a Slony-manager. In theory, most of the stuff that Slony is doing is within the database, and as such, could be configurable via stored procedures. I see a few problems with this. First off, it is not possible to configure external applications (such as erserver has a daemon) from within the database except through the modification of tables within the database which are monitored by said application. Second, it increases the footprint of Slony on the database. I am fairly uneasy about adding more tables, functions, and triggers to my (already quite taxed) production database. To add further functions for configuration, as well as related tables and triggers, makes my job managing the database more difficult. Additionally, those commands are queries. For something as trivial as configuration data, I would much rather not be issuing queries against an already very busy database. I am much more comfortable with the principle of external configuration files and programs. Lastly, and I may be the black sheep here, I don't find sql to be particularly useful for doing things that require a complex grammar. In this instance, I don't want to have to do something like: production=# select slony_config_setval( 'log_dir', '/data/slony_logs'); to manage the configuration. Obviously, this could be worse than the above example. I don't understand the opposition to an external set of tools (even a gui if need be). It seems to me, that until the postmaster has some kind of native replication, all replication efforts will be based on external programs. As such, they should be configured externally, and be treated as any other daemon would be. Alex -- alex@posixnap.net Alex J. Avriette, Unix Systems Gladiator "v shpxvat ungr jvaqbjf naq v ubcr ovyy tngrf oheaf va uryy." - Ronald O. Thompson, "13"
Alex J. Avriette wrote:
> On Fri, Mar 05, 2004 at 12:47:23AM +0100, Jochem van Dieten wrote:
>
>> >I personally don't think that a GUI tool should be the province of the
>> >Slony project. Seriously. I think that Slony should focus on a
>
> I very much agree with this, but this is Jan's baby, so I didn't say
> anything. I have personally never used a GUI with a postgres database
> (well, okay, I used one for a bit to troubleshoot a problem my boss
> was having with a pg node once), and I don't really plan to. I guess
> I was unaware this is a common usage pattern.
I was explicitly asking for opinions and input. I don't want this to be
"my baby". In the end I am a developer, not a DBA. I know how to do it,
but don't have the ultimate wisdom about how to manage it.
>
>> >command-line api and catalogs, and allow the existing GUI projects to
>> >build a slony-supporting interface.
>>
>> Why a command line api? I believe it would make sense to be able
>> to configure and control all nodes of the entire system from psql
>> connected to any of the nodes. That would also facilitate the
>> existing GUI projects in adding a Slony-manager.
>
> In theory, most of the stuff that Slony is doing is within the
> database, and as such, could be configurable via stored procedures. I
> see a few problems with this.
>
> First off, it is not possible to configure external applications (such
> as erserver has a daemon) from within the database except through the
> modification of tables within the database which are monitored by said
> application.
Which is exactly the way the Slony node daemons communicate with each
other and the way most of the admin activity is actually communicated
into the system.
The communication channels are "event" tables. The node daemons use
listen and notify to send messages from on to another. Messages are only
exchanged over this when the replication cluster configuration is
changed or every 10 seconds to tell "new replication data has
accumulated, come and get it". So I think the listen/notify protocol
suits well for that.
Some of the functionality happening on an event is already put into
stored procedures, and the replication engine as well as the (to be)
admin tools just call those. But that doesn't mean that using psql will
do the job. There are certain operations that need to be initiated (the
corresponding SP called) on a particular node, not just on any available
one. Also, these stored procedures take arguments, most of which are
just the ID numbers of configuration objects. Not the ideal user interface.
>
> Second, it increases the footprint of Slony on the database. I am
> fairly uneasy about adding more tables, functions, and triggers to my
> (already quite taxed) production database. To add further functions for
> configuration, as well as related tables and triggers, makes my job
> managing the database more difficult. Additionally, those commands are
> queries. For something as trivial as configuration data, I would much
> rather not be issuing queries against an already very busy database. I
> am much more comfortable with the principle of external configuration
> files and programs.
All tables, sequences and stored procdures/functions related to the
Slony replication system reside is a separate namespace. I found out
lately that (without replicating sequences yet), the whole replication
system can be "cleanly" removed from a database with just a DROP SCHEMA
... CASCADE.
The problem I have with external configurations is that they collide
with the hot subscribe capability. If node-3 subscribes to a set from
node-1, getting the data cascaded over node-2, the event to enable that
subscription has to travel from 1 over 2 to 3. When that is received
there, 3 has to copy over the current status of the data from 2 and then
catch up by replicating all changes that have happened during this copy,
which for large data sets can take a while. So node-2 must be aware of
this happening and not throw away any replication log since node-3
started copying, unless it is confirmed received by 3. The knowledge
that 3 exists must also cause other forwarding nodes to keep the log.
Imagine that after 3 successfully copied the data, while he's catching
up node-2 dies. At that moment, 3 can be reconfigured to get the rest of
the log from 1, or anyone else who has it, so that the copy effort is
not lost ... which at the time a node is failing in the system would
just add to the pain of the DBA.
>
> Lastly, and I may be the black sheep here, I don't find sql to be
> particularly useful for doing things that require a complex grammar. In
> this instance, I don't want to have to do something like:
>
> production=# select slony_config_setval( 'log_dir', '/data/slony_logs');
It currently looks more like
select "_MyCluster".storePath(2, 3, 'dbname=mydb host=node2', 30); select "_MyCluster".storeListen(2, 2, 3);
> to manage the configuration. Obviously, this could be worse than the
> above example.
So it "IS" worse! It is not supposed that the DBA uses the systems
internal API for configuration management. That is the whole reason for
the admin/config tools.
>
> I don't understand the opposition to an external set of tools (even a
> gui if need be). It seems to me, that until the postmaster has some
> kind of native replication, all replication efforts will be based on
> external programs. As such, they should be configured externally, and
> be treated as any other daemon would be.
There must be some external tools. And to be integrated into any
automated failover system, it needs to be commandline. So that one is a
given.
That still does not give an easy way to tell which of the existing
tables should be replicated, into how many independant sets they can be
divided, what nodes subscribe to what sets, what nodes do store and
forward of log data, all that stuff.
I have started on a small lex+yacc+libpq tool that will get me over the
immediate requirements I have to work on provider change and failover. I
will add that to the CVS (first as a subdirectory of ducttape) in a few
days.
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #
Jan Wieck said: > > The communication channels are "event" tables. The node daemons > use listen and notify to send messages from on to another. > Messages are only exchanged over this when the replication cluster > configuration is changed or every 10 seconds to tell "new > replication data has accumulated, come and get it". So I think > the listen/notify protocol suits well for that. > > Some of the functionality happening on an event is already put > into stored procedures, and the replication engine as well as the > (to be) admin tools just call those. But that doesn't mean that > using psql will do the job. There are certain operations that > need to be > initiated (the corresponding SP called) on a particular node, not > just on any available one. Also, these stored procedures take > arguments, most of which are just the ID numbers of configuration > objects. Not the ideal user interface. So some of the regular tasks can be performed from any of the nodes and some need to be done from a specific node. But if connected to the right node, they can all be done through sql and the management tool doesn't need shell access on the nodes. Right? > There must be some external tools. And to be integrated into any > automated failover system, it needs to be commandline. So that one > is a given. Would a database function that is called from the commandline like sudo -u postgres psql -c 'select "_MyCluster".useMaster(2,3,4);' qualify for that? Jochem
Jochem van Dieten wrote:
> Jan Wieck said:
>>
>> The communication channels are "event" tables. The node daemons
>> use listen and notify to send messages from on to another.
>> Messages are only exchanged over this when the replication cluster
>> configuration is changed or every 10 seconds to tell "new
>> replication data has accumulated, come and get it". So I think
>> the listen/notify protocol suits well for that.
>>
>> Some of the functionality happening on an event is already put
>> into stored procedures, and the replication engine as well as the
>> (to be) admin tools just call those. But that doesn't mean that
>> using psql will do the job. There are certain operations that
>> need to be
>> initiated (the corresponding SP called) on a particular node, not
>> just on any available one. Also, these stored procedures take
>> arguments, most of which are just the ID numbers of configuration
>> objects. Not the ideal user interface.
>
> So some of the regular tasks can be performed from any of the nodes
> and some need to be done from a specific node. But if connected to the
> right node, they can all be done through sql and the management tool
> doesn't need shell access on the nodes. Right?
No action so far needs shell acceess to any specific system. Everything
works through regular DB connections. It means however that the admin
tools will have postmaster port access to "every" node, from where they
are run. I don't think that this is an unreasonable requirement for the
DBA workstation in a datacenter.
>
>
>> There must be some external tools. And to be integrated into any
>> automated failover system, it needs to be commandline. So that one
>> is a given.
>
> Would a database function that is called from the commandline like
> sudo -u postgres psql -c 'select "_MyCluster".useMaster(2,3,4);'
> qualify for that?
That has no advantage over
sudo -u postgres slony_admtool -f MyCluster_emeergency_plan_A
Especially because psql does not allow multiple simultaneous DB
connections and has no control language for procedures.
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #
Jan Wieck wrote: > Alex J. Avriette wrote: > >> On Fri, Mar 05, 2004 at 12:47:23AM +0100, Jochem van Dieten wrote: >> >>> >I personally don't think that a GUI tool should be the province of >>> the >Slony project. Seriously. I think that Slony should focus on a >> >> >> I very much agree with this, but this is Jan's baby, so I didn't say >> anything. I have personally never used a GUI with a postgres database >> (well, okay, I used one for a bit to troubleshoot a problem my boss >> was having with a pg node once), and I don't really plan to. I guess >> I was unaware this is a common usage pattern. > > > I was explicitly asking for opinions and input. I don't want this to > be "my baby". In the end I am a developer, not a DBA. I know how to do > it, but don't have the ultimate wisdom about how to manage it. > If somebody likes to contribute a gui tool, I'm sure we could help to implement this in pgAdmin3. Regards, Andreas
In the last exciting episode, jochemd@oli.tudelft.nl (Jochem van Dieten) wrote: > Josh Berkus wrote: >> I personally don't think that a GUI tool should be the province of >> the Slony project. Seriously. I think that Slony should focus on >> a command-line api and catalogs, and allow the existing GUI >> projects to build a slony-supporting interface. > > Why a command line api? I believe it would make sense to be able to > configure and control all nodes of the entire system from psql > connected to any of the nodes. That would also facilitate the > existing GUI projects in adding a Slony-manager. Interesting... That would mean that the 'server' part of the application would be 'monitoring' NOTIFY requests on each of the nodes, right? Hmm... Queue up some records in the slony1.node_requests table, to indicate what needs to be changed, then NOTIFY "slony1". The server then has to look at _all_ the nodes for slony1.node_requests entries. It would be _very_ easy to write command line apps to manage this; no need to add any extra RPC scheme (e.g. - Java RMI, CORBA, talking to sockets), and no need to open extra firewall ports in addition to the ports already needed in order for Slony to communicate with the various databases. Further bonus: the "GUI project" need only have a database connection to one of the databases to control things. No need for ANYTHING else. After fleshing it out a little, that's a pretty slick approach. -- If this was helpful, <http://svcs.affero.net/rm.php?r=cbbrowne> rate me http://cbbrowne.com/info/multiplexor.html 0 7 * * * echo "...Linux is just a fad" | mail billg@microsoft.com \ -s "And remember..."
Christopher Browne wrote: > > Further bonus: the "GUI project" need only have a database connection > to one of the databases to control things. No need for ANYTHING else. > > After fleshing it out a little, that's a pretty slick approach. You miss the point, sorry. This "make GUI easy to write" approach leads to one major problem. When a central server in the cluster dies and the communication path's need to be redirected and the utility needs to contact all the remaining servers because they're not doing the big group chat always, but now their regular communication path is disrupted ... your GUI (the only thing wannabe-DBA's know) becomes useless and the whole plan with failover and backup systems falls apart. Jan -- #======================================================================# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #================================================== JanWieck@Yahoo.com #