replication hooks

On Thu, May 29, 2008 at 11:05:09PM +0300, Marko Kreen wrote:

There is this tiny matter of replicating schema changes asynchronously,
but I suspect nobody actually cares.

I know that Slony's users call this their number one irritant, so I
have my doubts nobody cares. But maybe nobody cares enough.

- The task cannot even be clearly defined (on technical level - how
the events should be represented).

Really? I've been in discussions where different people had clear
(but, alas, different) ideas of how to represent them.

- Any schema changes need to be carefully prepared anyway. Whether
to apply them to one or more servers does not make much difference.

One problem that designers of replication systems have is that they're
already thinking in the Serious Database Application world. But I
have recently had the pleasure of being reminded how many users of
database systems neither know nor care to know any of the details of
the underlying system. They already know how to make schema changes:
log into database, and start typing "ALTER TABLE. . ." You or I
agreeing that more careful preparation than that is important will not
change their mind. This is part of the reason MySQL looks so good:
you can "just do" these things. If it doesn't work out later, well,
you don't know that when your ALTER TABLE "just works".

- Major plus of async replica is ability to actually have different
schema on slaves.

I agree.

- People _do_ care about exact schema on single place - failover servers.

Yeah, but not only there. One of the things I was hoping to have
nailed down in the "hooks" discussion was, in fact, the use cases.
Half the time, people have such a clear idea of what _they_ want from
their replication that they come to believe "replication" means that.

Another thing I like about the current proposal is that it is very
clear about what it is (and isn't) aiming for.

A

--
Andrew Sullivan
ajs@commandprompt.com
+1 503 667 4564 x104
http://www.commandprompt.com/

On 5/29/08, Andrew Sullivan <ajs@commandprompt.com> wrote:

On Thu, May 29, 2008 at 11:05:09PM +0300, Marko Kreen wrote:

There is this tiny matter of replicating schema changes asynchronously,
but I suspect nobody actually cares.

I know that Slony's users call this their number one irritant, so I
have my doubts nobody cares. But maybe nobody cares enough.

Oh, users of course like their lives to be as easy as possible
and all tools be "do as i wish"-complete.

I meant no developer is interested after looking at the task complexity
and resulting payoff.

- The task cannot even be clearly defined (on technical level - how
the events should be represented).

Really? I've been in discussions where different people had clear
(but, alas, different) ideas of how to represent them.

Yeah. The main problem is that unless you do WAL based replication,
you cannot achieve transparency. So you need to pick few use cases
and tailor you solution for them, which gets uninteresting very fast
- user _will_ stumble upon spacial cases, and if they expect everything
"just work" the resulting conversation wont be funny.

- Any schema changes need to be carefully prepared anyway. Whether
to apply them to one or more servers does not make much difference.

One problem that designers of replication systems have is that they're
already thinking in the Serious Database Application world. But I
have recently had the pleasure of being reminded how many users of
database systems neither know nor care to know any of the details of
the underlying system. They already know how to make schema changes:
log into database, and start typing "ALTER TABLE. . ." You or I
agreeing that more careful preparation than that is important will not
change their mind. This is part of the reason MySQL looks so good:
you can "just do" these things. If it doesn't work out later, well,
you don't know that when your ALTER TABLE "just works".

Simple - use WAL-based replication.

Although - not so simple, as currently we don't provide it. The existing
PITR hooks expect users to write their own replication, which is not
a user-friendly approach...

Hopefully this will be fixed in 8.4.

- People _do_ care about exact schema on single place - failover servers.

Yeah, but not only there. One of the things I was hoping to have
nailed down in the "hooks" discussion was, in fact, the use cases.
Half the time, people have such a clear idea of what _they_ want from
their replication that they come to believe "replication" means that.

The main problem with replica-hooks-discuss list was lack of focus.

There are various replication methods - single-master, multi-master,
asynchronous, synchronous, WAL-based, trigger-based, changeset-based.

Any combination wants different hooks, putting them all together
makes people not care.

Eg - setting the topic to schema change logging for async trigger-based
replication would be better, but even there are various usage scenarios
that may not be compatible, so it people don't see a chance of common
hooks they don't bother. Actually I suspect this task is solvable,
main problem is that it's pretty low on anyones priority list.

Another thing I like about the current proposal is that it is very
clear about what it is (and isn't) aiming for.

Yes. And we can skip the "common hooks" discussion. ;)

--
marko

-----BEGIN PGP SIGNED MESSAGE-----
Hash: RIPEMD160

Yeah. The main problem is that unless you do WAL based replication,
you cannot achieve transparency. So you need to pick few use cases
and tailor you solution for them, which gets uninteresting very fast
- user _will_ stumble upon spacial cases,

Isn't that what PostGIS is for?

g,d,&r

- --
Greg Sabino Mullane greg@turnstep.com
PGP Key: 0x14964AC8 200805291840
http://biglumber.com/x/web?pk=2529DF6AB8F79407E94445B4BC9B906714964AC8
-----BEGIN PGP SIGNATURE-----

iEYEAREDAAYFAkg/MTQACgkQvJuQZxSWSsjfEACgr64IdjtfhidTAGg/dVVVTMOP
0HAAn2tkYoNleSryZ5EyiSMp0o2x9ZFL
=Fmc4
-----END PGP SIGNATURE-----

Hi Marko,

Replication requirements vary widely of course, but DDL support is shared by such a wide range of use cases it is very difficult to see how any real solution would fail to include it. This extends to change extraction APIs, however, defined. The question of what DDL to replicate is also quite clear-all of it with as few exceptions as possible.

For instance, it is almost impossible to set up and manage replicated systems easily if you cannot propagate schema changes in serialized order along with other updates from applications. The inconvenience of using alternative mechanisms like the SLONY 'execute script' is considerable and breaks most commonly used database management tools.

That said, SLONY at least serializes the changes. Non-serialized approaches lead to serious outages and can get you into distributed consensus problems, such as when is it 'safe' to change schema across different instances. These are very hard to solve practically and tend to run into known impossibility results like Brewer's Conjecture, which holds that it is impossible to keep distributed databases consistent while also remaining open for updates and handling network partitions.

I'll post back later on the question of the API. The key is to do something simple that avoids the problems discussed by Andrew and ties it accurately to use cases. However, this requires a more prepared response than my hastily written post from last night.

Cheers, Robert

On 5/29/08 9:05 PM, "Marko Kreen" <markokr@gmail.com> wrote:

On 5/29/08, Andrew Sullivan <ajs@commandprompt.com> wrote:

On Thu, May 29, 2008 at 12:05:18PM -0700, Robert Hodges wrote:

people are starting to get religion on this issue I would strongly
advocate a parallel effort to put in a change-set extraction API
that would allow construction of comprehensive master/slave
replication.

You know, I gave a talk in Ottawa just last week about how the last
effort to develop a comprehensive API for replication failed. I had
some ideas about why, the main one of which is something like this:
"Big features with a roadmap have not historically worked, so unless
we're willing to change the way we work, we won't get that."

I don't think an API is what's needed. It's clear proposals for
particlar features that can be delivered in small pieces. That's what
the current proposal offers. I think any kind of row-based approach
such as what you're proposing would need that kind of proposal too.

That isn't to say that I think an API is impossible or undesirable.
It is to say that the last few times we tried, it went nowhere; and
that I don't think the circumstances have changed.

I think the issue is simpler - API for synchronous replication is
undesirable - it would be too complex and hinder future development
(as I explained above).

And the API for asynchronous replication is already there - triggers,
txid functions for queueing.

There is this tiny matter of replicating schema changes asynchronously,
but I suspect nobody actually cares. Few random points about that:

- The task cannot even be clearly defined (on technical level - how
the events should be represented).
- Any schema changes need to be carefully prepared anyway. Whether
to apply them to one or more servers does not make much difference.
- Major plus of async replica is ability to actually have different
schema on slaves.
- People _do_ care about exact schema on single place - failover servers.
- But for failover server we want also synchronous replication.

So if we have synchronous WAL based replication for failover servers,
the interest in hooks to log schema changes will decrease even more.

--
marko

--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

--
Robert Hodges, CTO, Continuent, Inc.
Email: robert.hodges@continuent.com
Mobile: +1-510-501-3728 Skype: hodgesrm

Marko Kreen wrote:

There is this tiny matter of replicating schema changes asynchronously,
but I suspect nobody actually cares. Few random points about that:

I'm not sure I follow you - the Sybase 'warm standby' replication of
everything is really
useful for business continuity. The per-table rep is more effective for
publishing reference
data, but is painful to maintain.

Not having something that automagically reps a complete copy including
DDL (except
for temp tables) is a major weakness IMO.

James