Обсуждение: Postgres-R: tuple serialization
Hi,
yesterday, I promised to outline the requirements of Postgres-R for
tuple serialization, which we have been talking about before. There are
basically three types of how to serialize tuple changes, depending on
whether they originate from an INSERT, UPDATE or DELETE. For updates and
deletes, it saves the old pkey as well as the origin (a global
transaction id) of the tuple (required for consistent serialization on
remote nodes). For inserts and updates, all added or changed attributes
need to be serialized as well.
pkey+origin changes INSERT - x UPDATE x x DELETE x
-
Note, that the pkey attributes may never be null, so an isnull bit field
can be skipped for those attributes. For the insert case, all attributes
(including primary key attributes) are serialized. Updates require an
additional bit field (well, I'm using chars ATM) to store which
attributes have changed. Only those should be transferred.
I'm tempted to unify that, so that inserts are serialized as the
difference against the default vaules or NULL. That would make things
easier for Postgres-R. However, how about other uses of such a fast
tuple applicator? Does such a use case exist at all? I mean, for
parallelizing COPY FROM STDIN, one certainly doesn't want to serialize
all input tuples into that format before feeding multiple helper
backends. Instead, I'd recommend letting the helper backends do the
parsing and therefore parallelize that as well.
For other features, like parallel pg_dump or even parallel query
execution, this tuple serialization code doesn't help much, IMO. So I'm
thinking that optimizing it for Postgres-R's internal use is the best
way to go.
Comments? Opinions?
Regards
Markus
On Jul 22, 2008, at 3:04 AM, Markus Wanner wrote: > yesterday, I promised to outline the requirements of Postgres-R for > tuple serialization, which we have been talking about before. There > are basically three types of how to serialize tuple changes, > depending on whether they originate from an INSERT, UPDATE or > DELETE. For updates and deletes, it saves the old pkey as well as > the origin (a global transaction id) of the tuple (required for > consistent serialization on remote nodes). For inserts and updates, > all added or changed attributes need to be serialized as well. > > pkey+origin changes > INSERT - x > UPDATE x x > DELETE x - > > Note, that the pkey attributes may never be null, so an isnull bit > field can be skipped for those attributes. For the insert case, all > attributes (including primary key attributes) are serialized. > Updates require an additional bit field (well, I'm using chars ATM) > to store which attributes have changed. Only those should be > transferred. > > I'm tempted to unify that, so that inserts are serialized as the > difference against the default vaules or NULL. That would make > things easier for Postgres-R. However, how about other uses of such > a fast tuple applicator? Does such a use case exist at all? I mean, > for parallelizing COPY FROM STDIN, one certainly doesn't want to > serialize all input tuples into that format before feeding multiple > helper backends. Instead, I'd recommend letting the helper backends > do the parsing and therefore parallelize that as well. > > For other features, like parallel pg_dump or even parallel query > execution, this tuple serialization code doesn't help much, IMO. So > I'm thinking that optimizing it for Postgres-R's internal use is > the best way to go. > > Comments? Opinions? ISTM that both londiste and Slony would be able to make use of these improvements as well. A modular replication system should be able to use a variety of methods for logging data changes and then applying them on a subscriber, so long as some kind of common transport can be agreed upon (such as text). So having a change capture and apply mechanism that isn't dependent on a lot of extra stuff would be generally useful to any replication mechanism. -- Decibel!, aka Jim C. Nasby, Database Architect decibel@decibel.org Give your computer some brain candy! www.distributed.net Team #1828
Hi, Decibel! wrote: > ISTM that both londiste and Slony would be able to make use of these > improvements as well. A modular replication system should be able to use > a variety of methods for logging data changes and then applying them on > a subscriber, so long as some kind of common transport can be agreed > upon (such as text). So having a change capture and apply mechanism that > isn't dependent on a lot of extra stuff would be generally useful to any > replication mechanism. Hm.. yeah, that's a good hint. However, I'm not sure how londiste and Slony would interface with these internal methods. That would require some sort of special replication triggers or something. But when to fire them? After every statement (sync)? Just before commit (eager)? After commit (lazy)? (These are the points in Postgres-R, where the internal methods are called). I'm claiming that Postgres-R is modular (enough). But I'm unsure what interface it could provide to the outer world. Regards Markus Wanner
On Jul 22, 2008, at 4:43 PM, Markus Wanner wrote: > Decibel! wrote: >> ISTM that both londiste and Slony would be able to make use of >> these improvements as well. A modular replication system should be >> able to use a variety of methods for logging data changes and then >> applying them on a subscriber, so long as some kind of common >> transport can be agreed upon (such as text). So having a change >> capture and apply mechanism that isn't dependent on a lot of extra >> stuff would be generally useful to any replication mechanism. > > Hm.. yeah, that's a good hint. However, I'm not sure how londiste > and Slony would interface with these internal methods. That would > require some sort of special replication triggers or something. But > when to fire them? After every statement (sync)? Just before commit > (eager)? After commit (lazy)? (These are the points in Postgres-R, > where the internal methods are called). Currently, londiste triggers are per-row, not deferred. IIRC, londiste is the same. ISTM it'd be much better if we had per- statement triggers that could see what data had changed; that'd likely be a lot more efficient than doing stuff per-row. In any case, both replication systems should work with either sync or eager. I can't see them working with lazy. What about just making all three available? > I'm claiming that Postgres-R is modular (enough). But I'm unsure > what interface it could provide to the outer world. Yeah. I suspect that Postgres-R could end up taking the place of the replica-hooks mailing list (and more, of course). -- Decibel!, aka Jim C. Nasby, Database Architect decibel@decibel.org Give your computer some brain candy! www.distributed.net Team #1828
Hi, Decibel! wrote: > Currently, londiste triggers are per-row, not deferred. IIRC, londiste > is the same. ISTM it'd be much better if we had per-statement triggers > that could see what data had changed; that'd likely be a lot more > efficient than doing stuff per-row. Well, now that I think about it... there might be *lots* of changes. Certainly something you don't want to collect in memory. At the moment, Postgres-R cannot handle this, but I plan to add an upper limit on the change set size, and just send it out as soon as it exceeds that limit, then continue collecting. (Note for the GCS adept: this partial change set may be sent via reliable multicast, only the very last change set before the commit needs to be totally ordered.) That would mean, introducing another 'change set full' hook... > In any case, both replication systems should work with either sync or > eager. I can't see them working with lazy. Huh? AFAIK, londiste as well as Slony-I are both async. So what would hooks for sync replication be good for? Why not rather only lazy hooks? (Well, lazy hooks will pose yet another problem: those theoretically need to run somewhen *after* the commit, but at that time we don't have an open transaction, so where exactly shall we do this?) > What about just making all three available? Doh. Ehm. That's a lot of work for something we are not even sure it's good for anything. I'm certainly willing to help. And if other projects show enough interest, I might even add the appropriate triggers myself. But as long as this is all unclear, I certainly have more important things on my todo list for Postgres-R (see that TODO list ;-) ). > Yeah. I suspect that Postgres-R could end up taking the place of the > replica-hooks mailing list (and more, of course). Let's hope so, yeah! Regards Markus Wanner