Обсуждение: matview niceties: pick any two of these three
Tom has refactored where and how certain parts of the work get done for materialized views, reducing issues with modularity violations. I have been and will continue to review his changes to better understand how the pieces of the code fit together, so hopefully he won't need to do so much with future contributions. He has not, so far, changed functionality or the regression tests -- although he has complained about things, of which the below is the all of what remains as an issue for him as far as I know. I may have missed something; if so, it would be great to be reminded of what. What remains is a choice between three alternatives on which no consensus has been reached. There are three things that, as fas as I can tell, *everyone* agrees would be nice to be true about the matview implementation for 9.3, but it does not seem feasible to have more than two: (1) The ability to count on the results from a query which references a matview to reflect valid data from *some* point in time. (2) The ability to create unlogged materialized views. (3) The ability to consider a zero-length matview heap and a matview heap with allocated-but-empty pages as logically identical. Tom wants to ditch (2) to allow the others. Robert wants to ditch (1) to allow the others. I want to ditch (3) to allow the others. Andres wants (3) and has not expressed an opinion on which he would prefer to give up to get it. I believe Josh Berkus has mentioned how useful he thinks both (1) and (2) would be, without really commenting on (3). I am convinced that we can solve (3) in a later release without any significant impact on those already using matviews, including unlogged ones. Tom and Robert have suggested that the current implementation would paint us into a corner or pose a pg_upgrade hazard, without any clear indication of the mechanism of the problem. The logjam has caused me to hold back on finishing up the direction I prefer, for fear of conflicts with other work that someone else may be trying to do. Given that time is short, I'm going to apply the patch which inserts five lines (two of them comment lines) on the basis that at least the way that is currently implemented will have no known bugs, and it's just not that much to rip back out if we reach consensus on another direction. How to we resolve this impasse? -- Kevin Grittner EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
> Tom wants to ditch (2) to allow the others. Robert wants to ditch > (1) to allow the others. I want to ditch (3) to allow the others. > Andres wants (3) and has not expressed an opinion on which he would > prefer to give up to get it. I believe Josh Berkus has mentioned > how useful he thinks both (1) and (2) would be, without really > commenting on (3). As I understand it, we don't currently have any mechanism in Postgres which would cause allocated-but-empty pages. That we *might* have such a thing in 9.4 doesn't seem like a sufficient obstacle; we also might not. Further, I don't think that pg_upgrade is really a red card here. Matviews will be a new feature for 9.3. If we end up having to say "if you use pg_upgrade to upgrade to 9.4, you will need to rebuild your matviews afterwards", then that's what happens. People are used to some wonkiness in new features, and at this point the majority of our users don't use pg_upgrade. So, yes, I'd vote for (1) and (2) over (3), if that's the options which make sense. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
Josh Berkus <josh@agliodbs.com> writes: > As I understand it, we don't currently have any mechanism in Postgres > which would cause allocated-but-empty pages. That's not correct: the situation can easily arise after a database crash. (The scenario is that we've done smgrextend to add the first page to the file, but not yet completed or WAL-logged insertion of any data into it. This leaves us with an empty, all-zero page that will be ignored until we next want to add some data to the table.) The core problem here is that file extension is not a transactional operation, because it doesn't roll back on crash. The current matview design gets around this problem by requiring that transition between scannable and unscannable states involve a complete table rewrite, and thus the transactionality issue can be hidden behind a transactional update of the matview's pg_class.relfilenode field. IMO, that is obviously a dead-end design, because we are going to want scannability status updates associated with partial updates of the matview's contents. So Kevin's summary is leaving out one key desirable property: (4) ability to change scannability state without a full table rewrite. Putting the state into pg_class would preserve that property. regards, tom lane
Tom Lane <tgl@sss.pgh.pa.us> wrote: > The current matview design gets around this problem by requiring > that transition between scannable and unscannable states involve > a complete table rewrite, and thus the transactionality issue can > be hidden behind a transactional update of the matview's > pg_class.relfilenode field. IMO, that is obviously a dead-end > design, because we are going to want scannability status updates > associated with partial updates of the matview's contents. I don't think the change between populated and non-populated needs to change without a new heap being swapped in. To do incremental maintenance, one must be starting with a populated heap. What will change without a new heap is the contents of the matview and whether a populated matview is "fresh" enough to be scanned for some specific usage. There's a lot of bikeshedding left to be done on what options we need there and how they will be stored; but I don't see how that threatens the ability to use a zero-length heap to indicate that a matview has not been populated. Scannability and the populated state only happen to be the same thing for the moment, but improving the scannability test does not require a new mechanism for determining whether the matview has been populated. A new "populated" mechanism is, we all agree, desirable; but it's not necessary to improve scannability testing. We'll get into all kinds of design trouble if we conflate these two separate concepts. > Kevin's summary is leaving out one key desirable property: > > (4) ability to change scannability state without a full table > rewrite. > > Putting the state into pg_class would preserve that property. I could probably add a hundred other desirable properties for a materialized view implementation; I was trying to list the ones which were still up-for-grabs for the initial 9.3 implementation. If we want to make progress on catching up to other major databases on this, it will take progress in many releases; and every one of those will lack something that someone wants. OK, a lot of things that a lot of people want. But if we can't commit something that is less than complete or less than perfect to make incremental change, we're not going to be getting anywhere very fast. I'm not sure why you would want to make a non-populated matview scannable, anyway; which seems like the only reason to have (4) without more refined maintenance techniques. I plan to submit a patch to the first 9.4 CF for differential updates (generating the new contents in a temporary heap and transactionally applying the changes to the existing matview heap) as a REFRESH option. I'm working on estimates for the amount of work needed to support delta-driven incremental changes to matview contents based on the declaration of the matview for various types of matviews. On a preliminary basis, I think that we can have fast incremental updates in 9.4 without any imperative coding for all non-recursive views, although including support in 9.4 for incremental updates for views with window functions is still iffy, and some desirable optimizations for aggregates might not make it. I'm really very sure that incremental updates for recursive matviews will not be a 9.4 feature, since that is best handled by the "Delete and Rederive" (DRed) algorithm, while incremental maintenance for non-recursive matviews is best handled with the "counting" algorithm, and it seems unlikely that we'll get both into one major release. If I thought that the current hack for tracking the populated state would be permanent, or that it would be so hard to replace in a later release that it would impede further development, I would be the first one to rip it out. I haven't been able to see such a risk, and nobody has adequately explained where they think the risk lies. The "crash during extension of the first page" argument doesn't hold up unless you conflate the populated state with the scannable state, and it seems likely that the scannable will not only be separate, but able to differ for among backend processes. If adding unlogged materialized views requires reworking support for unlogged relations in general in a way that puts the data into the catalog, and that becomes my issue to solve, I can deal with that -- but that time would come out of time I would otherwise be spending working on supporting incremental update of certain classes of materialized views, and progress there will be slower than I've been hoping. It's all about trade-offs. If the unlogged recovery changes wind up on my plate, for example, it might mean that the next release won't be able to have incremental updates for matviews with a NOT EXISTS clause, since support for that looks to be roughly the same amount of work. I will certainly go along with the consensus, but from my perspective, living with the current hack for determining whether a matview has been populated is worth the benefit of having unlogged matviews in 9.3. I don't expect that hack to be viable for more that one or two releases; I just think that it's a practical matter of working with the tools we have *now* for unlogged relations to try to create a new type of unlogged relation. I'm all for better tools and will be happy to use them once they exist. -- Kevin Grittner EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Kevin Grittner <kgrittn@ymail.com> writes: > Tom Lane <tgl@sss.pgh.pa.us> wrote: >> The current matview design gets around this problem by requiring >> that transition between scannable and unscannable states involve >> a complete table rewrite, and thus the transactionality issue can >> be hidden behind a transactional update of the matview's >> pg_class.relfilenode field.� IMO, that is obviously a dead-end >> design, because we are going to want scannability status updates >> associated with partial updates of the matview's contents. > I don't think the change between populated and non-populated needs > to change without a new heap being swapped in. I think you are using the distinction between "populated" and "scannable" to dodge the issue. My point is that whenever we add some better control over whether matviews are considered too-stale-to-use, it is going to be important that the staleness state can be changed, in either direction, without a full rewrite of the matview's contents. Whether you call that "populated", or "scannable", or "stale", or "warm fuzzy" doesn't matter to me. If we can't do that then the feature is not actually going to be of any great use. And I want such updates to be transactional, because they'll very likely be connected to transactional updates of the matview's contents. So I continue to maintain that the current design is a dead end. > We'll get into all > kinds of design trouble if we conflate these two separate concepts. Perhaps I still don't understand what useful distinction is there. If there is a critical distinction, why is the current patch exposing only one state value to users? > I could probably add a hundred other desirable properties for a > materialized view implementation; I was trying to list the ones > which were still up-for-grabs for the initial 9.3 implementation. > If we want to make progress on catching up to other major databases > on this, it will take progress in many releases; and every one of > those will lack something that someone wants.� OK, a lot of things > that a lot of people want.� But if we can't commit something that > is less than complete or less than perfect to make incremental > change, we're not going to be getting anywhere very fast. Yes, exactly. I am not sure why you are so desperate to have unlogged matviews in 9.3 that you are willing to risk finding yourself boxed in and unable to make future improvements that everybody agrees are useful. ISTM that plain matviews are a perfectly good first-cut feature and we can add unlogged ones whenever we can think of a better way to do it. > If I thought that the current hack for tracking the populated state > would be permanent, or that it would be so hard to replace in a > later release that it would impede further development, I would be > the first one to rip it out. I haven't been able to see such a > risk, and nobody has adequately explained where they think the risk > lies. You can't update that state transactionally, and you can't update it without a complete rewrite. The only solutions you've offered to that are devices Rube Goldberg would be proud of (not that the code wouldn't fill his heart with joy as it is). This is not a good place to be for a first-cut implementation, at least not if we don't want to find ourselves needing to make a representation change that pg_upgrade can't readily cope with. The long and the short of it is this: having unlogged matviews in 9.3 is not worth taking that risk for. IMO anyway. > I will certainly go along with the consensus, but from my > perspective, living with the current hack for determining whether a > matview has been populated is worth the benefit of having unlogged > matviews in 9.3.� I don't expect that hack to be viable for more > that one or two releases; I just think that it's a practical matter > of working with the tools we have *now* for unlogged relations to > try to create a new type of unlogged relation. If you yourself admit that this hack has got a very short life expectancy, why are you so willing to buy into the assumption that you can get rid of it readily? I do not grant that assumption, and I'm not sure how come you think that hacking pg_upgrade to the point of being able to get rid of it will be zero-cost, or even acceptable to the community at all. regards, tom lane
On 5/2/13 6:00 PM, Kevin Grittner wrote: > (1) The ability to count on the results from a query which > references a matview to reflect valid data from *some* point in > time. > > (2) The ability to create unlogged materialized views. > > (3) The ability to consider a zero-length matview heap and a > matview heap with allocated-but-empty pages as logically identical. I have thought from the very beginning that violating (3) was totally weird. I would ditch (2) in favor of (3).
Tom Lane <tgl@sss.pgh.pa.us> wrote: > Kevin Grittner <kgrittn@ymail.com> writes: >> Tom Lane <tgl@sss.pgh.pa.us> wrote: >>> The current matview design gets around this problem by >>> requiring that transition between scannable and unscannable >>> states involve a complete table rewrite, and thus the >>> transactionality issue can be hidden behind a transactional >>> update of the matview's pg_class.relfilenode field. IMO, that >>> is obviously a dead-end design, because we are going to want >>> scannability status updates associated with partial updates of >>> the matview's contents. > >> I don't think the change between populated and non-populated >> needs to change without a new heap being swapped in. > > I think you are using the distinction between "populated" and > "scannable" to dodge the issue. My point is that whenever we add > some better control over whether matviews are considered > too-stale-to-use, it is going to be important that the staleness > state can be changed, in either direction, without a full rewrite > of the matview's contents. Whether you call that "populated", or > "scannable", or "stale", or "warm fuzzy" doesn't matter to me. > If we can't do that then the feature is not actually going to be > of any great use. I suspect that we will never support scanning of a non-populated matview, but that other considerations for what is scannable will be added. That's not a word game, it's an important semantic difference. > And I want such updates to be transactional, because they'll very > likely be connected to transactional updates of the matview's > contents. Yep, that is definitely my intent. > So I continue to maintain that the current design is a dead end. Not if you don't muddle multiple concepts as one thing. >> We'll get into all kinds of design trouble if we conflate these >> two separate concepts. > > Perhaps I still don't understand what useful distinction is > there. If there is a critical distinction, why is the current > patch exposing only one state value to users? We don't have any other criteria for what is scannable implemented yet in this release, but perhaps you're right that we should have a pg_relation_is_populated() function available to user in addition to the pg_relation_is_scannable() function. For the moment, the latter would just return the result of calling the former. I hadn't thought the former was useful at the SQL level at this point, but thinking about it now, I wonder whether "populated" wouldn't be the more appropriate test for pg_dump to use. I'm not sure just yet, but it bears some thinking. More on that tomorrow after I sleep on it. >> I could probably add a hundred other desirable properties for a >> materialized view implementation; I was trying to list the ones >> which were still up-for-grabs for the initial 9.3 >> implementation. If we want to make progress on catching up to >> other major databases on this, it will take progress in many >> releases; and every one of those will lack something that >> someone wants. OK, a lot of things that a lot of people want. >> But if we can't commit something that is less than complete or >> less than perfect to make incremental change, we're not going to >> be getting anywhere very fast. > > Yes, exactly. I am not sure why you are so desperate to have > unlogged matviews in 9.3 that you are willing to risk finding > yourself boxed in and unable to make future improvements that > everybody agrees are useful. ISTM that plain matviews are a > perfectly good first-cut feature and we can add unlogged ones > whenever we can think of a better way to do it. I think matviews are a useful feature without an unlogged option; but I think they will be significantly more useful with the option. I agree it's a judgment call whether the hacks needed to support it for now are worth the benefit of the option. Based on my reading of the user needs and the code, I think it is worth the trade-off, but I can understand arguments to the contrary. The lack of specifics on where problems would be has left me unconvinced. The most specific arguments so far have been with the "extend from zero page length during incremental maintenance" (which I hope I have made clear why I don't see a problem there, but maybe I need to say more...), and the argument that we might want to stuff metadata at the front of the heap to track this, which is out of the question for 9.3 and would require the same response later whether we use the zero-length hack for now or not. >> If I thought that the current hack for tracking the populated >> state would be permanent, or that it would be so hard to replace >> in a later release that it would impede further development, I >> would be the first one to rip it out. I haven't been able to >> see such a risk, and nobody has adequately explained where they >> think the risk lies. > > You can't update that state transactionally, and you can't update > it without a complete rewrite. Right, the *populated* state, indicating whether the query has been run to initially fill the matview with data, blocks other access. > The only solutions you've offered to that are devices Rube > Goldberg would be proud of (not that the code wouldn't fill his > heart with joy as it is). This is not a good place to be for > a first-cut implementation, at least not if we don't want to find > ourselves needing to make a representation change that pg_upgrade > can't readily cope with. I've already explained why I don't think pg_upgrade is an issue, short of the metadata in the first heap page change which was suggested, but I'll give it another go. pg_dump uses CREATE MATERIALIZED VIEW ... WITH NO DATA in the same phase as views are created, because each can reference the other. If we come up with a new mechanism for indicating non-populated matviews in a later release, the server from such a release which is digesting those commands will use its new mechanism. pg_upgrade need not know or care what the server is doing, or even that these commands are in the schema dump. So no problem there. For the pg_dump upgrade path, the REFRESH commands which may follow for some or all matviews will likewise do the right thing for the version of the server processing them. Still no problem. At the point where we move information about whether a matview is populate to the catalogs, we will probably, like at least one other database product, add an ALTER MATERIALIZED VIEW option to flag a matview as scannable, and the pg_dump of that new release could be modified to emit those where appropriate in the schema dump phase when the option used by pg_upgrade is set. That's some coding, but hardly a big deal. So where is the Rube Goldberg solution in that? > The long and the short of it is this: having unlogged matviews in > 9.3 is not worth taking that risk for. IMO anyway. We differ on that, but maybe you see a risk I'm missing, so I'm trying to understand what you see. >> I will certainly go along with the consensus, but from my >> perspective, living with the current hack for determining >> whether a matview has been populated is worth the benefit of >> having unlogged matviews in 9.3. I don't expect that hack to be >> viable for more that one or two releases; I just think that it's >> a practical matter of working with the tools we have *now* for >> unlogged relations to try to create a new type of unlogged >> relation. > > If you yourself admit that this hack has got a very short life > expectancy, why are you so willing to buy into the assumption > that you can get rid of it readily? I do not grant that > assumption, and I'm not sure how come you think that hacking > pg_upgrade to the point of being able to get rid of it will be > zero-cost, or even acceptable to the community at all. I have not heard a single suggestion of why pg_upgrade would need a single line changed. I'm believe that pg_dump will need some ALTER statements added when we add the catalog-based flagging of this, but that's true whether or not we use this technique in the meantime. The reason I think it can be readily abandoned is that the heap remains a valid heap according to the historical rules, and the rules which by all appearances would follow. So it is a question of whether the populated state can be transferred from a version using this technique to a version which is not with little or no pain (which I've explained above seems quite feasible to me) or whether we have to go so far as to tell people to REFRESH their unlogged materialized views in the release notes for the new release. Unlikely as that seems ot me, I think people would accept that. What do you see that I'm missing? -- Kevin Grittner EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Kevin Grittner <kgrittn@ymail.com> writes: > What do you see that I'm missing? TBH, if I had 20-20 foresight, we'd not be having this discussion: either I could see that you're right and this patch isn't going to cause us enormous pain, or I could put my finger on exactly where and why it's going to hurt us. But I can't do the latter today. Nonetheless, this patch terrifies me. It's ugly, it's a serious layering violation, and it flies in the face of very-long-standing assumptions about the semantics of heap storage. My gut tells me that we *will* regret shipping things this way. Nor am I impressed with the amount of functionality we're gaining by taking such a risk. regards, tom lane
... btw, I noticed a minor misfeature in the current implementation: regression=# select pg_relation_size('int8_tbl');pg_relation_size ------------------ 8192 (1 row) regression=# create materialized view mv1 as select * from int8_tbl; SELECT 5 regression=# select pg_relation_size('mv1');pg_relation_size ------------------ 16384 (1 row) So when populating a matview, we fail to make any use at all of the initially-added page. On the other hand, regression=# vacuum full mv1; VACUUM regression=# select pg_relation_size('mv1');pg_relation_size ------------------ 8192 (1 row) regression=# refresh materialized view mv1; REFRESH MATERIALIZED VIEW regression=# select pg_relation_size('mv1');pg_relation_size ------------------ 16384 (1 row) I haven't looked into why the VACUUM FULL code path is able to make use of the initially-created page while the CREATE/REFRESH code path can't. Possibly it's due to some innocent-looking difference in order of operations. The details aren't really too relevant though. Rather, my point is that IMO this sort of bug is an inevitable consequence of the layering violation that's at the heart of the current matview design. If we stick with this design, I'm afraid we'll be squashing bugs of this kind till kingdom come, and some of them may be much more painful to fix than the ones we've found to date. Layering violations tend to beget more layering violations. regards, tom lane
Tom Lane <tgl@sss.pgh.pa.us> wrote: > Kevin Grittner <kgrittn@ymail.com> writes: >> What do you see that I'm missing? > > TBH, if I had 20-20 foresight, we'd not be having this > discussion: either I could see that you're right and this patch > isn't going to cause us enormous pain, or I could put my finger > on exactly where and why it's going to hurt us. But I can't do > the latter today. Nonetheless, this patch terrifies me. It's > ugly, it's a serious layering violation, and it flies in the face > of very-long-standing assumptions about the semantics of heap > storage. My gut tells me that we *will* regret shipping things > this way. Nor am I impressed with the amount of functionality > we're gaining by taking such a risk. OK, I think there are more votes for removing unlogged matviews for 9.3 than for any other option, and it's time to make a call; so I'm conceding the point. Do you want me to take a shot at undoing that and straightening things out, or given the short time and your superior grasp of the layer boundaries, would you prefer to take it? -- Kevin Grittner EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On 05/05/13 10:41, Kevin Grittner wrote: [...] > removing unlogged matviews for 9.3 [...] So you are going to unlog the unlogged matviews from 9.3? :-) More seriously, I think this will result in a better solution for 9.4 - as it seems to me to be too rushed to fix it now Cheers, Gavin.
Tom Lane <tgl@sss.pgh.pa.us> writes: > The long and the short of it is this: having unlogged matviews in 9.3 > is not worth taking that risk for. IMO anyway. FWIW, +1 Regards, -- Dimitri Fontaine http://2ndQuadrant.fr PostgreSQL : Expertise, Formation et Support