Обсуждение: Re: [GENERAL] Buglist
On Friday 22 August 2003 16:23, Manfred Koizar wrote: > On Fri, 22 Aug 2003 12:15:33 +0530, "Shridhar Daithankar" > > <shridhar_daithankar@persistent.co.in> wrote: > >> Which leads us to a zero gravity vacuum, that does the lazy vacuum for > >> pages currently available in the buffer cache only. [...] > > > >Since autovacuum issues vacuum analyze only, is it acceptable to say that > > this is taken care of already? > > Even a plain VACUUM (without FULL) scans the whole relation to find > the (possibly few) pages that need to be changed. We are trying to > find a way to avoid those needless reads of clean pages, because (a) > they are IOs competing with other disk operations and (b) they push > useful pages out of OS cache and (c) of PG shared buffers. The latter > might become a non-issue with LRU-k, 2Q or ARC. But (a) and (b) > remain. Umm.. What does FSM does then? I was under impression that FSM stores page pointers and vacuum work on FSM information only. In that case, it wouldn't have to waste time to find out which pages to clean. Shridhar
On Fri, 22 Aug 2003 16:27:53 +0530, Shridhar Daithankar
<shridhar_daithankar@persistent.co.in> wrote:
>What does FSM does then?
FSM = Free Space Map. VACUUM writes information into the FSM, INSERTs
consult the FSM to find pages with free space for new tuples.
> I was under impression that FSM stores page
>pointers and vacuum work on FSM information only. In that case, it wouldn't
>have to waste time to find out which pages to clean.
This has been discussed yesterday here and on -hackers ("Decent VACUUM
(was: Buglist)"). We were talking about inventing a second data
structure: RSM.
Servus
Manfred
Shridhar Daithankar wrote: > On Friday 22 August 2003 16:23, Manfred Koizar wrote: >> On Fri, 22 Aug 2003 12:15:33 +0530, "Shridhar Daithankar" >> >> <shridhar_daithankar@persistent.co.in> wrote: >> >> Which leads us to a zero gravity vacuum, that does the lazy vacuum for >> >> pages currently available in the buffer cache only. [...] >> > >> >Since autovacuum issues vacuum analyze only, is it acceptable to say that >> > this is taken care of already? >> >> Even a plain VACUUM (without FULL) scans the whole relation to find >> the (possibly few) pages that need to be changed. We are trying to >> find a way to avoid those needless reads of clean pages, because (a) >> they are IOs competing with other disk operations and (b) they push >> useful pages out of OS cache and (c) of PG shared buffers. The latter >> might become a non-issue with LRU-k, 2Q or ARC. But (a) and (b) >> remain. > > Umm.. What does FSM does then? I was under impression that FSM stores page > pointers and vacuum work on FSM information only. In that case, it wouldn't > have to waste time to find out which pages to clean. It's the other way around! VACUUM scan's the tables to find and reclaim free space and remembers that free space in the FSM. The regular backends that need storage space to insert tuples then use the free space in the pages that are recorded in the FSM instead of adding new pages at the end of the relations. Jan > > Shridhar > > > ---------------------------(end of broadcast)--------------------------- > TIP 7: don't forget to increase your free space map settings -- #======================================================================# # 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 <JanWieck@Yahoo.com> writes:
> Shridhar Daithankar wrote:
>> Umm.. What does FSM does then? I was under impression that FSM stores page
>> pointers and vacuum work on FSM information only. In that case, it wouldn't
>> have to waste time to find out which pages to clean.
> It's the other way around! VACUUM scan's the tables to find and reclaim
> free space and remembers that free space in the FSM.
Right. One big question mark in my mind about these "partial vacuum"
proposals is whether they'd still allow adequate FSM information to be
maintained. If VACUUM isn't looking at most of the pages, there's no
very good way to acquire info about where there's free space.
regards, tom lane
On 22 Aug 2003 at 10:45, Tom Lane wrote: > Jan Wieck <JanWieck@Yahoo.com> writes: > > Shridhar Daithankar wrote: > >> Umm.. What does FSM does then? I was under impression that FSM stores page > >> pointers and vacuum work on FSM information only. In that case, it wouldn't > >> have to waste time to find out which pages to clean. > > > It's the other way around! VACUUM scan's the tables to find and reclaim > > free space and remembers that free space in the FSM. > > Right. One big question mark in my mind about these "partial vacuum" > proposals is whether they'd still allow adequate FSM information to be > maintained. If VACUUM isn't looking at most of the pages, there's no > very good way to acquire info about where there's free space. Somehow it needs to get two types of information. A. If any transaction is accessing a page B. If a page contains any free space. Vacuum needs to look for pages not in A but in B. Can storage manager maintain two lists/hashes with minimal cost? In that case, all unlocked and not in transaction pages could be a much smaller subset. Does it sound bizzare? Bye Shridhar -- Chemicals, n.: Noxious substances from which modern foods are made.
Tom Lane wrote:
> Jan Wieck <JanWieck@Yahoo.com> writes:
>> Shridhar Daithankar wrote:
>>> Umm.. What does FSM does then? I was under impression that FSM stores page
>>> pointers and vacuum work on FSM information only. In that case, it wouldn't
>>> have to waste time to find out which pages to clean.
>
>> It's the other way around! VACUUM scan's the tables to find and reclaim
>> free space and remembers that free space in the FSM.
>
> Right. One big question mark in my mind about these "partial vacuum"
> proposals is whether they'd still allow adequate FSM information to be
> maintained. If VACUUM isn't looking at most of the pages, there's no
> very good way to acquire info about where there's free space.
That's why I think it needs one more pg_stat column to count the number
of vacuumed tuples. If one does
tuples_updated + tuples_deleted - tuples_vacuumed
he'll get approximately the number of tuples a regular vacuum might be
able to reclaim. If that number is really small, no need for autovacuum
to cause any big trouble by scanning the relation.
Another way to give autovacuum some hints would be to return some number
as commandtuples from vacuum. like the number of tuples actually
vacuumed. That together with the new number of reltuples in pg_class
will tell autovacuum how frequent a relation really needs scanning.
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: > Another way to give autovacuum some hints would be to return some number > as commandtuples from vacuum. like the number of tuples actually > vacuumed. That together with the new number of reltuples in pg_class > will tell autovacuum how frequent a relation really needs scanning. Which actually would be much better because it'd work without the statistics collector configured for gathering IO stats. 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 22 Aug 2003 at 11:03, Jan Wieck wrote: > Tom Lane wrote: > > > Jan Wieck <JanWieck@Yahoo.com> writes: > >> Shridhar Daithankar wrote: > >>> Umm.. What does FSM does then? I was under impression that FSM stores page > >>> pointers and vacuum work on FSM information only. In that case, it wouldn't > >>> have to waste time to find out which pages to clean. > > > >> It's the other way around! VACUUM scan's the tables to find and reclaim > >> free space and remembers that free space in the FSM. > > > > Right. One big question mark in my mind about these "partial vacuum" > > proposals is whether they'd still allow adequate FSM information to be > > maintained. If VACUUM isn't looking at most of the pages, there's no > > very good way to acquire info about where there's free space. > > That's why I think it needs one more pg_stat column to count the number > of vacuumed tuples. If one does > > tuples_updated + tuples_deleted - tuples_vacuumed > > he'll get approximately the number of tuples a regular vacuum might be > able to reclaim. If that number is really small, no need for autovacuum > to cause any big trouble by scanning the relation. > > Another way to give autovacuum some hints would be to return some number > as commandtuples from vacuum. like the number of tuples actually > vacuumed. That together with the new number of reltuples in pg_class > will tell autovacuum how frequent a relation really needs scanning. This kind of information does not really help autovacuum. If we are talking about modifying backend stat collection algo., so that vacuum does minimum work, is has translate to cheaper vacuum analyze so that autovacuum can fire it at will any time. In the best case, another resident process like stat collector can keep cleaning the deads. This information must be in terms of pages and actually be maintained as per page stat. Looking at number of tuples values does not give any idea to vacuum how it is going to flush cache lines, either in postgresql or on OS. I doubt it will help vacuum command in itself to be any lighter or more efficient. If it is easy to do, I would favour maitaining two page maps as I mentioned in another mail. One for pages in cache but not locked by any transaction and another for pages which has some free space. If it is rare for a page to be full, we can skip the later one. I think that could be good enough. Bye Shridhar -- Office Automation: The use of computers to improve efficiency in the office by removing anyone you would want to talk with over coffee.
On Fri, 2003-08-22 at 11:08, Jan Wieck wrote: > > Another way to give autovacuum some hints would be to return some number > > as commandtuples from vacuum. like the number of tuples actually > > vacuumed. That together with the new number of reltuples in pg_class > > will tell autovacuum how frequent a relation really needs scanning. > > Which actually would be much better because it'd work without the > statistics collector configured for gathering IO stats. Which is certainly a good thing. Using the stats system is a measurable performance hit. I still want to play with pg_autovacuum ignoring the stats system and just looking at the FSM data.
On Fri, 2003-08-22 at 11:17, Shridhar Daithankar wrote: > On 22 Aug 2003 at 11:03, Jan Wieck wrote: > > That's why I think it needs one more pg_stat column to count the number > > of vacuumed tuples. If one does > > > > tuples_updated + tuples_deleted - tuples_vacuumed > > > > he'll get approximately the number of tuples a regular vacuum might be > > able to reclaim. If that number is really small, no need for autovacuum > > to cause any big trouble by scanning the relation. > > > > Another way to give autovacuum some hints would be to return some number > > as commandtuples from vacuum. like the number of tuples actually > > vacuumed. That together with the new number of reltuples in pg_class > > will tell autovacuum how frequent a relation really needs scanning. > > This kind of information does not really help autovacuum. If we are talking > about modifying backend stat collection algo., so that vacuum does minimum > work, is has translate to cheaper vacuum analyze so that autovacuum can fire it > at will any time. In the best case, another resident process like stat > collector can keep cleaning the deads. I believe what Jan is talking about is knowing when to use a normal vacuum, and when to do a vacuum decent. So his proposal is working under the assumption that there would be a cheaper vacuum analyze that can be run most of the time.
On Fri, 2003-08-22 at 10:45, Tom Lane wrote: > Jan Wieck <JanWieck@Yahoo.com> writes: > Right. One big question mark in my mind about these "partial vacuum" > proposals is whether they'd still allow adequate FSM information to be > maintained. If VACUUM isn't looking at most of the pages, there's no > very good way to acquire info about where there's free space. Well, pg_autovacuum really needs to be looking at the FSM anyway. It could look at the FSM, and choose to to do a vacuum normal when there the amount of FSM data becomes inadequate. Of course I'm not sure how you would differentiate a busy table with "inadequate" FSM data and an inactive table that doesn't even register in the FSM. Perhaps you would still need to consult the stats system.
On Fri, 22 Aug 2003 10:45:50 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote: >One big question mark in my mind about these "partial vacuum" >proposals is whether they'd still allow adequate FSM information to be >maintained. If VACUUM isn't looking at most of the pages, there's no >very good way to acquire info about where there's free space. VACUUM has accurate information about the pages it just visited. Free space information for pages not touched by VACUUM is still in the FSM, unless free space on a page is too low to be interesting. VACUUM has to merge these two lists and throw away entries with little free space if running out of room. Thus we might end up with new almost full pages in the FSM while there are pages with more free space lying around that a previous VACUUM failed to register because there was more free space at that time. Considering that . FSM is lossy per definitionem . we are targeting at relations with large passive areas . decent VACUUM shall not replace lazy VACUUM I see no problem here. Future advice could be: "VACCUM DECENT every hour, VACUUM daily, VACUUM FULL once a year" where the first two could be scheduled by autovacuum ... Servus Manfred
Okay, my proposal would be to have a VACUUM mode where it tells the buffer manager to only return a page if it is already in memory, and some "not cached" if it would have to read it from disk, and simply skip the page in that case. Probably needs some modifications in vacuums FSM handling, but basically that's it. It'll still cause IO for the resulting index bulk cleaning, so I don't know how efficient it'll be after all. The number of vacuumed tuples returned will tell the autovacuum how useful this vacuum scan was. The less useful it is, the less frequent it'll be scheduled. There is no point in vacuuming a 50M row table every hour when the average number of tuples reclaimed is in the hundreds. I don't intend to avoid a full table scan completely. I only intend to lower the frequency of them. It will require some fuzzy logic in autovacuum to figure out if a CACHEONLY vacuum for a table needs to be more or less frequent to find more tuples though. So far for what I have in mind. Now what are you proposing down there? Where do you intend to hold that "per page stat" and what exactly is maintaining it? And please don't give us any vague "some other resident process". This only indicates you don't really know what it requires for a process to be able to read or write data in PostgreSQL. Jan Shridhar Daithankar wrote: > On 22 Aug 2003 at 11:03, Jan Wieck wrote: > >> Tom Lane wrote: >> >> > Jan Wieck <JanWieck@Yahoo.com> writes: >> >> Shridhar Daithankar wrote: >> >>> Umm.. What does FSM does then? I was under impression that FSM stores page >> >>> pointers and vacuum work on FSM information only. In that case, it wouldn't >> >>> have to waste time to find out which pages to clean. >> > >> >> It's the other way around! VACUUM scan's the tables to find and reclaim >> >> free space and remembers that free space in the FSM. >> > >> > Right. One big question mark in my mind about these "partial vacuum" >> > proposals is whether they'd still allow adequate FSM information to be >> > maintained. If VACUUM isn't looking at most of the pages, there's no >> > very good way to acquire info about where there's free space. >> >> That's why I think it needs one more pg_stat column to count the number >> of vacuumed tuples. If one does >> >> tuples_updated + tuples_deleted - tuples_vacuumed >> >> he'll get approximately the number of tuples a regular vacuum might be >> able to reclaim. If that number is really small, no need for autovacuum >> to cause any big trouble by scanning the relation. >> >> Another way to give autovacuum some hints would be to return some number >> as commandtuples from vacuum. like the number of tuples actually >> vacuumed. That together with the new number of reltuples in pg_class >> will tell autovacuum how frequent a relation really needs scanning. > > This kind of information does not really help autovacuum. If we are talking > about modifying backend stat collection algo., so that vacuum does minimum > work, is has translate to cheaper vacuum analyze so that autovacuum can fire it > at will any time. In the best case, another resident process like stat > collector can keep cleaning the deads. > > This information must be in terms of pages and actually be maintained as per > page stat. Looking at number of tuples values does not give any idea to vacuum > how it is going to flush cache lines, either in postgresql or on OS. I doubt it > will help vacuum command in itself to be any lighter or more efficient. > > If it is easy to do, I would favour maitaining two page maps as I mentioned in > another mail. One for pages in cache but not locked by any transaction and > another for pages which has some free space. If it is rare for a page to be > full, we can skip the later one. I think that could be good enough. > > > > > Bye > Shridhar > > -- > Office Automation: The use of computers to improve efficiency in the office by > removing anyone you would want to talk with over coffee. > > > ---------------------------(end of broadcast)--------------------------- > TIP 7: don't forget to increase your free space map settings -- #======================================================================# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #================================================== JanWieck@Yahoo.com #
On Fri, 22 Aug 2003 12:18:02 -0400, Jan Wieck <JanWieck@Yahoo.com> wrote: >Okay, my proposal would be to have a VACUUM mode where it tells the >buffer manager to only return a page if it is already in memory But how can it know? Yes, we know exactly what we have in PG shared buffers. OTOH we keep telling people that they should configure moderate values for shared_buffers because the OS is better at caching. Your CACHEONLY VACUUM wouldn't catch those pages that are in the OS cache but not in the shared buffers, although they are retrievable at almost the same low cost. We should not try to avoid _any_ physical disk access. It's good enough to avoid useless reads. Hence my proposal for a reclaimable space list ... Servus Manfred
Jan Wieck <JanWieck@Yahoo.com> writes:
> Okay, my proposal would be to have a VACUUM mode where it tells the
> buffer manager to only return a page if it is already in memory, and
> some "not cached" if it would have to read it from disk, and simply skip
> the page in that case.
Since no such call is available at the OS level, this would only work
well with very large shared_buffers settings (ie, you try to rely on
PG shared buffers to the exclusion of kernel disk cache). AFAIK the
general consensus is that that's not a good way to run Postgres.
regards, tom lane
Tom Lane wrote: > Jan Wieck <JanWieck@Yahoo.com> writes: >> Okay, my proposal would be to have a VACUUM mode where it tells the >> buffer manager to only return a page if it is already in memory, and >> some "not cached" if it would have to read it from disk, and simply skip >> the page in that case. > > Since no such call is available at the OS level, this would only work > well with very large shared_buffers settings (ie, you try to rely on > PG shared buffers to the exclusion of kernel disk cache). AFAIK the > general consensus is that that's not a good way to run Postgres. Oh-kay ... so yes Manfred, your RSM is probably the better way. 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: > Tom Lane wrote: > > > Jan Wieck <JanWieck@Yahoo.com> writes: > >> Okay, my proposal would be to have a VACUUM mode where it tells the > >> buffer manager to only return a page if it is already in memory, and > >> some "not cached" if it would have to read it from disk, and simply skip > >> the page in that case. > > > > Since no such call is available at the OS level, this would only work > > well with very large shared_buffers settings (ie, you try to rely on > > PG shared buffers to the exclusion of kernel disk cache). AFAIK the > > general consensus is that that's not a good way to run Postgres. > > Oh-kay ... so yes Manfred, your RSM is probably the better way. Added to TODO. -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073