Обсуждение: optimizing a query over tree-like structure
hi. sorry for the vague syntax used below, but the query is huge so i've tried to present it in simple terms. And sorry if i'm doing obviously stupid things, i have lots of years programming behind me but NO sql involved. i have a somewhat tree-like structure of objects that link to each other via many2many associations. it looks like: (N is "root")N links to R,P,FR links to F P links to O,FO links to O,F #recursivelyF links to Z All links to F but the one in O are "shortcuts", to avoid looking it up recursively. each of these objects has some associated values (again many2many, ownership). what i want is to get all the values related to a given N and its sublevels, in one query. one variant of what i've invented so far is (~pseudocode, no recursion on O): SELECT ownership.*, value.* FROM Nazn, mm_N2P, mm_P2O, mm_O2O, mm_O2O AS mm_O2O1, mm_N2Z, ownership JOIN value ON ownership.value = value.dbid WHERE (N.dbid = ownership.NORN.dbid = mm_N2R.left AND mm_N2R.right = ownership.RORN.dbid = mm_N2P.left AND ( mm_N2P.right= ownership.P OR mm_N2P.right = mm_P2O.left AND ( mm_P2O.right = ownership.O OR mm_P2O.right= mm_O2O.left AND ( mm_O2O.right = ownership.O OR mm_O2O.right = mm_O2O1.leftAND mm_O2O1.right = ownership.O)))ORNazn.dbid = mm_N2F.left AND ( mm_N2F.right = ownership.F OR mm_N2Z.right = ownership.Z) ) AND ownership.value = value.dbid AND N.obj = whatever-filter-by-N ---------------- this scales very poor. it uses the shortcut to F present in N. for just 200 rows with related associations, it takes 4 seconds to get result. if i use the shortcut to F present in P, it takes 2 seconds - but thats still inacceptable. seems that the number or consequtive ORs on same level is killing it. EXPLAIN gives nested loops all over. What am i doing wrong here? should i expand the A-to-B links of the sort mm_N2P.right = mm_P2O.left into mm_N2P.right = P.dbid and P.dbid == mm_P2O.left ? the query is generated via sqlalchemy and a layer on top, so i can tweak it any way required (and it has many other sub/filterings which i've ommited for brevity - they dont make it better/worse). any pointers of how such queries should be written are appreciated - e.g. what is considered fine, what doable and what is a no-no. thanks ahead ciao svil www.svilendobrev.com dbcook.sf.net
Hi, Svil I 'd like to first fully understand the background of your problem before figurin out if I can be of any help (or not). You have a tree, of which N is the root, is this correct? Then, what are the next sublevel? F, P and R? If so, why is R linked to a sibling (F) ? And the next one? O and Z? Is O connected to itself? And i am not understanding your concept of "shortcuts". Could you please explain ? What kind of tree do you have exactly? Binary? Trenary? The mm_* tables keep relations between nodes, I guess.... If so , the mm_N2Z one is empty, in this example, right?As there is no edge from N to Z (not direct). But what is the Nazn table? What records does it keep? And what is the ownership table? And the value? Could you tell which columns these tables have, at least the relevant ones for your problem ? Please kindly advice me on these issues. I am not very specialized in optimizing queries, but I see you have a lot of cartesian products on your FROM clause, which, from my own experience, I guess it has tendency to be slow... Best, Oliveiros ----- Original Message ----- From: <az@svilendobrev.com> To: <pgsql-sql@postgresql.org> Sent: Tuesday, September 30, 2008 9:32 AM Subject: [SQL] optimizing a query over tree-like structure > hi. > sorry for the vague syntax used below, but the query is huge so i've > tried to present it in simple terms. And sorry if i'm doing obviously > stupid things, i have lots of years programming behind me but NO sql > involved. > i have a somewhat tree-like structure of objects that link to each > other via many2many associations. it looks like: > (N is "root") > N links to R,P,F > R links to F > P links to O,F > O links to O,F #recursively > F links to Z > All links to F but the one in O are "shortcuts", to avoid looking it > up recursively. > each of these objects has some associated values (again many2many, > ownership). > > what i want is to get all the values related to a given N and its > sublevels, in one query. > > one variant of what i've invented so far is (~pseudocode, no recursion > on O): > > SELECT ownership.*, value.* > FROM Nazn, mm_N2P, mm_P2O, mm_O2O, mm_O2O AS mm_O2O1, mm_N2Z, > ownership JOIN value ON ownership.value = value.dbid > WHERE ( > N.dbid = ownership.N > OR > N.dbid = mm_N2R.left AND mm_N2R.right = ownership.R > OR > N.dbid = mm_N2P.left AND ( > mm_N2P.right = ownership.P > OR > mm_N2P.right = mm_P2O.left AND ( > mm_P2O.right = ownership.O > OR > mm_P2O.right = mm_O2O.left AND ( > mm_O2O.right = ownership.O > OR > mm_O2O.right = mm_O2O1.left AND > mm_O2O1.right = ownership.O > ))) > OR > Nazn.dbid = mm_N2F.left AND ( > mm_N2F.right = ownership.F > OR > mm_N2Z.right = ownership.Z > ) > ) AND ownership.value = value.dbid AND N.obj = whatever-filter-by-N > > ---------------- > this scales very poor. > it uses the shortcut to F present in N. > for just 200 rows with related associations, it takes 4 seconds to get > result. > if i use the shortcut to F present in P, it takes 2 seconds - but > thats still inacceptable. > seems that the number or consequtive ORs on same level is killing it. > EXPLAIN gives nested loops all over. > What am i doing wrong here? > should i expand the A-to-B links of the sort > mm_N2P.right = mm_P2O.left > into > mm_N2P.right = P.dbid and P.dbid == mm_P2O.left ? > > the query is generated via sqlalchemy and a layer on top, so i can > tweak it any way required (and it has many other sub/filterings which > i've ommited for brevity - they dont make it better/worse). > > any pointers of how such queries should be written are appreciated - > e.g. what is considered fine, what doable and what is a no-no. > > thanks ahead > ciao > svil > > www.svilendobrev.com > dbcook.sf.net > > -- > Sent via pgsql-sql mailing list (pgsql-sql@postgresql.org) > To make changes to your subscription: > http://www.postgresql.org/mailpref/pgsql-sql
another idea i just got, to decrease the number of tables in one FROM, is to represent the first-level ORs with unions. but i'm not sure how exactly to do it: are these equivalent? FROM N,ownership,mm_N2R where ( N.dbid = ownership.NORN.dbid = mm_N2R.left AND mm_N2R.right = ownership.ROR ... ) AND N.obj = whatever-filter-by-N -------- FROM ( (ownership join N on ownership.N = N.dbid)union (ownership join mm_N2R on ownership.R = mm_N2R.right join N on mm_N2R.left = N.dbid )OR ) ... and should i bundle the filtering-by-N/Employment in every of above union-members? On Tuesday 30 September 2008 15:21:09 Oliveiros Cristina wrote: > Hi, Svil > > I 'd like to first fully understand the background of your problem > before figurin out if I can be of any help (or not). > > You have a tree, of which N is the root, is this correct? > Then, what are the next sublevel? > F, P and R? If so, why is R linked to a sibling (F) ? > And the next one? > O and Z? > Is O connected to itself? > And i am not understanding your concept of "shortcuts". Could you > please explain ? > What kind of tree do you have exactly? Binary? Trenary? > The mm_* tables keep relations between nodes, I guess.... If so , > the mm_N2Z one is empty, in this example, right?As there is no edge > from N to Z (not direct). > But what is the Nazn table? What records does it keep? > And what is the ownership table? And the value? > Could you tell which columns these tables have, at least the > relevant ones for your problem ? > > Please kindly advice me on these issues. > I am not very specialized in optimizing queries, but I see you have > a lot of cartesian products on your FROM clause, which, from my own > experience, I guess it has tendency to be slow... > > Best, > > Oliveiros > > ----- Original Message ----- > From: <az@svilendobrev.com> > To: <pgsql-sql@postgresql.org> > Sent: Tuesday, September 30, 2008 9:32 AM > Subject: [SQL] optimizing a query over tree-like structure > > > hi. > > sorry for the vague syntax used below, but the query is huge so > > i've tried to present it in simple terms. And sorry if i'm doing > > obviously stupid things, i have lots of years programming behind > > me but NO sql involved. > > i have a somewhat tree-like structure of objects that link to > > each other via many2many associations. it looks like: > > (N is "root") > > N links to R,P,F > > R links to F > > P links to O,F > > O links to O,F #recursively > > F links to Z > > All links to F but the one in O are "shortcuts", to avoid looking > > it up recursively. > > each of these objects has some associated values (again > > many2many, ownership). > > > > what i want is to get all the values related to a given N and its > > sublevels, in one query. > > > > one variant of what i've invented so far is (~pseudocode, no > > recursion on O): > > > > SELECT ownership.*, value.* > > FROM Nazn, mm_N2P, mm_P2O, mm_O2O, mm_O2O AS mm_O2O1, mm_N2Z, > > ownership JOIN value ON ownership.value = value.dbid > > WHERE ( > > N.dbid = ownership.N > > OR > > N.dbid = mm_N2R.left AND mm_N2R.right = ownership.R > > OR > > N.dbid = mm_N2P.left AND ( > > mm_N2P.right = ownership.P > > OR > > mm_N2P.right = mm_P2O.left AND ( > > mm_P2O.right = ownership.O > > OR > > mm_P2O.right = mm_O2O.left AND ( > > mm_O2O.right = ownership.O > > OR > > mm_O2O.right = mm_O2O1.left AND > > mm_O2O1.right = ownership.O > > ))) > > OR > > Nazn.dbid = mm_N2F.left AND ( > > mm_N2F.right = ownership.F > > OR > > mm_N2Z.right = ownership.Z > > ) > > ) AND ownership.value = value.dbid AND N.obj = > > whatever-filter-by-N > > > > ---------------- > > this scales very poor. > > it uses the shortcut to F present in N. > > for just 200 rows with related associations, it takes 4 seconds > > to get result. > > if i use the shortcut to F present in P, it takes 2 seconds - but > > thats still inacceptable. > > seems that the number or consequtive ORs on same level is killing > > it. EXPLAIN gives nested loops all over. > > What am i doing wrong here? > > should i expand the A-to-B links of the sort > > mm_N2P.right = mm_P2O.left > > into > > mm_N2P.right = P.dbid and P.dbid == mm_P2O.left ? > > > > the query is generated via sqlalchemy and a layer on top, so i > > can tweak it any way required (and it has many other > > sub/filterings which i've ommited for brevity - they dont make it > > better/worse). > > > > any pointers of how such queries should be written are > > appreciated - e.g. what is considered fine, what doable and what > > is a no-no. > > > > thanks ahead > > ciao > > svil > > > > www.svilendobrev.com > > dbcook.sf.net > > > > -- > > Sent via pgsql-sql mailing list (pgsql-sql@postgresql.org) > > To make changes to your subscription: > > http://www.postgresql.org/mailpref/pgsql-sql
Svil, Please advice me, You have values and one table for N,R,P,F and O and Z, right? And you have ownership which is a "catch-all" associative table between values and whatever other table, is this correct? You want to retrieve the values for a certain N, and all to all the other entities that belong to that N, is this correct? Fine, I have a few questions on your associative tablesmm_N2P -- An N can be associated to P in a many to many relationship? N Can have one or none P, am I right? It can have many Ps ? And a P can belong to many Ns ? mm_P2O, -- I understand an O can have many Positions, but a Positon can be in more than one O? is this a many to many relationship? mm_O2O, -- if it is a hierarchical tree do you need an associative table? An O can have many sub-departments but a sub-department (O) can just belong to a (super-)department, -- isn't it so?mm_N2Z -- this one was mistake of yours, it doesn't exist, right? You wrote that giant query by yourself or was it generated by some automated tool? (sqlalchemy , I have no idea :p ) Your ultimate goal is to retrieve the values associated with a certain Department( and all its sub-departments) which is associated with a certain Position, which values are also to obtain, such position being associated with a certain N, which values are also to obtain, is my understandin correct? Explain me the simplest case, where you have an N not associated with any R or P, just with an F. Which info is to be retrieved, exactly in this case ? Best, Oliveiros ----- Original Message ----- From: <az@svilendobrev.com> To: <pgsql-sql@postgresql.org> Sent: Tuesday, September 30, 2008 3:24 PM Subject: Re: [SQL] optimizing a query over tree-like structure > another idea i just got, to decrease the number of tables in one FROM, > is to represent the first-level ORs with unions. but i'm not sure how > exactly to do it: are these equivalent? > > FROM N,ownership,mm_N2R > where ( > N.dbid = ownership.N > OR > N.dbid = mm_N2R.left AND mm_N2R.right = ownership.R > OR > ... > ) AND N.obj = whatever-filter-by-N > > -------- > > FROM ( > (ownership join N on ownership.N = N.dbid) > union > (ownership join mm_N2R on ownership.R = mm_N2R.right join N on > mm_N2R.left = N.dbid ) > OR > ) ... > > and should i bundle the filtering-by-N/Employment in every of above > union-members? > > On Tuesday 30 September 2008 15:21:09 Oliveiros Cristina wrote: >> Hi, Svil >> >> I 'd like to first fully understand the background of your problem >> before figurin out if I can be of any help (or not). >> >> You have a tree, of which N is the root, is this correct? >> Then, what are the next sublevel? >> F, P and R? If so, why is R linked to a sibling (F) ? >> And the next one? >> O and Z? >> Is O connected to itself? >> And i am not understanding your concept of "shortcuts". Could you >> please explain ? >> What kind of tree do you have exactly? Binary? Trenary? >> The mm_* tables keep relations between nodes, I guess.... If so , >> the mm_N2Z one is empty, in this example, right?As there is no edge >> from N to Z (not direct). >> But what is the Nazn table? What records does it keep? >> And what is the ownership table? And the value? >> Could you tell which columns these tables have, at least the >> relevant ones for your problem ? >> >> Please kindly advice me on these issues. >> I am not very specialized in optimizing queries, but I see you have >> a lot of cartesian products on your FROM clause, which, from my own >> experience, I guess it has tendency to be slow... >> >> Best, >> >> Oliveiros >> >> ----- Original Message ----- >> From: <az@svilendobrev.com> >> To: <pgsql-sql@postgresql.org> >> Sent: Tuesday, September 30, 2008 9:32 AM >> Subject: [SQL] optimizing a query over tree-like structure >> >> > hi. >> > sorry for the vague syntax used below, but the query is huge so >> > i've tried to present it in simple terms. And sorry if i'm doing >> > obviously stupid things, i have lots of years programming behind >> > me but NO sql involved. >> > i have a somewhat tree-like structure of objects that link to >> > each other via many2many associations. it looks like: >> > (N is "root") >> > N links to R,P,F >> > R links to F >> > P links to O,F >> > O links to O,F #recursively >> > F links to Z >> > All links to F but the one in O are "shortcuts", to avoid looking >> > it up recursively. >> > each of these objects has some associated values (again >> > many2many, ownership). >> > >> > what i want is to get all the values related to a given N and its >> > sublevels, in one query. >> > >> > one variant of what i've invented so far is (~pseudocode, no >> > recursion on O): >> > >> > SELECT ownership.*, value.* >> > FROM Nazn, mm_N2P, mm_P2O, mm_O2O, mm_O2O AS mm_O2O1, mm_N2Z, >> > ownership JOIN value ON ownership.value = value.dbid >> > WHERE ( >> > N.dbid = ownership.N >> > OR >> > N.dbid = mm_N2R.left AND mm_N2R.right = ownership.R >> > OR >> > N.dbid = mm_N2P.left AND ( >> > mm_N2P.right = ownership.P >> > OR >> > mm_N2P.right = mm_P2O.left AND ( >> > mm_P2O.right = ownership.O >> > OR >> > mm_P2O.right = mm_O2O.left AND ( >> > mm_O2O.right = ownership.O >> > OR >> > mm_O2O.right = mm_O2O1.left AND >> > mm_O2O1.right = ownership.O >> > ))) >> > OR >> > Nazn.dbid = mm_N2F.left AND ( >> > mm_N2F.right = ownership.F >> > OR >> > mm_N2Z.right = ownership.Z >> > ) >> > ) AND ownership.value = value.dbid AND N.obj = >> > whatever-filter-by-N >> > >> > ---------------- >> > this scales very poor. >> > it uses the shortcut to F present in N. >> > for just 200 rows with related associations, it takes 4 seconds >> > to get result. >> > if i use the shortcut to F present in P, it takes 2 seconds - but >> > thats still inacceptable. >> > seems that the number or consequtive ORs on same level is killing >> > it. EXPLAIN gives nested loops all over. >> > What am i doing wrong here? >> > should i expand the A-to-B links of the sort >> > mm_N2P.right = mm_P2O.left >> > into >> > mm_N2P.right = P.dbid and P.dbid == mm_P2O.left ? >> > >> > the query is generated via sqlalchemy and a layer on top, so i >> > can tweak it any way required (and it has many other >> > sub/filterings which i've ommited for brevity - they dont make it >> > better/worse). >> > >> > any pointers of how such queries should be written are >> > appreciated - e.g. what is considered fine, what doable and what >> > is a no-no. >> > >> > thanks ahead >> > ciao >> > svil >> > >> > www.svilendobrev.com >> > dbcook.sf.net >> > >> > -- >> > Sent via pgsql-sql mailing list (pgsql-sql@postgresql.org) >> > To make changes to your subscription: >> > http://www.postgresql.org/mailpref/pgsql-sql > > > > -- > Sent via pgsql-sql mailing list (pgsql-sql@postgresql.org) > To make changes to your subscription: > http://www.postgresql.org/mailpref/pgsql-sql