Thread: pg_locks view and user locks
I've been knocking around a bit with user level locks and I have a few suggestions/questions: First, the pg_locks view lists user level locks but doesn't really say much about them. For example, for traditional locks we get the xid but for user locks we don't really get anything useful. I looked in lockfuncs.c and didn't see that there would be any real penalty to displaying the offset/block of the user lock, and returning it in request from pg_locks. Is this view frequently queried by system processes? Why would we want to do this? Well it makes resolving a user lock to a specific pid easier...admittedly of dubious value but helpful to me...although if this doesn't fly I can always create an alternate view which serves my purpose (and perhaps add a high level function to the /contrib userlock module). Second, Is there a reason why user level locks are completely undocumented? AFAICT, There is no mention of them in anywhere in the docs, particularly 12.4, which describes methods for application managed concurrency. The availability of cooperative long term locks is (IMO) a really nice feature, particularly for people porting legacy applications which depend on explicit locking (there is some good info in the user lock module which is unfortunately under the GPL). Not complaining or looking for help, just curious why they seemed to have slipped through the cracks. Merlin
Merlin Moncure wrote: > Second, > Is there a reason why user level locks are completely undocumented? > AFAICT, There is no mention of them in anywhere in the docs, > particularly 12.4, which describes methods for application managed > concurrency. The availability of cooperative long term locks is > (IMO) a really nice feature, particularly for people porting legacy > applications which depend on explicit locking (there is some good > info in the user lock module which is unfortunately under the GPL). That's the whole problem. I don't think anyone objects to the user lock principle, but as long as it's GPL, we won't support it. It was probably a mistake to accept this module in the first place. I believe some people have been trying to get the module relicensed, but that evidently never happened. -- Peter Eisentraut http://developer.postgresql.org/~petere/
On Mon, Sep 13, 2004 at 01:34:05PM -0400, Merlin Moncure wrote: > First, the pg_locks view lists user level locks but doesn't really say > much about them. For example, for traditional locks we get the xid but > for user locks we don't really get anything useful. I looked in > lockfuncs.c and didn't see that there would be any real penalty to > displaying the offset/block of the user lock, and returning it in > request from pg_locks. ISTM it would be better to have pg_locks show only system locks, and another view, say pg_userlocks, to show user locks. This would allow to show different data; for example, the PID of the process involved. What's more, user locks and system locks never conflict, so it seems wrong to show them together. Another option would be to add another column to pg_locks to say what lockmethod (1 for system, 2 for user) is used in each lock. Maybe we want to separate locks even more, and use one table to show xact locks, another to show table locks, and yet another to show user locks. But maybe this is a bad idea right from the start. > Is this view frequently queried by system processes? I don't think there is a lot of server-side stuff (aside from possible UDFs) that use pg_locks directly. It's there only for user convenience. -- Alvaro Herrera (<alvherre[a]dcc.uchile.cl>) "Someone said that it is at least an order of magnitude more work to do production software than a prototype. I think he is wrong by at least an order of magnitude." (Brian Kernighan)
> Merlin Moncure wrote: > > Second, > > Is there a reason why user level locks are completely undocumented? > > AFAICT, There is no mention of them in anywhere in the docs, > > particularly 12.4, which describes methods for application managed > > concurrency. The availability of cooperative long term locks is > > (IMO) a really nice feature, particularly for people porting legacy > > applications which depend on explicit locking (there is some good > > info in the user lock module which is unfortunately under the GPL). > > That's the whole problem. I don't think anyone objects to the user lock > principle, but as long as it's GPL, we won't support it. It was > probably a mistake to accept this module in the first place. I believe > some people have been trying to get the module relicensed, but that > evidently never happened. Well, the lock module itself is nothing more than some wrappers for the system lock manager functions (LockAcquire,etc) in C and wrappers for those in SQL, plus a makefile and some basic documentation...in other words almost nothing. If the 'userlock' module author can't be found or refuses to open his module, why not simply rewrite it? It's extremely small. This also gives an opportunity to add some missing parts, like resolving a lock to a pid and better error handling. In any case, IMHO it is worthwhile to flesh out 12.3, adding user locks (simply listing as a lock type), 12.4 with usage of above, perhaps with an example. Also a mention of userlocks in 43.32 and possibly a new entry in this section, depending on how they get handled per my suggestion wrt pg_locks. Also possibly a note in 16.4.8 (postgresql.conf entry for max_locks_per_transaction) describing the interaction between this parameter and user locks (aside: is there any?). Certain aspects of these changes would of course depend on the feasibility of moving some of the lock interfaces into the main project. I'd be willing to have a stab at the above, is it worthwhile? Merlin
Peter Eisentraut <peter_e@gmx.net> writes: > That's the whole problem. I don't think anyone objects to the user lock > principle, but as long as it's GPL, we won't support it. It was > probably a mistake to accept this module in the first place. I believe > some people have been trying to get the module relicensed, but that > evidently never happened. I said some time ago that I would pester contrib authors about fixing license issues, but it still hasn't gotten to the top of my to-do list :-( AFAIK nothing at all has been done on this score. If Merlin is hot about contrib/userlock right now, he's welcome to try to contact the original author and get this dealt with. The point about not importing documentation from a mislicensed contrib module is valid IMHO. regards, tom lane
> I said some time ago that I would pester contrib authors about fixing > license issues, but it still hasn't gotten to the top of my to-do list :-( > AFAIK nothing at all has been done on this score. > > If Merlin is hot about contrib/userlock right now, he's welcome to try > to contact the original author and get this dealt with. > > The point about not importing documentation from a mislicensed contrib > module is valid IMHO. I'll try to contact the author and see what comes up. Supposing the userlock module is re-licensed (or a suitable replacement is contributed and accepted), is there any merit to: Promoting the user lock wrappers out of contrib (plus fleshing them out a bit)? Adjusting the pg_locks view to return the user lock specific information (or, as Alvaro suggested, to add a new view)? A general documentation update wrt user locks? Merlin
"Merlin Moncure" <merlin.moncure@rcsonline.com> writes: > ... is there any merit to promoting the user lock wrappers out of contrib Dunno. Yours is the first message I can recall in quite a long time indicating that anyone was using userlocks. I thought the code was kind of dying on the vine. Of course it's hard to tell whether that's not just *because* it's in contrib and not mainstream. But personally I'd like to see some more evidence of use before we promote it. (And yeah, the API could probably use some cleanup first.) regards, tom lane
Alvaro Herrera <alvherre@dcc.uchile.cl> writes: > ISTM it would be better to have pg_locks show only system locks, and > another view, say pg_userlocks, to show user locks. This would allow to > show different data; for example, the PID of the process involved. I think this is a bad idea, at least for the base-level view, because it would imply that it's impossible to get a truly simultaneous view of the state of all the locks. It would certainly be a bad idea to so separate xact and table locks. I'd prefer to extend the present approach and add columns that are NULL when the type of lock isn't relevant to them. > Another option would be to add another column to pg_locks to say what > lockmethod (1 for system, 2 for user) is used in each lock. How about a text column with values "TABLE", "XACT", "USER"? regards, tom lane
This is the first I have ever heard "user locks," but I have more than once wanted a lock that would persist beyond the end of a transaction. Do these do that? Paul >"Merlin Moncure" <merlin.moncure@rcsonline.com> writes: > > >>... is there any merit to promoting the user lock wrappers out of contrib >> >> > >Dunno. Yours is the first message I can recall in quite a long time >indicating that anyone was using userlocks. I thought the code was kind >of dying on the vine. Of course it's hard to tell whether that's not >just *because* it's in contrib and not mainstream. But personally I'd >like to see some more evidence of use before we promote it. (And yeah, >the API could probably use some cleanup first.) > > regards, tom lane > >---------------------------(end of broadcast)--------------------------- >TIP 5: Have you checked our extensive FAQ? > > http://www.postgresql.org/docs/faqs/FAQ.html > > > >
On Mon, Sep 13, 2004 at 08:06:08PM -0400, Paul Tillotson wrote: > This is the first I have ever heard "user locks," but I have more than > once wanted a lock that would persist beyond the end of a transaction. > Do these do that? Yes. -- Alvaro Herrera (<alvherre[a]dcc.uchile.cl>) "El destino baraja y nosotros jugamos" (A. Schopenhauer)
> "Merlin Moncure" <merlin.moncure@rcsonline.com> writes: > > ... is there any merit to promoting the user lock wrappers out of > contrib > > Dunno. Yours is the first message I can recall in quite a long time > indicating that anyone was using userlocks. I thought the code was kind > of dying on the vine. Of course it's hard to tell whether that's not > just *because* it's in contrib and not mainstream. But personally I'd > like to see some more evidence of use before we promote it. (And yeah, > the API could probably use some cleanup first.) Well, what's interesting about user locks is that all the real substantive stuff is already built into the lock manager. The userlock contrib module, depending on how you count it, consists of about 10 lines of actual code and only (minimally) exposes the lock manager. It's like some GPL code has parked itself on top of an existing server feature (and the copyright on the code in this module is dubious anyways, it being defined strictly by function). Plus, there are some important missing parts lock being able to resolve a locktag back to a pid and being able to query a lock without acquiring it. Anyways, there is a reasonable argument to be made for keeping user locks a loadable module because of the potential for misuse...it's just too easy to bring down a server with them in place. A remedy might involve some combination of the following: 1. rename max_locks_per_transaction to something more reasonable...this leads one to believe persistent locks are not managed here (but they are). 2. consider bumping out of memory condition in lock table to 'fatal' to better guard against a process in runaway lock acquirement...at least there is a chance for recovery now. 3. enforce a limit to max#locks for each process that is less than the total maximum size (for example, no single process can acquire more than half of the entire lock table). IMO, with some safety and usabilty aspects rolled in, documentation changes, and re-packaging, plus throw out the gpl stuff, this makes a reasonable (if somewhat esoteric) bullet point feature for 8.1. Merlin