Visibility map page pinned for too long ? - Mailing list pgsql-hackers

From Pavan Deolasee
Subject Visibility map page pinned for too long ?
Date
Msg-id CABOikdOqcZW5ra-uE41HUFn2_tENC8_muY5gLBWVFg1+XJf9UA@mail.gmail.com
Whole thread Raw
Responses Re: Visibility map page pinned for too long ?  (Simon Riggs <simon@2ndQuadrant.com>)
List pgsql-hackers
I was looking at the code when the following tiny bit caught my attention. In vacuumlazy.c, we release the pin on the final VM page at line number 972.

 954     if (vacrelstats->num_dead_tuples > 0)
 955     {
 956         /* Log cleanup info before we touch indexes */
 957         vacuum_log_cleanup_info(onerel, vacrelstats);
 958 
 959         /* Remove index entries */
 960         for (i = 0; i < nindexes; i++)
 961             lazy_vacuum_index(Irel[i],
 962                               &indstats[i],
 963                               vacrelstats);
 964         /* Remove tuples from heap */
 965         lazy_vacuum_heap(onerel, vacrelstats);
 966         vacrelstats->num_index_scans++;
 967     }
 968 
 969     /* Release the pin on the visibility map page */
 970     if (BufferIsValid(vmbuffer))
 971     {
 972         ReleaseBuffer(vmbuffer);
 973         vmbuffer = InvalidBuffer;
 974     }

So we are holding the pin right through the index vacuuming and the second pass over the heap; both can take a very long time. We can and should really be releasing the pin *before* those steps. In fact, it would be appropriate to do it right after the preceding big for-loop.  

While it may or may not matter from the performance or correctness perspective, I think we should fix that.

Thanks,
Pavan

--
Pavan Deolasee
http://www.linkedin.com/in/pavandeolasee

pgsql-hackers by date:

Previous
From: Magnus Hagander
Date:
Subject: Re: Tablespaces in the data directory
Next
From: Dimitri Fontaine
Date:
Subject: Re: Review: create extension default_full_version