Re: logical decoding - GetOldestXmin

Hi,

Robert Haas <robertmhaas(at)gmail(dot)com> schrieb:

>On Fri, Dec 14, 2012 at 7:19 PM, Andres Freund <andres(at)2ndquadrant(dot)com>
>wrote:
>> On 2012-12-14 14:01:30 -0500, Robert Haas wrote:
>>> On Fri, Dec 14, 2012 at 6:46 AM, Andres Freund
><andres(at)2ndquadrant(dot)com> wrote:
>>> > Just moving that tidbit inside the lock seems to be the pragmatic
>>> > choice. GetOldestXmin is called
>>> >
>>> > * once per checkpoint
>>> > * one per index build
>>> > * once in analyze
>>> > * twice per vacuum
>>> > * once for HS feedback messages
>>> >
>>> > Nothing of that occurs frequently enough that 5 instructions will
>make a
>>> > difference. I would be happy to go an alternative path, but right
>now I
>>> > don't see any nice one. A "already_locked" parameter to
>GetOldestXmin
>>> > seems to be a cure worse than the disease.
>>>
>>> I'm not sure that would be so bad, but I guess I question the need
>to
>>> do it this way at all. Most of the time, if you need to advertise
>>> your global xmin, you use GetSnapshotData(), not GetOldestXmin(),
>and
>>> I guess I'm not seeing why that wouldn't also work here. Am I dumb?
>>
>> I wondered upthread whether that would be better:
>>
>> On 2012-12-13 21:03:44 +0100, Andres Freund wrote:
>>> Another alternative to this would be to get a snapshot with
>>> GetSnapshotData(), copy the xmin to the logical slot, then call
>>> ProcArrayEndTransaction(). But that doesn't really seem to be nicer
>to
>>> me.
>>
>> Not sure why I considered it ugly anymore, but it actually has a
>> noticeable disadvantage. GetOldestXmin is nicer is than
>GetSnapshotData
>> as the latter set a fairly new xid as xmin whereas GetOldestXmin
>returns
>> the actual current xmin horizon. Thats preferrable because it allows
>us
>> to start up more quickly. snapbuild.c can only start building a
>snapshot
>> once it has seen a xl_running_xact with oldestRunningXid >=
>> own_xmin. Otherwise we cannot be sure that no relevant catalog tuples
>> have been removed.
>
>I'm a bit confused. Are you talking about the difference between
>RecentGlobalXmin and RecentXmin? I think GetSnapshotData() updates
>both.

The problem is that at the time GetSnapshotData returns the xmin horizon might have gone upwards and tuples required for decoding might get removed by other backends. That needs to be prevented while holding the procarray lock exclusively.

Does it make more sense now?

Andres

---
Please excuse the brevity and formatting - I am writing this on my mobile phone.