Re: buffer cache and Rollback

Comments embedded.

<plbiju_at_gmail.com> wrote in message
news:1185712439.456536.54960_at_l70g2000hse.googlegroups.com...
> On Jul 27, 8:32 pm, "Richard Foote" <richard.fo..._at_bigpond.nospam.com>
> wrote:
>> <plb..._at_gmail.com> wrote in message
>>
>> news:1185578894.204112.202140_at_q75g2000hsh.googlegroups.com...
>>
>> > Question:
>>
>> > During updating any table, it updates the block in the buffercache and
>> > not the actual datafile. In a scenerio where the user executes the
>> > update and don't commit, Does the select query of other users will be
>> > retreived from the rollback segment(data file) or from the same
>> > datablock which is cached in the buffer cache.How does it work. During
>> > any updates does it copy the old image to the rollback
>> > segment(datafile), if so any update it requires disk access... please
>> > clarify...
>>
>> > Thanks
>>
>> Basically and simplistically, a query that reads data that has been
>> updated
>> but not committed by another session will read the "dirty" block,
>> determine
>> from the header and locking info that the row is in a dirty state, go to
>> the
>> specific RBS specified in the transaction slot of the updating
>> transaction,
>> find the corresponding undo (previous state of changed data), reconstruct
>> a
>> new read consistent block to the point of time prior to the update
>> occurring, see if the block is now at a point of time prior to the select
>> statement starting (if not repeat the process again and peel off another
>> layer of the onion by reconstructing read consistent images until the
>> block
>> is at a state as at the start of the transaction) and retrieve the
>> corresponding consistent data. This consistency check is always performed
>> during a read by comparing the relevant header info in a block to the
>> point
>> of a query commencing.
>>
>> A rollback/undo block is just another block, if it's in the buffer cache
>> great read it from there, if not then read it from disk. Note that the
>> first
>> rollback/undo block at least is likely to be in cache as it's from a
>> current
>> transaction but if it's a long transaction either in terms of the number
>> of
>> updated blocks or time it takes to issue the commit, then even this might
>> not be the case.
>>
>> Hope it makes sense.
>>
>> Cheers
>>
>> Richard
>
>
> Thanks Richard for your response.
> If I understand correctly, the buffer cache works like below:
> 1) User1 does a select query for row R1 which is in Block B1.
> 2) Now B1 is in Buffer cache, any further select query for row R1 will
> be served from Buffer cache.
> 3) Now User1 executes update R1.... and not commiting, so In buffer
> cache oracle will copy the old image of B1 to B1-old and will have a
> pointer in B1 to point to B1-old (No I/O Activity).

Note that B1-old doesn't exist as such. The changed data (and just the changed data plus some overhead details) are actually stored in the rollback segment associated with this transaction. The "pointer" from B1 which is part of the transaction information stored in the header of B1 points to this rollback block. It's B1 and the rollback data that together may be used to create a consistent image of B1 to a prior point of time for read consistency purposes.

> 4) Now User2 executes select query for R1, which will be served from
> B1-old(No I/O Activity).

Is served by creating a new consistent image of B1 (called B1-old if you wish) which restores the changed data stored in the rollback block and makes B1-old look as it did prior to the update statement.

> 5) If there are no space in the buffer cache than B1 will be moved to
> Datafile and B1-old will be moved to Rollback segment.
>

No. B1 will eventually be written to disk as it eventually ages. The rollback block will eventually be written to disk if it ages. B1-old will not be written to disk as it's only a consistent image and will simply be over-written once it ages.

> Could you please correct me if any of the above statement is wrong..

Did my best.

Cheers

Richard Received on Sun Jul 29 2007 - 08:10:15 CDT