Re: [REVIEW] Re: Compression of full-page-writes

On Tue, Mar 3, 2015 at 9:34 AM, Michael Paquier
<michael(dot)paquier(at)gmail(dot)com> wrote:
> On Tue, Mar 3, 2015 at 9:24 AM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
>> On 2015-03-03 08:59:30 +0900, Michael Paquier wrote:
>>> Already mentioned upthread, but I agree with Fujii-san here: adding
>>> information related to the state of a block image in
>>> XLogRecordBlockHeader makes little sense because we are not sure to
>>> have a block image, perhaps there is only data associated to it, and
>>> that we should control that exclusively in XLogRecordBlockImageHeader
>>> and let the block ID alone for now.
>>
>> This argument doesn't make much sense to me. The flag byte could very
>> well indicate 'block reference without image following' vs 'block
>> reference with data + hole following' vs 'block reference with
>> compressed data following'.
>
> Information about the state of a block is decoupled with its
> existence, aka in the block header, we should control if:
> - record has data
> - record has a block
> And in the block image header, we control if the block is:
> - compressed or not
> - has a hole or not.

Are there any other flag bits that we should or are planning to add into
WAL header newly, except the above two? If yes and they are required by even
a block which doesn't have an image, I will change my mind and agree to
add something like chunk ID to a block header. But I guess the answer of the
question is No. Since the flag bits now we are thinking to add are required
only by a block having an image, adding them into a block header (instead of
block image header) seems a waste of bytes in WAL. So I concur with Michael.

Regards,

--
Fujii Masao