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

On Mon, Feb 23, 2015 at 5:28 PM, Rahila Syed <rahilasyed90(at)gmail(dot)com> wrote:
> Hello,
>
> Attached is a patch which has following changes,
>
> As suggested above block ID in xlog structs has been replaced by chunk ID.
> Chunk ID is used to distinguish between different types of xlog record
> fragments.
> Like,
> XLR_CHUNK_ID_DATA_SHORT
> XLR_CHUNK_ID_DATA_LONG
> XLR_CHUNK_BKP_COMPRESSED
> XLR_CHUNK_BKP_WITH_HOLE
>
> In block references, block ID follows the chunk ID. Here block ID retains
> its functionality.
> This approach increases data by 1 byte for each block reference in an xlog
> record. This approach separates ID referring different fragments of xlog
> record from the actual block ID which is used to refer block references in
> xlog record.

I've not read this logic yet, but ISTM there is a bug in that new WAL format
because I got the following error and the startup process could not replay
any WAL records when I set up replication and enabled wal_compression.

LOG: record with invalid length at 0/30000B0
LOG: record with invalid length at 0/3000518
LOG: Invalid block length in record 0/30005A0
LOG: Invalid block length in record 0/3000D60
...

Regards,

--
Fujii Masao