On 09/02/2015 08:22 PM, Venky Shankar wrote:
On Wed, Sep 2, 2015 at 12:42 PM, Aravinda <avishwan@xxxxxxxxxx> wrote:
Geo-replication and Sharding Team today discussed about the approach
to make Sharding aware Geo-replication. Details are as below
Participants: Aravinda, Kotresh, Krutika, Rahul Hinduja, Vijay Bellur
- Both Master and Slave Volumes should be Sharded Volumes with same
configurations.
- In Changelog record changes related to Sharded files also. Just like
any regular files.
- Sharding should allow Geo-rep to list/read/write Sharding internal
Xattrs if Client PID is gsyncd(-1)
- Sharding should allow read/write of Sharded files(that is in .shards
directory) if Client PID is GSYNCD
- Sharding should return actual file instead of returning the
aggregated content when the Main file is requested(Client PID
GSYNCD)
For example, a file f1 is created with GFID G1.
When the file grows it gets sharded into chunks(say 5 chunks).
f1 G1
.shards/G1.1 G2
.shards/G1.2 G3
.shards/G1.3 G4
.shards/G1.4 G5
In Changelog, this is recorded as 5 different files as below
CREATE G1 f1
DATA G1
META G1
CREATE G2 PGS/G1.1
DATA G2
META G1
CREATE G3 PGS/G1.2
DATA G3
META G1
CREATE G4 PGS/G1.3
DATA G4
META G1
CREATE G5 PGS/G1.4
DATA G5
META G1
Where PGS is GFID of .shards directory.
Geo-rep will create these files independently in Slave Volume and
syncs Xattrs of G1. Data can be read only when all the chunks are
synced to Slave Volume. Data can be read partially if main/first file
and some of the chunks synced to Slave.
So, before replicating data to the salve, all shards needs to be created there?
No. each files will be synced independently. But for reading complete
file all the shards should be present, else partial data is read.
Please add if I missed anything. C & S Welcome.
regards
Aravinda
On 08/11/2015 04:36 PM, Aravinda wrote:
Hi,
We are thinking different approaches to add support in Geo-replication for
Sharded Gluster Volumes[1]
Approach 1: Geo-rep: Sync Full file
- In Changelog only record main file details in the same brick where it
is created
- Record as DATA in Changelog whenever any addition/changes to the
sharded file
- Geo-rep rsync will do checksum as a full file from mount and syncs as
new file
- Slave side sharding is managed by Slave Volume
Approach 2: Geo-rep: Sync sharded file separately
- Geo-rep rsync will do checksum for sharded files only
- Geo-rep syncs each sharded files independently as new files
- [UNKNOWN] Sync internal xattrs(file size and block count) in the main
sharded file to Slave Volume to maintain the same state as in Master.
- Sharding translator to allow file creation under .shards dir for
gsyncd. that is Parent GFID is .shards directory
- If sharded files are modified during Geo-rep run may end up stale data
in Slave.
- Files on Slave Volume may not be readable unless all sharded files sync
to Slave(Each bricks in Master independently sync files to slave)
First approach looks more clean, but we have to analize the Rsync checksum
performance on big files(Sharded in backend, accessed as one big file from
rsync)
Let us know your thoughts. Thanks
Ref:
[1]
http://www.gluster.org/community/documentation/index.php/Features/sharding-xlator
--
regards
Aravinda
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
regards
Aravinda
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
