Rados - metadata-only journal mode¶
Currently the Ceph community is thinking of eliminating the double write
penalty of write ahead logging, newstore is a great design which implements
create, append operations in an copy on read way, while maintaining all
the original semantics. This makes newstore a general purpose optimization,
especially suitable for the write once scenarios. Metadata-only journal mode
intends to do in a more aggressive way, that is, not journal object data at all.
This applies to two major kinds of situations, one is that the atomicity for
object data modification may not need, for example, RBD to simulate a disk
in cloud platform. The second is those double journaling situations, for example,
cache tiering, while cache pool has already provided the durability, when dirty
objects are written back, they theoretically need not go through the journaling
process of base pool, since the flusher could always replay the write operation.
Metadata-only journal mode, to some extent, resembles the data=ordered journal
mode in ext4. With such journal mode is on, object data are written directly to
their ultimate location, when data written finished, metadata are written into the
journal. It guarantees the consistency in terms of RADOS name space, and the data
consistency among object copies. However, the object data may not be correct.
Later we will demonstrate that this rarely happens.
If you are interested in contributing to this blueprint, or want to be a "speaker" during the Summit session, list your name here.
Please describe the current status of Ceph as it relates to this blueprint. Is there something that this replaces? Are there current features that are related?
We have two options,
The first option
1 Write object data
2 Commit metadata to journal
3 Submit metadata
This option makes minimal revision to current framework, enjoy the simplicity. In most cases, this will not introduce problem relying on the powerful peering and client resent mechanism. The only problematic situation is PG down as a whole, and client also down, in that case, the guest fs in the vm will possibly recover it to consistent by fsck and journal replaying. So it just to leave scrub to find and fix this by randomly synchronizing one of the copies to others.
The second option
1 Commit transaction A into journal, to add a UNSTABLE flag in object metadata, add a pglog record, and update the pglog version
2 Write data to object
3 Commit transaction B into journal, to update metadata, delete the flag, add a pglog record, and update the pglog version
4 Submit metadata
As long as one osd in the pg has succeeded, the pg will be recovered to
a consistent and correct state by peering; If the pg down as a whole,
there are the following situations,
(1) None of the osds finishes step 1, nothing happen;
(2) At least one of the osds finishes step 3, journaling and
peering will recover the pg to a consistent and correct state;
(3) none of the osds has finished step (3), and at least one of the osds has
finished step (1), this is the only potentially problematical situation,
in this case, peering will synchronize the UNSTABLE flag to other osds in the pg.
For object read and write, if found UNSTABLE flag in object metadata, it will let the operation wait and
start a recovery to randomly choose one osd to synchronize its
content of written area to other copies. During scrub, it will also check the flag and do the recvery.
This section should contain a list of work tasks created by this blueprint. Please include engineering tasks as well as related build/release and documentation work. If this blueprint requires cleanup of deprecated features, please list those tasks as well.
Build / release tasks