Rados - metadata-only journal mode » History » Revision 15
Revision 14 (Li Wang, 07/01/2015 10:22 AM) → Revision 15/16 (Li Wang, 07/02/2015 03:17 AM)
h1. Rados - metadata-only journal mode *Summary* 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. *Owners* Li Wang (liwang@ubuntukylin.com) Yunchuan Wen (yunchuanwen@ubuntukylin.com) Name *Interested Parties* If you are interested in contributing to this blueprint, or want to be a "speaker" during the Summit session, list your name here. Name (Affiliation) Name (Affiliation) Name *Current Status* 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? *Detailed Description* We have two options, The first option 1 Write The only revision lies in that it does not journal object data, when the transaction is committing, the data 2 Commit metadata to journal 3 Submit metadata This option makes minimal revision to current framework, enjoy the simplicity. are written into object directly. 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 copy to others. The second option 1 Commit Submit transaction A into journal, to add a UNSTABLE flag record for <offset, length> non-journaling data write in object metadata, and add a pglog record, update the pglog version omap 2 Write data to object 3 Commit Submit transaction B into journal, to update the metadata, delete the flag, and add a pglog record, update revert the pglog version 4 Submit metadata operations of transaction A 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 omap record to other osds in the pg. For object read and write, if found UNSTABLE flag in object metadata, the record, 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 record and do the recvery. *Work items* 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. *Coding tasks* Task 1 Task 2 Task 3 *Build / release tasks* Task 1 Task 2 Task 3 *Documentation tasks* Task 1 Task 2 Task 3 *Deprecation tasks* Task 1 Task 2 Task 3