This page is to talk about Fscking and autorecovery of ledgers and bookies, so that we can discuss and get a clear story of what needs to be done, from which we can then derive a list of JIRAs.https://cwiki.apache.org/confluence/pages/editpage.action?pageId=27844384^{Image Removed}

State of the plan (as of

...

31 Aug 2012)

BookKeeper Auto-Recovery

When any Bookie goes down in the BookKeeper cluster, there is no way to recover
the lost data from that Bookie server. For example, if we have 2 replicas for a
ledger in BK cluster, and a node goes down from it, we will be running the cluster
with single replica. Running clusters with single or no replicas will be a risk, as
nodes may fail in general.
To avoid such situations, we need a mechanism for recovering the data to new
bookies for meeting the enough replica criteria (quorum size) and it is called as
Auto-Recovery in BookKeeper.

Working

Auto-Recovery has two main modules:

•  Auditor
•  ReplicationWorker

AuditorPeer is an Auto-recovery node, which internally initializes and starts Auditor
and ReplicationWorker threads. So, each Auto-recovery node will have two threads
running.

This Auto-recovery node has to be started in each Bookie machine. All recovery
nodes will participate in leader election and one Auditor may become as leader and
others will just watch the elected auditor failure to participate again in next election.

Auditor:

Once Auditor thread is started, auditor elector will go for the election to win the
auditing job for Bookie cluster. Here, auditing job would be that, it has to detect the
under-replicated ledgers in the cluster due to Bookie failures.

Image Added

Auditor will keep watch on the available Bookies in the cluster. Bookie will add its
entry in the available bookie during Bookie server startup. So, when the Bookie is
crashed or killed, Auditor will get a notification about the children change in
available Bookie list. Auditor will immediately scan the complete ledger list related
to that failed Bookie.
On getting the details of ledgers, Auditor node publishes the under-replicated
ledgers in under-replication znode path in Zookeeper. After this, again it will be
watching Bookie failures (by resetting Zookeeper watcher) on available Bookie list.

ReplicationWorker:

ReplicationWorker will get the under replicated ledgers one after the other for rereplication.
If there are no ledgers in under-replicated state, worker will wait for the
ledgers to be added into under-replicated list in ZK by Auditor.

Image Added

Once the auditor publishes about the under replicated ledgers in ZK,
ReplicationWorker picks one ledger and get a lock on it. Here it will get the lock by
just adding an ephemeral znode with ledger name in underreplication/locks folder.
Then replication worker scans the ledger and gets the under replicated fragments.
If the ledger fragment already contains the local Bookie address in its ensemble, it
will skip the replication for that ledger as ReplicationWorker will treat local Bookies
as a target bookie to copy or replicate ledger fragment to it.
If replication is not completed for all the under-replicated fragments in the ledger,
ReplicationWorker will just release the lock for the ledger. Technically, releasing
lock will be deleting the held lock in /underreplication/locks folder. So, that other
replication worker can immediately get notified about the lock deletion and may
pick this ledger for replicating the pending fragments from the ledger. This process
works well for the closed ledgers.
If the ledger is in the open state and if the last fragment of the ledger in under
replication state, there will be a risk of data loss if we replicate the last fragment
straight away like above process.

How ReplicationWorker handled this data loss scenario?
Scenario: The last fragment of the ledger is in under replicated state; replication
worker replicates it and updates the ledger metadata with local Bookies address.
Immediately, the failed Bookie started and running. Now the client resumed for
adding some more entries, and it can continue with adding entries with the old
Bookie. But ReplicationWorker already change the metadata for that fragment with
local Bookie. That means, that client unnecessarily adding the entries to the old
bookie whose address is already removed from fragment ensemble. So, this can
create data loss if other bookie goes down and even though old Bookie is running
fine.
To prevent this situation, ReplicationWorker will postpone the replications by adding
that ledger to the pending replications. Pending ReplicationWorker will check the
timeout of the ledgers which are there in pending replications. This timeout is
configurable. Once the timeout happens, pending ReplicationWorker will force fence
the ledger if it is still in open state and will inform the replication worker for the
replication.
So, any under-replicated last fragment ledger will not be kept open for long time if
the client is idle and not reforming ensemble for long (more than pending
replication timeout.)

We have two new entities running on bookies. The auditor and the recovery worker.

The Auditor

Within the system there is only one auditor. Each bookie runs an auditor thread and they use zookeeper to elect which one gets to be auditor. If the auditor fails, then the election is run again. As per the latest discussion Auditor may be started as separate process instead of running it along with the bookies.

The roll of the auditor is to watch for bookie failure, and when a failure does occur, mark all ledgers with fragments on that bookie to be rereplicated.

There is a znode, /ledgers/underreplicated under which the ledgers to be rereplicated are placed.

The recovery worker

Each bookie in the cluster runs a recovery worker. The recovery worker watches /ledgers/underreplicated for new ledgers to appear. When a ledger does appear, the recovery worker will lock it, and run rereplication on it. If the recovery worker fails to acquire the lock, it tries the next ledger.
On successful rereplication, recovery worker deletes the ledger from /ledgers/underreplicated and also releases the lock.

Open Questions

• We should also periodically check ledgers are available. Where should this run from?

...