This guide contains detailed information about using BookKeeper for logging. It discusses the basic operations BookKeeper supports, and how to create logs and perform basic read and write operations on these logs.
BookKeeper is a replicated service to reliably log streams of records. In BookKeeper, servers are "bookies", log streams are "ledgers", and each unit of a log (aka record) is a "ledger entry". BookKeeper is designed to be reliable; bookies, the servers that store ledgers, can crash, corrupt data, discard data, but as long as there are enough bookies behaving correctly the service as a whole behaves correctly.
The initial motivation for BookKeeper comes from the namenode of HDFS. Namenodes have to log operations in a reliable fashion so that recovery is possible in the case of crashes. We have found the applications for BookKeeper extend far beyond HDFS, however. Essentially, any application that requires an append storage can replace their implementations with BookKeeper. BookKeeper has the advantage of writing efficiently, replicating for fault tolerance, and scaling throughput with the number of servers through striping.
At a high level, a bookkeeper client receives entries from a client application and stores it to sets of bookies, and there are a few advantages in having such a service:
BookKeeper implements highly available logs, and it has been designed with write-ahead logging in mind. Besides high availability due to the replicated nature of the service, it provides high throughput due to striping. As we write entries in a subset of bookies of an ensemble and rotate writes across available quorums, we are able to increase throughput with the number of servers for both reads and writes. Scalability is a property that is possible to achieve in this case due to the use of quorums. Other replication techniques, such as state-machine replication, do not enable such a property.
An application first creates a ledger before writing to bookies through a local BookKeeper client instance. Upon creating a ledger, a BookKeeper client writes metadata about the ledger to ZooKeeper. Each ledger currently has a single writer. This writer has to execute a close ledger operation before any other client can read from it. If the writer of a ledger does not close a ledger properly because, for example, it has crashed before having the opportunity of closing the ledger, then the next client that tries to open a ledger executes a procedure to recover it. As closing a ledger consists essentially of writing the last entry written to a ledger to ZooKeeper, the recovery procedure simply finds the last entry written correctly and writes it to ZooKeeper.
Note that currently this recovery procedure is executed automatically upon trying to open a ledger and no explicit action is necessary. Although two clients may try to recover a ledger concurrently, only one will succeed, the first one that is able to create the close znode for the ledger.
BookKeeper uses four basic elements:
A set of bookies implements BookKeeper, and we use a quorum-based protocol to replicate data across the bookies. There are basically two operations to an existing ledger: read and append. Here is the complete API list (mode detail here):
There is only a single client that can write to a ledger. Once that ledger is closed or the client fails, no more entries can be added. (We take advantage of this behavior to provide our strong guarantees.) There will not be gaps in the ledger. Fingers get broken, people get roughed up or end up in prison when books are manipulated, so there is no deleting or changing of entries.
p. A simple use of BooKeeper is to implement a write-ahead transaction log. A server maintains an in-memory data structure (with periodic snapshots for example) and logs changes to that structure before it applies the change. The application server creates a ledger at startup and store the ledger id and password in a well known place (ZooKeeper maybe). When it needs to make a change, the server adds an entry with the change information to a ledger and apply the change when BookKeeper adds the entry successfully. The server can even use asyncAddEntry to queue up many changes for high change throughput. BooKeeper meticulously logs the changes in order and call the completion functions in order.
When the application server dies, a backup server will come online, get the last snapshot and then it will open the ledger of the old server and read all the entries from the time the snapshot was taken. (Since it doesn't know the last entry number it will use MAX_INTEGER). Once all the entries have been processed, it will close the ledger and start a new one for its use.
A client library takes care of communicating with bookies and managing entry numbers. An entry has the following fields:
|Ledger number||long||The id of the ledger of this entry|
|Entry number||long||The id of this entry|
|last confirmed ( LC )||long||id of the last recorded entry|
|data||byte||the entry data (supplied by application)|
|authentication code||byte||Message authentication code that includes all other fields of the entry|
The client library generates a ledger entry. None of the fields are modified by the bookies and only the first three fields are interpreted by the bookies.
To add to a ledger, the client generates the entry above using the ledger number. The entry number will be one more than the last entry generated. The LC field contains the last entry that has been successfully recorded by BookKeeper. If the client writes entries one at a time, LC is the last entry id. But, if the client is using asyncAddEntry, there may be many entries in flight. An entry is considered recorded when both of the following conditions are met:
LC seems mysterious right now, but it is too early to explain how we use it; just smile and move on.
Once all the other fields have been field in, the client generates an authentication code with all of the previous fields. The entry is then sent to a quorum of bookies to be recorded. Any failures will result in the entry being sent to a new quorum of bookies.
To read, the client library initially contacts a bookie and starts requesting entries. If an entry is missing or invalid (a bad MAC for example), the client will make a request to a different bookie. By using quorum writes, as long as enough bookies are up we are guaranteed to eventually be able to read an entry.
There are some meta data that needs to be made available to BookKeeper clients:
We maintain this information in ZooKeeper. Bookies use ephemeral nodes to indicate their availability. Clients use znodes to track ledger creation and deletion and also to know the end of the ledger and the bookies that were used to store the ledger. Bookies also watch the ledger list so that they can cleanup ledgers that get deleted.
The process of closing out the ledger and finding the last entry is difficult due to the durability guarantees of BookKeeper:
If the ledger was closed gracefully, ZooKeeper will have the last entry and everything will work well. But, if the BookKeeper client that was writing the ledger dies, there is some recovery that needs to take place.
The problematic entries are the ones at the end of the ledger. There can be entries in flight when a BookKeeper client dies. If the entry only gets to one bookie, the entry should not be readable since the entry will disappear if that bookie fails. If the entry is only on one bookie, that doesn't mean that the entry has not been recorded successfully; the other bookies that recorded the entry might have failed.
The trick to making everything work is to have a correct idea of a last entry. We do it in roughly three steps:
This section gives an overview of how a bookie manages its ledger fragments.
Bookies manage data in a log-structured way, which is implemented using three kind of files:
Since updating index files would introduce random disk I/O, for performance consideration, index files are updated lazily by a Sync Thread running in the background. Before index pages are persisted to disk, they are gathered in LedgerCache for lookup.
When a bookie receives entries from clients to be written, these entries will go through the following steps to be persisted to disk:
Ledger index pages are flushed to index files in the following two cases:
Besides flushing index pages, Sync Thread is responsible for rolling journal files in case that journal files use too much disk space.
The data flush flow in Sync Thread is as follows:
Using the above data flush mechanism, it is safe for the Sync Thread to skip data flushing when the bookie shuts down. However, in Entry Logger, it uses BufferedChannel to write entries in batches and there might be data buffered in BufferedChannel upon a shut down. The bookie needs to ensure Entry Logger flushes its buffered data during shutting down. Otherwise, Entry Log files become corrupted with partial entries.
As described above, EntryLogger#flush is invoked in the following two cases:
* in Sync Thread : used to ensure entries added before LastLogMark are persisted to disk.
* in ShutDown : used to ensure its buffered data persisted to disk to avoid data corruption with partial entries.