Berkeley DB recoverability

Berkeley DB recovery is based on write-ahead logging. This means that when a change is made to a database page, a description of the change is written into a log file. This description in the log file is guaranteed to be written to stable storage before the database pages that were changed are written to stable storage. This is the fundamental feature of the logging system that makes durability and rollback work.

If the application or system crashes, the log is reviewed during recovery. Any database changes described in the log that were part of committed transactions and that were never written to the actual database itself are written to the database as part of recovery. Any database changes described in the log that were never committed and that were written to the actual database itself are backed-out of the database as part of recovery. This design allows the database to be written lazily, and only blocks from the log file have to be forced to disk as part of transaction commit.

There are two interfaces that are a concern when considering Berkeley DB recoverability:

  1. The interface between Berkeley DB and the operating system/filesystem.
  2. The interface between the operating system/filesystem and the underlying stable storage hardware.

Berkeley DB uses the operating system interfaces and its underlying filesystem when writing its files. This means that Berkeley DB can fail if the underlying filesystem fails in some unrecoverable way. Otherwise, the interface requirements here are simple: The system call that Berkeley DB uses to flush data to disk (normally fsync or fdatasync), must guarantee that all the information necessary for a file's recoverability has been written to stable storage before it returns to Berkeley DB, and that no possible application or system crash can cause that file to be unrecoverable.

In addition, Berkeley DB implicitly uses the interface between the operating system and the underlying hardware. The interface requirements here are not as simple.

First, it is necessary to consider the underlying page size of the Berkeley DB databases. The Berkeley DB library performs all database writes using the page size specified by the application, and Berkeley DB assumes pages are written atomically. This means that if the operating system performs filesystem I/O in blocks of different sizes than the database page size, it may increase the possibility for database corruption. For example, assume that Berkeley DB is writing 32KB pages for a database, and the operating system does filesystem I/O in 16KB blocks. If the operating system writes the first 16KB of the database page successfully, but crashes before being able to write the second 16KB of the database, the database has been corrupted and this corruption may or may not be detected during recovery. For this reason, it may be important to select database page sizes that will be written as single block transfers by the underlying operating system. If you do not select a page size that the underlying operating system will write as a single block, you may want to configure the database to use checksums (see the DB->set_flags() flag for more information). By configuring checksums, you guarantee this kind of corruption will be detected at the expense of the CPU required to generate the checksums. When such an error is detected, the only course of recovery is to perform catastrophic recovery to restore the database.

Second, if you are copying database files (either as part of doing a hot backup or creation of a hot failover area), there is an additional question related to the page size of the Berkeley DB databases. You must copy databases atomically, in units of the database page size. In other words, the reads made by the copy program must not be interleaved with writes by other threads of control, and the copy program must read the databases in multiples of the underlying database page size. Generally, this is not a problem, as operating systems already make this guarantee and system utilities normally read in power-of-2 sized chunks, which are larger than the largest possible Berkeley DB database page size.

One problem we have seen in this area was in some releases of Solaris where the cp utility was implemented using the mmap system call rather than the read system call. Because the Solaris' mmap system call did not make the same guarantee of read atomicity as the read system call, using the cp utility could create corrupted copies of the databases. Another problem we have seen is implementations of the tar utility doing 10KB block reads by default, and even when an output block size was specified to that utility, not reading from the underlying databases in multiples of the block size. Using the dd utility instead of the cp or tar utilities (and specifying an appropriate block size), fixes these problems. If you plan to use a system utility to copy database files, you may want to use a system call trace utility (for example, ktrace or truss) to check for an I/O size smaller than or not a multiple of the database page size and system calls other than read.

Third, it is necessary to consider the behavior of the system's underlying stable storage hardware. For example, consider a SCSI controller that has been configured to cache data and return to the operating system that the data has been written to stable storage, when, in fact, it has only been written into the controller RAM cache. If power is lost before the controller is able to flush its cache to disk, and the controller cache is not stable (that is, the writes will not be flushed to disk when power returns), the writes will be lost. If the writes include database blocks, there is no loss because recovery will correctly update the database. If the writes include log file blocks, it is possible that transactions that were already committed may not appear in the recovered database, although the recovered database will be coherent after a crash.

If the underlying hardware can fail in any way so that only part of the block was written, the failure conditions are the same as those described previously for an operating system failure that writes only part of a logical database block. In such cases, configuring the database for checksums will ensure the corruption is detected.

For these reasons, it may be important to select hardware that does not do partial writes and does not cache data writes (or does not return that the data has been written to stable storage until it has either been written to stable storage or the actual writing of all of the data is guaranteed, barring catastrophic hardware failure -- that is, your disk drive exploding).

If the disk drive on which you are storing your databases explodes, you can perform normal Berkeley DB catastrophic recovery, because it requires only a snapshot of your databases plus the log files you have archived since those snapshots were taken. In this case, you should lose no database changes at all.

If the disk drive on which you are storing your log files explodes, you can also perform catastrophic recovery, but you will lose any database changes made as part of transactions committed since your last archival of the log files. Alternatively, if your database environment and databases are still available after you lose the log file disk, you should be able to dump your databases. However, you may see an inconsistent snapshot of your data after doing the dump, because changes that were part of transactions that were not yet committed may appear in the database dump. Depending on the value of the data, a reasonable alternative may be to perform both the database dump and the catastrophic recovery and then compare the databases created by the two methods.

Regardless, for these reasons, storing your databases and log files on different disks should be considered a safety measure as well as a performance enhancement.

Finally, you should be aware that Berkeley DB does not protect against all cases of stable storage hardware failure, nor does it protect against simple hardware misbehavior (for example, a disk controller writing incorrect data to the disk). However, configuring the database for checksums will ensure that any such corruption is detected.