*** DRAFT ***

Powersafe Overwrite

"Powersafe overwrite" is a term used by the SQLite team to describe a behavior of some filesystems and disk-controllers related to data preservation during a power loss. Powersafe overwrite is a boolean property: either the storage system has it or it does not.

We say that a system has the powersafe overwrite property if the following statement is true:

When an application writes a range of bytes in a file, no bytes outside of that range will change, even if the write occurs just before a crash or power failure.

The powersafe overwrite property says nothing about the state of the bytes that were written. Those bytes might contain their old values, their new values, random values, or some combination of these. The powersafe overwrite property merely states that writes cannot change bytes outside of the range of bytes written.

In other words, powersafe overwrite means that there is no "collateral damage" when a power loss occurs while writing. Only those bytes actually being written might be damaged.

In practical terms, what the powersafe write property means is that when the disk controller detects an impending power loss, it finishes writing whatever sector it is working on prior to parking the heads. It means that individual sector writes will complete once started, even if there is a power loss.

Consider what would happen if disk sector writes are interrupted by a power loss. If an application writes two or three bytes in the middle of some file, the operating system will implement this by first reading the entire sector containing those bytes, making the change to the sector in memory, then writing the entire sector back to the disk. If a power loss occurs during the writeback and the sector was not completely written, then on the next read after reboot, error correcting codes in the sector will probably detect irreparable damage and the disk controller will read out the sector as all zeros or all ones. Thus values will have changed outside of the range of the two or three bytes that were written at the application level - a violation of the powersafe overwrite property.

SQLite Assumptions About Powersafe Overwrite

All versions of SQLite up to and including version 3.7.9 assume that the filesystem does not provide powersafe overwrite. SQLite has traditionally assumed that when any one byte of a file changes, all other bytes within the same sector of that byte have the potential of being corrupted on a power loss. When writing, SQLite has made sure to journal all bytes in the same sector of any modifications and it pads journal files out to the next sector boundary so that subsequent appends to that journal cannot damage prior records. SQLite understands the sector size to be the value returned by the xSectorSize method in the VFS. The SQLite team has often referred to the value returned by xSectorSize as the "blast radius" of a write, since it expresses the range of bytes that might be damaged if a power loss occurs during the write. The default VFSes for unix and windows have always returned 512 as the sector size (or blast radius) for all versions of SQLite up to and including version 3.7.9.

Newer disk drives have begun using 4096 byte sectors however. Beginning with SQLite version 3.7.10, the SQLite development team experimented with changes xSectorSize to report 4096 bytes as the blast radius. This had the effect of increasing write overhead on many databases. For a database with a PRAGMA page_size of 1024 (a very common choice) making a change to a single page in the database now requires SQLite to backup three other adjacent pages to the rollback journal, whereas formerly it only had to backup the one page that was changing. In WAL mode, each transaction had to be padded out to the next 4096-byte boundary in the WAL file, rather than the next 512-byte boundary, resulting in thousands of extra bytes being written per transaction.

The extra write overhead prompted a reexamination of assumptions about powersafe overwrite. With modern disk drives, the capacity has become so large and the data density so great that a single sector is very small and writing a single sector takes very little time. We know that disk drives can detect an impending power loss and continue to operate for some small amount of time on residual energy because those drives are able to park their heads before spinning down. And so if an impending power loss is detectable by the disk controller, it seems reasonable that the controller will finish writing whatever sector it is current working on when the imminent power loss is first detected, prior to parking the heads, as long as doing so does not take too long, which it should not with small and dense sectors. Hence it seems reasonable to assume powersafe overwrite for modern disks. Indeed, BerkeleyDB has made this assumption for decades, we are told. Caution is advised though. As Roger Binns noted on the SQLite developers mailing list: "'poorly written' should be the main assumption about drive firmware."

Torn Pages

A torn page occurs when a database page is larger than a disk sector, the database page is written to disk, but a power loss occurs prior to all sectors of the database page being written. Then, upon recovery, part of the database page will have the old content while some other parts of the page will have the new content. Some database engines assume that page writes are atomic and hence a torn page is an unrecoverable error.

SQLite never assumes that database page writes are atomic, regardless of the PSOW setting.(1) And hence SQLite is always able to automatically recover from torn pages induced by a crash. Enabling PSOW does not decrease SQLite's ability to recover from a torn page.

Changes In SQLite Version 3.7.10

The VFS for SQLite version 3.7.10 adds a new device characteristic named SQLITE_IOCAP_POWERSAFE_OVERWRITE. Database files that report this characteristic are assumed to reside on storage systems that have the powersafe overwrite property. The default unix and windows VFSes now report SQLITE_IOCAP_POWERSAFE_OVERWRITE if SQLite is compiled with -DSQLITE_POWERSAFE_OVERWRITE=1 or they make the legacy assumption that storage does not have the powersafe overwrite property if compiled with -DSQLITE_POWERSAFE_OVERWRITE=0. For now, the default is for powersafe overwrite to be turned on, though we may revisit this in the future and default it off.

The powersafe overwrite property for individual databases can be specified as the database is opened using the "psow" query parameter with a URI filename. For example, to always assume powersafe overwrite for a file (perhaps to ensure maximum write performance), open it as

file:somefile.db?psow=1

Or to be extra safe with a database and to force SQLite to assume the database lacks powersafe overwrite, open it using

file:somefile.db?psow=0

There is also a new SQLITE_FCNTL_POWERSAFE_OVERWRITE opcode for the sqlite3_file_control() that allows an application to query the powersafe overwrite property for a database file.


Notes:

  1. SQLite never assumes atomic page writes in its default configurations. But a custom VFS can set one of the SQLITE_IOCAP_ATOMIC bits in the result of the xDeviceCharacteristic() method and then SQLite will assume that page writes are atomic. The application must supply a custom VFS to accomplish this, however, since none of the standard VFSes will ever set any of the atomic bits in the xDeviceCharacteristics() vector.

*** DRAFT ***