xtrabackup implementation details

This page contains notes on various internal aspects of the xtrabackup tool’s operation.

File permissions

xtrabackup opens the source data files in read-write mode, although it does not modify the files. This means that you must run xtrabackup as a user who has permission to write the data files. The reason for opening the files in read-write mode is that xtrabackup uses the embedded InnoDB libraries to open and read the files, and InnoDB opens them in read-write mode because it normally assumes it is going to write to them.

Tune the OS buffers

Because xtrabackup reads large amounts of data from the filesystem, it uses posix_fadvise() where possible, to instruct the operating system not to try to cache the blocks it reads from disk. Without this hint, the operating system would prefer to cache the blocks, assuming that xtrabackup is likely to need them again, which is not the case. Caching such large files can place pressure on the operating system’s virtual memory and cause other processes, such as the database server, to be swapped out. The xtrabackup tool avoids this with the following hint on both the source and destination files:

posix_fadvise(file, 0, 0, POSIX_FADV_DONTNEED)

In addition, xtrabackup asks the operating system to perform more aggressive read-ahead optimizations on the source files:

posix_fadvise(file, 0, 0, POSIX_FADV_SEQUENTIAL)

Copy data files

When copying the data files to the target directory, xtrabackup reads and writes 1 MB of data at a time. This is not configurable. When copying the log file, xtrabackup reads and writes 512 bytes at a time. This is also not possible to configure, and matches InnoDB’s behavior (workaround exists in Percona Server for MySQL because it has an option to tune innodb_log_block_size for XtraDB, and in that case Percona XtraBackup will match the tuning).

After reading from the files, xtrabackup iterates over the 1MB buffer a page at a time, and checks for page corruption on each page with InnoDB’s buf_page_is_corrupted() function. If the page is corrupt, it re-reads and retries up to 10 times for each page. It skips this check on the doublewrite buffer.