Table of contents
This document is intended to provide general hints, tips and advice to administrators of Xapian systems. It assumes that you have installed Xapian on your system, and are familiar with the basics of creating and searching Xapian databases.
The intended audience is system administrators who need to be able to perform general management of a Xapian database, including tasks such as taking backups and optimising performance. It may also be useful introductory reading for Xapian application developers.
The document is up-to-date for Xapian version 1.4.21.
Xapian databases hold all the information needed to perform searches in a set of tables. The default database backend for the 1.4 release series is called glass. The default backend for the 1.2 release series was called chert, and this is also supported by 1.4.
The following table always exists:
- The postlist table holds a list of all the documents indexed by each term in the database (postings), and also chunked streams of the values in each value slot.
The following table exists by default, but you can choose not to have it:
- The termlist table holds a list of all the terms which index each document, and also the value slots used in each document. Without this, some features aren't supported - see Xapian::DB_NO_TERMLIST for details.
And the following optional tables exist only when there is data to store in them:
- The docdata table holds the document data associated with each document in the database.
- The position table stores all the word positions in each document which each term occurs at.
- The spelling table holds data for suggesting spelling corrections.
- The synonym table holds a synonym dictionary.
Each of the tables is held in a separate file with extension .glass (e.g. postlist.glass), allowing an administrator to see how much data is being used for each of the above purposes.
The .glass file actually stores the data, and is structured as a tree of blocks, which have a default size of 8KB (though this can be set, either through the Xapian API, or with the xapian-compact tool described below).
Changing the blocksize may have performance implications, but it is hard to know whether these will be positive or negative for a particular combination of hardware and software without doing some profiling.
The .baseA and .baseB files you may remember if you've worked with older Xapian database backends no longer exist in glass databases - the information about unused blocks is stored in a freelist (itself stored in unused blocks in the .glass file, and the other information is stored in the iamglass file.
Glass also supports databases stored in a single file - currently these only support read operations, and have to be created by compacting an existing glass database. It won't save you diskspace, but it means only one file needs to be opened to open the database so reduces initialisation overhead a little, and a single file is more convenient if you need to copy it around. You can even embed the database in another file so you can ship a single file containing content and a Xapian database which provides a search of it.
The following tables always exist:
- The postlist holds a list of all the documents indexed by each term in the database, and also chunked streams of the values in each value slot.
- The record holds the document data associated with each document in the database.
- The termlist holds a list of all the terms which index each document, and also the value slots used in each document.
And the following optional tables exist only when there is data to store in them:
- The position holds a list of all the word positions in each document which each term occurs at.
- The spelling holds data for suggesting spelling corrections.
- The synonym holds a synonym dictionary.
Each of the tables is held in a separate file, allowing an administrator to see how much data is being used for each of the above purposes. It is not always necessary to fully populate these tables: for example, if phrase searches are never going to be performed on the database, it is not necessary to store any positionlist information.
If you look at a Xapian database, you will see that each of these tables actually uses 2 or 3 files. For example, for a "chert" format database the termlist table is stored in the files termlist.baseA, termlist.baseB and termlist.DB.
The .DB file actually stores the data, and is structured as a tree of blocks, which have a default size of 8KB (though this can be set, either through the Xapian API, or with some of the tools detailed later in this document).
The .baseA and .baseB files are used to keep track of where to start looking for data in the .DB file (the root of the tree), and which blocks are in use. Often only one of the .baseA and .baseB files will be present; each of these files refers to a revision of the database, and there may be more than one valid revision of the database stored in the .DB file at once.
Changing the blocksize may have performance implications, but it is hard to tell whether these will be positive or negative for a particular combination of hardware and software without doing some profiling.
Xapian ensures that all modifications to its database are performed atomically. This means that:
- From the point of view of a separate process (or a separate database object in the same process) reading the database, all modifications made to a database are invisible until the modifications is committed.
- The database on disk is always in a consistent state.
- If the system is interrupted during a modification, the database should always be left in a valid state. This applies even if the power is cut unexpectedly, as long as the disk does not become corrupted due to hardware failure.
Committing a modification requires several calls to the operating system to make it flush any cached modifications to the database to disk. This is to ensure that if the system fails at any point, the database is left in a consistent state. Of course, this is a fairly slow process (since the system has to wait for the disk to physically write the data), so grouping many changes together will speed up the throughput considerably.
Many modifications can be explicitly grouped into a single transaction, so that lots of changes are applied at once. Even if an application doesn't explicitly protect modifications to the database using transactions, Xapian will group modifications into transactions, applying the modifications in batches.
Note that it is not currently possible to extend Xapian's transactions to cover multiple databases, or to link them with transactions in external systems, such as an RDBMS.
Finally, note that it is possible to compile Xapian such that it doesn't make modifications in an atomic manner, in order to build very large databases more quickly (search the Xapian mailing list archives for "DANGEROUS" mode for more details). This isn't yet integrated into standard builds of Xapian, but may be in future, if appropriate protections can be incorporated.
Xapian implements a "single writer, multiple reader" model. This means that, at any given instant, there is only permitted to be a single object modifying a database, but there may (simultaneously) be many objects reading the database at once.
Xapian enforces this restriction using by having a writer lock the database. Each Xapian database directory contains a lock file named flintlock (we've kept the same name as flint used, since the locking technique is the same).
This lock-file will always exist, but will be locked using fcntl() when the database is open for writing. A major advantage of fnctl() locks is that if a writer exits without being given a chance to clean up (for example, if the application holding the writer is killed), any fcntl() locks held will be automatically released by the operating system so stale locks can't happen.
Unfortunately, fcntl() locking has some unhelpful semantics (if a process closes ANY open file descriptor on the file that releases the lock) so on most POSIX platforms we spawn a child process to hold the lock for each database opened for writing, which then exec-s cat, so you will see a cat subprocess of any writer process in the output of ps, top, etc.
"Open File Description" locks are like traditional fcntl() locks but with this problem addressed, and Xapian will use these if available and avoid these extra child processes. At the time of writing it seems only Linux (since kernel 3.15) actually supports these, but they've been accepted for POSIX issue 8.
Under Microsoft Windows, we use a different locking technique which doesn't require a child process, but still means the lock is released automatically when the writing process exits.
Xapian databases contain a revision number. This is essentially a count of the number of modifications since the database was created, and is needed to implement the atomic modification functionality. It is stored as a 32 bit integer, so there is a chance that a very frequently updated database could cause this to overflow. The consequence of such an overflow would be to throw an exception reporting that the database has run out of revision numbers.
This isn't likely to be a practical problem, since it would take nearly a year for a database to reach this limit if 100 modifications were committed every second, and no normal Xapian system will commit more than once every few seconds. However, if you are concerned, you can use the xapian-compact tool to make a fresh copy of the database with the revision number set to 1.
The revision number of each table can be displayed by the xapian-check tool.
Xapian should work correctly over a network file system. However, there are various potential issues with such file systems, so we recommend extensive testing of your particular network file system before deployment.
Be warned that Xapian is heavily I/O dependent, and therefore performance over a network file system is likely to be slow unless you've got a very well tuned setup.
Xapian needs to be able to lock a file in a database directory when modifications are being performed. On some network files systems (e.g., NFS) this requires a lock daemon to be running.
As of release 1.4.0, you should generally use the glass format (which is now the default).
Support for the pre-1.0 quartz format (deprecated in 1.0) was removed in 1.1.0. See below for how to convert a quartz database to a flint one.
The flint backend (the default for 1.0, and still supported by 1.2.x) was removed in 1.3.0. See below for how to convert a flint database to a chert one.
The chert backend (the default for 1.2) is still supported by 1.4.x, but deprecated - only use it if you already have databases in this format; and plan to migrate away.
If you wish to store meta-data or other information relating to the Xapian database, it is reasonable to wish to put this in files inside the Xapian database directory, for neatness. For example, you might wish to store a list of the prefixes you've applied to terms for specific fields in the database.
Current Xapian backends don't do anything which will break this technique, so as long as you don't choose a filename that Xapian uses itself, there should be no problems. However, be aware that new versions of Xapian may use new files in the database directory, and it is also possible that new backend formats may not be compatible with the technique. And of course you can't do this with a single-file glass database.
- The simplest way to perform a backup is to temporarily halt modifications, take a copy of all files in the database directory, and then allow modifications to resume. Read access can continue while a backup is being taken.
- If you have a filesystem which allows atomic snapshots to be taken of directories (such as an LVM filesystem), an alternative strategy is to take a snapshot and simply copy all the files in the database directory to the backup medium. Such a copy will always be a valid database.
- Progressive backups are not easily possible; modifications are typically spread throughout the database files.
Even though Xapian databases are often automatically generated from source data which is stored in a reliable manner, it is usually desirable to keep backups of Xapian databases being run in production environments. This is particularly important in systems with high-availability requirements, since re-building a Xapian database from scratch can take many hours. It is also important in the case where the data stored in the database cannot easily be recovered from external sources.
Xapian databases are managed such that at any instant in time, there is at least one valid revision of the database written to disk (and if there are multiple valid revisions, Xapian will always open the most recent). Therefore, if it is possible to take an instantaneous snapshot of all the database files (for example, on an LVM filesystem), this snapshot is suitable for copying to a backup medium. Note that it is not sufficient to take a snapshot of each database file in turn - the snapshot must be across all database files simultaneously. Otherwise, there is a risk that the snapshot could contain database files from different revisions.
If it is not possible to take an instantaneous snapshot, the best backup strategy is simply to ensure that no modifications are committed during the backup procedure. While the simplest way to implement this may be to stop whatever processes are used to modify the database, and ensure that they close the database, it is not actually necessary to ensure that no writers are open on the database; it is enough to ensure that no writer makes any modification to the database.
Because a Xapian database can contain more than one valid revision of the database, it is actually possible to allow a limited number of modifications to be performed while a backup copy is being made, but this is tricky and we do not recommend relying on it. Future versions of Xapian are likely to support this better, by allowing the current revision of a database to be preserved while modifications continue.
Progressive backups are not recommended for Xapian databases: Xapian database files are block-structured, and modifications are spread throughout the /database file. Therefore, a progressive backup tool will not be able to take a backup by storing only the new parts of the database. Modifications will normally be so extensive that most parts of the database have been modified, however, if only a small number of modifications have been made, a binary diff algorithm might make a usable progressive backup tool.
When designing an indexing strategy, it is often useful to be able to check the contents of the database. Xapian includes a simple command-line program, xapian-delve, to allow this (prior to 1.3.0, xapian-delve was usually called delve, though some packages were already renaming it).
For example, to display the list of terms in document "1" of the database "foo", use:
xapian-delve foo -r 1
It is also possible to perform simple searches of a database. Xapian includes another simple command-line program, quest, to support this. quest is only able to search for un-prefixed terms, the query string must be quoted to protect it from the shell. To search the database "foo" for the phrase "hello world", use:
quest -d foo '"hello world"'
If you have installed the "Omega" CGI application built on Xapian, this can also be used with the built-in "godmode" template to provide a web-based interface for browsing a database. See Omega's documentation for more details on this.
Xapian databases normally have some spare space in each block to allow new information to be efficiently slotted into the database. However, the smaller a database is, the faster it can be searched, so if there aren't expected to be many further modifications, it can be desirable to compact the database.
Xapian includes a tool called xapian-compact for compacting databases. This tool makes a copy of a database, and takes advantage of the sorted nature of the source Xapian database to write the database out without leaving spare space for future modifications. This can result in a large space saving.
The downside of compaction is that future modifications may take a little longer, due to needing to reorganise the database to make space for them. However, modifications are still possible, and if many modifications are made, the database will gradually develop spare space.
There's an option (-F) to perform a "fuller" compaction. This option compacts the database as much as possible, but it violates the design of the Btree format slightly to achieve this, so it is not recommended if further modifications are at all likely in future. If you do need to modify a "fuller" compacted database, we recommend you run xapian-compact on it without -F first.
You can specify the blocksize to use for the compacted database (which should be a power of 2 between 2KB and 64KB, with the default being 8KB).
Making the blocksize a multiple of (or the same as) both the sector size of the device and the blocksize of the filing system which the database is on is a good idea, but sector size seems to always be 4K or less (at least according to https://en.wikipedia.org/wiki/Disk_sector) and FS block size still seems to be 4K by default (the widely used Linux ext4 FS potentially supports up to 64K but only up to the system page size which is 4K on e.g. x86 and x86-64). So in practice a Xapian blocksize of 4KB or more will satisfy this.
The main benefits a larger blocksize gives are slightly more efficient packing and reduced total per-block overheads (and the additional gains here are likely to be smaller for each extra block size doubling), while the downside is needing to read/write more data to read/write a single block. The extra data is at least contiguous (at least in file offset terms - maybe not always on disk if the file is fragmented) but there are potentially significant negative factors like added pressure on the drive cache and OS file cache. The additional losses are likely to grow for each extra block size doubling.
In general for most people just using the default block size is sensible. It's something you might tune when you either care more about reducing size over anything else, or if you're prepared to profile your complete system with different block sizes to see what works best for your own situation.
If profiling different blocksizes including the 8KB default, remember to use a compacted version for the 8KB block size database or else you won't get a fair comparison.
When building an index for a very large amount of data, it can be desirable to index the data in smaller chunks (perhaps on separate machines), and then merge the chunks together into a single database. This can be performed using the xapian-compact tool, simply by supplying several source database paths.
Normally, merging works by reading the source databases in parallel, and writing the contents in sorted order to the destination database. This will work most efficiently if excessive disk seeking can be avoided; if you have several disks, it may be worth placing the source databases and the destination database on separate disks to obtain maximum speed.
The xapian-compact tool supports an additional option, --multipass, which is useful when merging more than three databases. This will cause the postlist tables to be grouped and merged into temporary tables, which are then grouped and merged, and so on until a single postlist table is created, which is usually faster, but requires more disk space for the temporary files.
Xapian includes a command-line tool to check that a database is self-consistent. This tool, xapian-check, runs through the entire database, checking that all the internal nodes are correctly connected. It can also be used on a single table, for example, this command will check the termlist table of database "foo":
xapian-check foo/termlist.DB
The "xapian-check" tool is capable of fixing corrupted databases in certain limited situations. Currently it only supports this for chert, where it is capable of:
- Regenerating a damaged iamchert file (if you've lost yours completely just create an invalid one, e.g. with touch iamchert).
- Regenerating damaged or lost base files from the corresponding DB files. This was developed for the scenario where the database is freshly compacted but should work provided the last update was cleanly applied. If the last update wasn't actually committed, then it is possible that it will try to pick the root block for the partial update, which isn't what you want. If you are in this situation, come and talk to us - with a testcase we should be able to make it handle this better.
To fix such issues, run xapian-check like so:
xapian-check /path/to/database F
This can be done using the copydatabase example program included with Xapian. This is a lot slower to run than xapian-compact, since it has to perform the sorting of the term occurrence data from scratch, but should be faster than a re-index from source data since it doesn't need to perform the tokenisation step. It is also useful if you no longer have the source data available.
The following command will copy a database from "SOURCE" to "DESTINATION", creating the new database at "DESTINATION" as a chert database:
copydatabase SOURCE DESTINATION
By default copydatabase will renumber your documents starting with docid 1. If the docids are stored in or come from some external system, you should preserve them by using the --no-renumber option:
copydatabase --no-renumber SOURCE DESTINATION
The chert format changed in 1.1.4 - at that point the format hadn't been finalised, but a number of users had already deployed it, and it wasn't hard to write an updater, so we provided one called xapian-chert-update which makes a copy with the updated format:
xapian-chert-update SOURCE DESTINATION
It works much like xapian-compact so should take a similar amount of time (and results in a compact database). The initial version had a few bugs, so use xapian-chert-update from Xapian 1.2.5 or later.
The xapian-chert-update utility was removed in Xapian 1.3.0, so you'll need to install Xapian 1.2.x to use it.
It is possible to convert a flint database to a chert database by installing Xapian 1.2.x (since this has support for both flint and chert) using the copydatabase example program included with Xapian. This is a lot slower to run than xapian-compact, since it has to perform the sorting of the term occurrence data from scratch, but should be faster than a re-index from source data since it doesn't need to perform the tokenisation step. It is also useful if you no longer have the source data available.
The following command will copy a database from "SOURCE" to "DESTINATION", creating the new database at "DESTINATION" as a chert database:
copydatabase SOURCE DESTINATION
By default copydatabase will renumber your documents starting with docid 1. If the docids are stored in or come from some external system, you should preserve them by using the --no-renumber option (new in Xapian 1.2.5):
copydatabase --no-renumber SOURCE DESTINATION
It is possible to convert a quartz database to a flint database by installing Xapian 1.0.x (since this has support for both quartz and flint) and using the copydatabase example program included with Xapian. This is a lot slower to run than xapian-compact, since it has to perform the sorting of the term occurrence data from scratch, but should be faster than a re-index from source data since it doesn't need to perform the tokenisation step. It is also useful if you no longer have the source data available.
The following command will copy a database from "SOURCE" to "DESTINATION", creating the new database at "DESTINATION" as a flint database:
copydatabase SOURCE DESTINATION
In 0.9.x, flint was the development backend.
Due to a bug in the flint position list encoding in 0.9.x which made flint databases non-portable between platforms, we had to make an incompatible change in the flint format. It's not easy to write an upgrader, but you can convert a database using the following procedure (although it might be better to rebuild from scratch if you want to use the new UTF-8 support in Xapian::QueryParser, Xapian::Stem, and Xapian::TermGenerator).
Run the following command in your Xapian 0.9.x installation to copy your 0.9.x flint database "SOURCE" to a new quartz database "INTERMEDIATE":
copydatabase SOURCE INTERMEDIATE
Then run the following command in your Xapian 1.0.y installation to copy your quartz database to a 1.0.y flint database "DESTINATION":
copydatabase INTERMEDIATE DESTINATION