NAME

ovdb - Overview storage method for INN

DESCRIPTION

The ovdb overview is a storage method that uses the Berkeley DB library to store overview data. It requires version 4.4 or later of the Berkeley DB library (4.7+ is recommended because older versions suffer from various issues).

The ovdb overview method makes use of the full transaction/logging/locking functionality of the Berkeley DB environment. Berkeley DB may be downloaded from <http://www.oracle.com/technetwork/database/database-technologies/berkeleydb/overview/index.html> and is needed to build the ovdb backend.

UPGRADING

There are several versions of the ovdb storage method:

Version 1, the initial version shipped with INN 2.3.0 up to INN 2.3.5.
Version 2, with improved performance, since INN 2.4.0.
Version 3, corresponding to version 2 with compression enabled, starting with INN 2.5.0.

If you have a database created with a previous version of ovdb, your database will need to be upgraded using ovdb_init. See the ovdb_init(8) man page for upgrade instructions, as well as the COMPRESSION section below.

Note that when the Berkeley DB library is updated to a newer version, the ovdb database also needs being upgraded.

INSTALLATION

If the Berkeley DB library is found at configure time, INN will be built with Berkeley DB support unless the --without-bdb flag is explicitly passed to configure. By default, configure will search for Berkeley DB in standard locations; there will be a message in the configure output indicating the pathname that will be used.

You can override this pathname by adding a path to the option, for instance --with-bdb=/usr/BerkeleyDB.4.4. This directory is expected to have subdirectories include and lib (lib32 and lib64 are also checked), containing respectively db.h, and the library itself. In case non-standard paths to the Berkeley DB libraries are used, one or both of the options --with-bdb-include and --with-bdb-lib can be given to configure with a path.

The ovdb database may take up more disk space for a given spool than the other overview methods. Plan on needing at least 1.1 KB for every article in your spool (not counting crossposts). So, if you have 5 million articles, you'll need at least 5.5 GB of disk space for ovdb. With compression enabled, this estimate changes to 0.9 KB per article, so you'll need at least 4.5 GB of disk space for 5 million articles. See the COMPRESSION section below. Plus, you'll need additional space for transaction logs: at least 100 MB. By default, the transaction logs go in the same directory as the database. To improve performance, they can be placed on a different disk -- see the DB_CONFIG section.

CONFIGURATION

To enable the ovdb overview method, set the ovmethod parameter in inn.conf to "ovdb". The ovdb database is stored in the directory specified by the pathoverview parameter in inn.conf. This is the "DB_HOME" directory. To start out, this directory should be empty (other than an optional DB_CONFIG file; see DB_CONFIG for details), and innd (or makehistory) will create the files as necessary in that directory. Also, make sure the directory is owned by the news user.

Other parameters for configuring ovdb are in the ovdb.conf configuration file. The following parameters can be set in that file:

compress

If INN was compiled with zlib, and this compress parameter is true, ovdb will compress overview records that are longer than 600 bytes. See the COMPRESSION section below.

cachesize

Size of the memory pool cache, in kilobytes. The cache will have a backing store file in the DB directory which will be at least as big. In general, the bigger the cache, the better. Use "ovdb_stat -m" to see cache hit percentages. To make a change of this parameter take effect, shut down and restart INN (be sure to kill all of the nnrpd processes when shutting down). Default is 8000 (KB), which is adequate for small to medium-sized servers. Large servers will probably need at least 20000 (KB).

ncache

Number of regions across which to split the cache. The region size is equal to cachesize divided by ncache. Default is 1 for ncache, that is to say the cache will be allocated contiguously in memory.

numdbfiles

Overview data is split between this many files. Currently, innd will keep all of the files open, so don't set this too high or innd may run out of file descriptors. nnrpd only opens one at a time, regardless. May be set to one, or just a few, but only do that if your OS supports large (> 2 GB) files. Changing this parameter has no effect on an already-established database. Default is 32.

txn_nosync

If txn_nosync is set to false, Berkeley DB flushes the log after every transaction. This minimizes the number of transactions that may be lost in the event of a crash, but results in significantly degraded performance. Default is true.

useshm

If useshm is set to true, Berkeley DB will use shared memory instead of mmap for its environment regions (cache, lock, etc). With some platforms, this may improve performance. Default is false.

shmkey

Sets the shared memory key used by Berkeley DB when useshm is true. Berkeley DB will create several (usually 5) shared memory segments, using sequentially numbered keys starting with "shmkey". Choose a key that does not conflict with any existing shared memory segments on your system. Default is 6400.

pagesize

Sets the page size for the DB files (in bytes). Must be a power of 2. Best choices are 4096 or 8192. The default is 8192. Changing this parameter has no effect on an already-established database.

minkey

Sets the minimum number of keys per page. See the Berkeley DB documentation for more information. Default is based on page size and whether compression is enabled:

   default_minkey = MAX(2, pagesize / 2600) if compress is false
   default_minkey = MAX(2, pagesize / 1500) if compress is true

The lowest allowed minkey is 2. Setting minkey higher than the default is not recommended, as it will cause the databases to have a lot of overflow pages. Changing this parameter has no effect on an already-established database.

maxlocks

Sets the Berkeley DB lk_max parameter, which is the maximum number of locks that can exist in the database at the same time. Default is 4000.

nocompact

The nocompact parameter affects the behaviour of expireover. The expireover function in ovdb can do its job in one of two ways: by simply deleting expired records from the database; or by re-writing the overview records into a different location leaving out the expired records. The first method is faster, but it leaves 'holes' that result in space that can not immediately be reused. The second method 'compacts' the records by rewriting them.

If this parameter is set to 0, expireover will compact all newsgroups; if set to 1, expireover will not compact any newsgroups; and if set to a value greater than one, expireover will only compact groups that have less than that number of articles.

Experience has shown that compacting has minimal effect (other than making expireover take longer) so the default is 1. This parameter will probably be removed in the future.

readserver

When the readserver parameter is set to false, each nnrpd process directly accesses the Berkeley DB environment. The process of attaching to the database (and detaching when finished) is fairly expensive, and can result in high loads in situations when there are lots of reader connections of relatively short duration.

When the readserver parameter is set to true, the nnrpd processes will access overview via a helper server (ovdb_server -- which is started by ovdb_init). All ovdb reads will then be funnelled through a single process with a cleaner interface to the underlying Berkeley DB database. This will result in cleaner shutdowns for the database, improving stability and avoiding deadlocks, timing issues and corrupted databases. That's why you should try to set this parameter to true if you are experiencing any instability in the ovdb overview method.

Default value is true.

numrsprocs

This parameter is only used when readserver is true. It sets the number of ovdb_server processes. As each ovdb_server can process only one transaction at a time, running more servers can improve reader response times. Default is 5.

maxrsconn

This parameter is only used when readserver is true. It sets a maximum number of readers that a given ovdb_server process will serve at one time. This means the maximum number of readers for all of the ovdb_server processes is (numrsprocs * maxrsconn). This does not limit the actual number of readers, since nnrpd will fall back to opening the database directly if it can't connect to an ovdb_server. Default is 0, which means an unlimited number of connections is allowed.

COMPRESSION

The ovdb storage method has the ability to compress overview data before it is stored into the database. In addition to consuming less disk space, compression keeps the average size of the database keys smaller. This in turn increases the average number of keys per page, which can significantly improve performance and also helps keep the database more compact. This feature requires that INN be built with zlib. Only records larger than 600 bytes get compressed, because that is the point at which compression starts to become significant.

If compression is not enabled (either from the compress option in ovdb.conf or INN was not built with zlib support), the database will be backward compatible with older versions of ovdb. However, if compression is enabled, the database is marked with a newer version that will prevent older versions of ovdb from opening the database.

You can upgrade an existing database to use compression simply by setting compress to true in ovdb.conf. Note that existing records in the database will remain uncompressed; only new records added after enabling compression will be compressed.

If you disable compression on a database that previously had it enabled, new records will be stored uncompressed, but the database will still be incompatible with older versions of ovdb (and will also be incompatible with this version of ovdb if INN was not built with zlib support). So to downgrade to a completely uncompressed database, you will have to rebuild the database using makehistory.

DB_CONFIG

A file called DB_CONFIG may be placed in the database directory (pathoverview in inn.conf) to customize where the various database files and transaction logs are written. By default, all of the files are written in the "DB_HOME" directory. One way to improve performance is to put the transaction logs on a different disk. To do this, put:

    DB_LOG_DIR /path/to/logs

in the DB_CONFIG file. If the pathname you give starts with a "/", it is treated as an absolute path; otherwise, it is relative to the "DB_HOME" directory. Make sure that any directories you specify exist and have proper ownership/mode before starting INN, because they won't be created automatically. Also, don't change the DB_CONFIG file while anything that uses ovdb is running.

Another thing that you can do with this file is to split the overview database across multiple disks. In the DB_CONFIG file, you can list directories that Berkeley DB will search when it goes to open a database.

For example, let's say that you have pathoverview set to /mnt/overview and you have four additional file systems created on /mnt/ovX. You would create a file /mnt/overview/DB_CONFIG containing the following lines:

    set_data_dir /mnt/overview
    set_data_dir /mnt/ov1
    set_data_dir /mnt/ov2
    set_data_dir /mnt/ov3
    set_data_dir /mnt/ov4

Distribute your ovNNNNN files into the four filesystems (say, 8 each). When called upon to open a database file, the db library will look for it in each of the specified directories (in order). If said file is not found, one will be created in the first of those directories.

Whenever you change DB_CONFIG or move database files around, make sure all news processes that use the database are shut down first (including nnrpd processes).

The DB_CONFIG functionality is part of Berkeley DB itself, rather than something provided by ovdb. See the Berkeley DB documentation for complete details for the version of Berkeley DB that you're running.

RUNNING

When starting the news system, rc.news will invoke the ovdb_init program. See the ovdb_init(8) man page for information about the tasks it performs. ovdb_init must be run before using the database.

And when stopping INN, rc.news kills the ovdb_monitor processes after the other INN processes have been shut down.

DIAGNOSTICS

Problems relating to ovdb are logged to news.err with "OVDB" in the error message.

INN programs that use overview will fail to start up if the ovdb_monitor processes aren't running. Be sure to run ovdb_init before running anything that accesses overview.

Also, INN programs that use overview will fail to start up if the user running them is not the news user.

If a program accessing the database crashes, or otherwise exits uncleanly, it might leave a stale lock in the database. This lock could cause other processes to deadlock on that stale lock. To fix this, shut down all news processes (using "kill -9" if necessary) and then restart. ovdb_init should perform a recovery operation which will remove the locks and repair damage caused by killing the deadlocked processes.

FILES

pathetc/inn.conf: The ovmethod and pathoverview parameters are relevant to ovdb.
pathetc/ovdb.conf: Optional configuration file for tuning. See CONFIGURATION above.
pathoverview: Directory where the database goes. Berkeley DB calls it the "DB_HOME" directory.
pathoverview/DB_CONFIG: Optional file to configure the layout of the database files.
pathrun/ovdb.sem: A file that gets locked by every process that is accessing the database. This is used by ovdb_init to determine whether the database is active or quiescent.
pathrun/ovdb_monitor.pid: Contains the process ID of ovdb_monitor.

TO DO

Implement a way to limit how many databases can be open at once (to reduce file descriptor usage); maybe using something similar to the cache code in legacy ov3.c file.

HISTORY

Written by Heath Kehoe <hakehoe@avalon.net> for InterNetNews.