I'm looking for a dbm-like library that I can use in place of Berkeley DB, which I'm currently using. My main reason for switching is the licensing fees for BDB are pretty high (free for open source apps, but my employer does not want to open source this particular app for various reasons).

I've looked briefly at qdbm but it doesn't look like it will fill my needs -- lots of keys (several million) and large data items (> 1-5 megabytes). Before I continue my search I figured I'd ask because it seems there are tons of dbm-like libraries out there.

Simon Schmidt

C/C++

http://sphia.org/ a Key/Value Store, Written in C and BSD-Licensed.
https://code.google.com/p/leveldb/ Written in C++, also BSD-Licensed
http://luxio.sourceforge.net/ Written in C++, LGPL, faster than Berkeley DB
http://symas.com/mdb/ Written in C, License is OpenLDAP Public License (BSD-Like)
https://github.com/davidar/sdbm Written in C, Public Domain, Reimplementation of "ndbm"
http://repetae.net/computer/vsdb/ Written in C, License unknown, Experimental
https://github.com/stevedekorte/skipdb Written in C, BSD-Licensed, Key-value Store using Skip-List indexes
http://unqlite.org/ written in C/C++, Embedded NoSQL Database Engine, BSD-Licensed, Similar to SQLite.
http://hamsterdb.com/ Written in C/C++, Embedded NoSQL Database Engine, Apache License 2.0.

Java

https://github.com/dain/leveldb A Java port of LevelDB, Apache License 2.0
https://github.com/jingwei/krati A former LinkedIn Project, Apache License 2.0
http://www.mapdb.org/ An Embedded Java Database, Apache License 2.0, supports Key-Value Storage, very efficient Object serialisation
https://github.com/OpenHFT/Chronicle-Map, LGPL v3, embedded pure Java key-value store, faster than Berkeley DB
https://github.com/jetbrains/xodus, Apache License 2.0, a Java transactional schema-less embedded database, faster than Berkeley DB
https://github.com/lmdbjava/lmdbjava, JNI binding for LMDB, Apache License 2.0

Phillip B Oldham

You could look at Tokyo Cabinet. Its the successor to qdbm/gdbm, and if you decide to scale has a nice network front-end available.

Edit:

Another variant is Kyoto Cabinet; developed by the same person, but supposedly easier to use.

SQLite is public domain, meaning you can use it for any purpose whatsoever, and is widely used and supported.

You can get much improved performance out of any dbm (even qdbm) and improved parallelism with a simple level of indirection: Simply take your keys and hash them, and use data_dir/H(key)/ as the database to store those keys. Limit the hash output to a small value (say) 255 for best results.

This approach has a number of benefits, easily summarized: