The original source https://github.com/dmbarbour/haskell-lmdb
See ./CONCORDIUM-FORK.md for reasons for the fork.
Bindings to LMDB from Haskell. This will be released as just the lmdb package on hackage, since the Haskell aspect is implicit. To install haskell-lmdb, you'll need to first install the lmdb development files (e.g. sudo apt-get install liblmdb-dev works in at least my 14.04 version of Ubuntu).
These bindings are developed against LMDB 0.9.10. If LMDB is updated in a significant way that e.g. adds new features or relaxes old constraints, and it isn't clear that I've already noticed, please file an issue report!
Lightning MDB (LMDB) is a Berkeley DB replacement developed for the OpenLDAP project, primarily by Howard Chu and Symas Corp. Besides its performance, which is impressive, LMDB has much to offer:
- commercial-friendly BSD-style Open LDAP license
- minimal configuration and administration
- stable latencies suitable for real-time
- tight memory usage without double-buffering
- simplicity, transparency, comprehensibility
While key-value databases are becoming more popular, it seems many of them are adding lots of features (new table types, clever optimizations with thread-local storage, etc.). This makes me wary: new features frequently come with new bugs. At least so far, it seems LMDB has avoided this trend and focused instead on refining the relatively few features it has. LMDB does one thing - memory-mapped B-trees - and does so very well.
LMDB is read-optimized, and outperforms most existing databases for read-only transactions. Write performance is also good, outperforming most other B-tree databases such as Berkeley DB. Log structured merge trees, such as LevelDB, are write-optimized and will tend to outperform LMDB in that area. LMDB is a very good choice for a read-heavy workload.
Known weaknesses of LMDB:
- Read-only vs read-write transactions must be decided ahead of time.
- No concurrency for writers. No merging non-conflicting writes.
- The database does not shrink. No direct support for compacting.
- Must set a maximum database size when opening. Cannot grow dynamically.
From a Haskell programmer's perspective, LMDB's API is very ad-hoc. Functions do many different things based on flags, and it's difficult to tell what a function does or returns just by studying the type signatures. Transactions and cursors and databases all have many different, implicit 'types' - e.g. there are MDB_DUPSORT databases that allow multiple values for a key, and MDB_DUPFIXED databases with fixed-size elements allowing calls to MDB_MULTIPLE. The same transaction type is used for both read-only and read-write transactions.
At the moment, I only have a 'raw' interface, Database.LMDB.Raw, which is mostly faithful to #include <lmdb.h> in C. The raw interface does follow common Haskell FFI conventions, e.g. representing flag fields with lists, enumerations with datatypes, and errors with exceptions. However, there are many safety caveats for the raw interface:
- write transactions must use an OS bound thread
- don't open the same environment file twice!
- an
MDB_valmustn't be used outside its transaction - the normal LMDB API flags and functions cautions
- very low level cursor 'get' binding
LMDB's writer mutex, which is held for the full duration of a write transaction, causes all sorts of problems for Haskell's lightweight threads. I currently introduce a simple MVar mutex at the MDB_env layer. Opening the same environment twice (concurrently) would bypass this. Fortunately, the expected use-case for LMDB is to open an environment once when the application starts up, then keep it open until the application shuts down or crashes.
Safe FFI calls are enough overhead to significantly impact a microbenchmark reading the LMDB database (especially if we read it sequentially). This might not be a big deal for 'real' code, where we perform any significant processing of returned values. But, in the interest of performance freaks like me, the 'raw' interface exports the read/write/cursor operations twice: once with MDB_dbi and again for MDB_dbi' (and similar for cursors). The difference between these is that the MDB_dbi allows user-defined comparisons but requires 'safe' FFI calls (which adds ~100ns overhead), while MDB_dbi' uses unsafe FFI calls but forbids user-defined comparisons (which adds ~10ns overhead).
The common case for LMDB is to not define your own comparison functions. Thus, use of MDB_dbi' will frequently be the preferable option.
My original plan was to develop a monadic interface using bytestrings and ST-like references to keep the MDB_val within a transaction. I'd also like to further wrap the database types, e.g. to distinguish maps from multimaps at the type level, i.e. using types to distinguish MDB_DUPSORT databases from other databases. And I'd like to have a clean model above the cursors, perhaps developing a cursor monad. However, none of the above will happen anytime soon. I find myself much more interested in moving on and leveraging LMDB for my other projects rather than further developing the LMDB API.
If someone else wishes to take over development of a higher level interface, you're welcome to do so, either via separate package or pull requests or just by asking nicely to become an lmdb cabal package maintainer. :)
If LMDB's API ever becomes unstable, or if there is enough demand for it, I've also considered internalizing a copy of LMDB source code into this project, rather than relying on an external library. I have mixed feelings about that option, but I haven't been able to pin them down with logic quite yet.