Storage

TIP

• If the database does not shrink after deleting documents or indexes, it is expected behavior. You are not losing space, MeiliSearch is keeping this space for performance reasons.
• You should have the same amount of RAM than the space taken on disk by MeiliSearch for optimal performances.

MeiliSearch is a database. It stores the indexed documents along with the data needed to perform lightning search.

Writing a database is hard, and we do not want to reinvent the wheel, so MeiliSearch uses a storage engine under the hood. Using a storage engine allows MeiliSearch to focus on improving search relevancy and search performance while abstracting the complicated task of creating, reading and updating documents on disk, and in memory.

LMDB

The storage engine of MeiliSearch is a Lightning Memory-Mapped Database (opens new window) (LMDB for short). LMDB is a transactional key-value store written in C that was developed for OpenLDAP, and it has ACID properties.

We chose LMDB after we successfully (or not) tried MeiliSearch with Sled (opens new window) and RocksDB (opens new window) and decided to move on with LMDB because it is the best combination of performance and stability for MeiliSearch.

Memory mapping

LMDB stores its data in a memory-mapped file (opens new window). All data fetched from LMDB is returned straight from the memory map, which means there is no memory allocation or memory copy during data fetches.

All documents stored on disk are automatically loaded in memory when MeiliSearch asks for them. This ensures LMDB will always make the best use of the RAM available to retrieve the documents.

For the best performance, it is recommended to provide the same amount of RAM as the size the database takes on disk, so all the data structures can fit in memory.

Understanding LMDB

The choice of LMDB comes with certain pros and cons. In order to understand this choice, its upsides and downsides, we need to have an insight on how LMDB impact size and memory usage. This is well explained in a blogpost of LMDB (opens new window) and we are trying to summarize it here.

Database size

When freeing entries from the database (in our case, removing documents from MeiliSearch), one can observe that no space disk is released. The space previously used by the entry is marked as free for LMDB but not made available for the operating system.
Unlike other storage engines, LMDB chooses this design for performance issues as there is no need for a compaction phase.

As a result, you may see that the disk occupied by LMDB and therefore by MeiliSearch keeps growing even if you are deleting indexes or documents. This is normal behavior, and you can note that the database will not grow again if you write some data after deleting indexes or documents.

Memory usage

Since LMDB is memory mapped, it is the operating system that manages the real memory allocated (or not) to MeiliSearch.

Thus, if you run MeiliSearch as a standalone program on a server, LMDB will use the maximum RAM it can use.

On the other hand, if you run MeiliSearch along with other programs, the OS will manage memory based on everyone’s need. This makes MeiliSearch’s memory usage quite flexible when used in development.

TIP

Virtual Memory != Real Memory
Virtual memory is the disk space a program requests from the OS. It is not the memory that the program will actually use.

MeiliSearch will always demand a certain amount of space to use as a memory map. This space will be used as virtual memory, but the amount of real memory (RAM) used will be much smaller.

Measured disk usage

The following measurements were taken using an 8.6 MB JSON dataset containing 19,553 documents.

After indexed, the dataset size in LMDB is about 122MB.

This means the database is using 305 MB of RAM and 224 MB of disk space. Note that virtual memory (opens new window) refers only to disk space allocated by your computer for MeiliSearch—it does not mean that it’s actually in use by the database. See Memory Usage for more details.

WARNING

These metrics are highly dependent on the machine that is running MeiliSearch. Running this test in significantly underpowered machines is likely to give different results.

It is important to note that there is no reliable way to predict the final size of a database. This is true for just about any search engine on the market—we’re just the only ones saying it out loud.

Database size is affected by a large number of criteria, including settings, relevancy rules, use of facets, the number of different languages present, and more.