several existing strategies to clustering Lucene
from Lucene in a cluster
Lucene is a highly optimized inverted index search engine. It stored a number of inverted indexes in a custom file format that is highly optimized to ensure that the indexes can be loaded by searchers quickly and searched efficiently. These structures are create so that they are almost completely pre-computed.
To store the index, Lucene uses an implementation of a 'Directory' interface, not to be confused with anything in java.io. . The standard implementation if FSDirectory that stored the search index on a file system. There a number of other implementations that can bused including ones to split the index on the filesystem into smaller chunks, and ones to distribute the index throughout a cluster using Map Reduce (see google). There is additionally a database implementation that stored the index as blocks in a database.
Lucene derives its speed from this index structure, and to work really well it needs to be able to seek efficiently into the blocks of the segments that make up the index. This is trivial where the underlying storage mechanism supports seek, but less trivial if the storage mechanism does not. The FSDirectory is based on files, and is efficient in this area. If the files are on a local file system, pure seeks can be used. If the index is on a shared file system , there will always be some latency and potentially increased IO traffic. The Database implementation is highly dependent on the the blob implementation in the target database and will nearly always be slower than the FSDirectory. Some databases support seekable blobs (Oracle), some emulate this behavior (MySQL with emulateLocators=true), others just don't support it and so are really slow. (and I mean really slow)
All of this impacts how Lucene works in a cluster. Each node performing the search needs access to the index. To make search work in a clustered environment we must provide this. There are 3 ways of doing this.
- Use a shared file system between all nodes, and use FSDirectory.
- Use indexes on the nodes local file system and a synchronization strategy.
- Use a database using JDBCDirectory
- Use a distributed file system (eg Google File System, Nutch Distributed File System)
- Use a local cache with backup in the Database
Shared filesystem
There are a number of issues with a shared file system. Performance is lower than a local file system (obviously), unless a SAN is used, but a SAN shared file system must be a true SAN file system (eg Redhat Global File System, Apple XSan) as modifications to the file system blocks must be mirrored instantly in the block cache of all connected nodes, otherwise they will see a corrupted file system. Remember a SAN is just a networked block device, that without additional help cannot be shared by multiple compute nodes at the same time. Provided the performance of the shared file system is sufficient, Lucene works well like this with no modifications using the FSDirectory implementation. The implementation of the lock managed in the Sakai Search component eliminates problems with locks reported by the Lucene community.
This mechanism is available now in Sakai Search.
Synchronized Local indexes.
Where the architecture of the cluster is a shared nothing architecture, the Lucene indexes can be written to local disk and synchronized at the end of each index cycle. This is an optimal deployment of Lucene in a cluster as it ensures that all the IO is from the local disk and is hence fast. To ensure that there is always a back up copy of the index, the synchronization would also target a backup location.
The difficulty with this approach is that without support in the implementation of the search engine, it requires some deployment support. This may involve include making hard link mirrors to speed up the synchronization process. Lucene indexes are suitable for synchronizing with rsync which is a block based synchronization mechanism.
The main drawback of this approach is that the full index is present on the local machine. In large search environments, this duplication will be wastefully, however in search engine terms, a single deployment of Sakai will probably never get into the large space ( large > 100M documents, 2TB index)
This mechanism is available, but requires local configuration
Database hosted search index.
Where a simple cluster setup is required, a database hosted search index is straightforward option. There are however significant drawbacks with this approach, most notable being the drop in performance. The index is stored as blocks in blobs inside the database. These blobs are stored in a block structure to eliminate most of the unnecessary loading however each blob bypasses any local disk block cache on the local machine and has to be streamed over the network. If the database supports seekable blobs, within the database itself, it is possible to minimize unnecessary network traffic. Oracle has this support. However where the database only emulated this behavior (MySQL) the performance is poor as the complete blob needs to be streamed over the network. In addition to this the speed of access is slower since a SQL statement has to be executed for each data access.
The net result is slower performance.
This mechanims is available, but performance is probably unacceptable
Distributed File System
Real Search Engines use a distributed file system that provides a self healing file system where the data itself is distributed across multiple nodes in such a way that the file system can recover from the loss of one or more nodes. The original file system of this form is the Google File System and the Nutch Distributed File System is modeled on Google File System. Both implementations use a gather scatter algorithm detailed by Google in Map-Reduce (see Google labs).
This approach results in every node containing a part of the file system. Where the index size has grown to such an extent to make the storage of the complete index on every node in the cluster, this approach becomes more attractive.
At the moment there are no plans to provide an implementation of a distributed file system within Sakai.
Database Clustered Local Search
In this approach, indexes are used from local disk, but backed up to the database as Lucene Segments. A cluster app node is installed, it synconizes the local copy of the search index with the database. When new content is added by one of the cluster app nodes, it updates the backup copy in the database. On reciept of the index reload events, all cluster app nodes resyncronize the with the database downloading changed and new search segments.
This mechanism in in the process of being tested, I exhibits the same performance as a local basaed search for a 200MB index with 80,000 documents.
Once this mechanism is completely tested it will become the default OOTB mechanism, as it works where there is a single cluster node or more than one cluster node. The added advantage of this mechanism is that the index is stored in the database.
It will also be possible to implement this mechanism with a shared filestore acting as the backup location.