NOTE: Development of Distributed Collections for Scala is discontinued. The repository will remain online for existing references. The reasons for discontinuations are:
Incompatibility with Scala Collections - Distributed collections require different interface for collections. Methods like
immutable.Mapwhich is not compatible with distributed collections. Furthermore, in some cases data manifests, or similar constructs are needed for efficient serialization but those would require significant changes in the common signature.
Rich signature of Scala Collections - The Scala collections hierarchy contains methods whose signature is intended for in-memory operation. By default Scala Collections eagerly execute the code which is for distributed collections overly expensive. Furthermore, abstractions like
Setrequire a barrier after every operation making them very slow even for basic operations like
map. Finally, the Scala Collections expose methods like
zipWithIndexwhich are used frequently but are expensive in the distributed case. Exposing these methods would lure users into using them but their performance on large data would be unacceptably low.
Use cases - Abstractions like
Setare widely used in shared memory environment for fast lookups. In case of distributed data this functionality is rarely needed and requires traversal of the whole data set. However, maintenance of these abstractions in the distributed case is very high.
Distributed Collections for Scala is a library for large scale data processing that uses different cluster computing frameworks as the back-end. Library inherits Scala 2.9.1 collections generic interface enriched with additional methods like
Currently the library uses only Hadoop as the back-end processing engine. However, we are aiming to extend the library to work with other frameworks like Spark, HaLoop and Nephele.
Project is built by Simple Build Tool and can be packaged by invoking
sbt update package from project root. Building is also available through Maven by invoking
After packaging you need to copy the library and all dependencies to the $HADOOP-HOME/lib:
Project consists of 4 modules:
- distriubted-collections - Distributed Collections Library
- hadoop-backend - back-end that executes collections operations on Hadoop MapReduce
- beckend-api - api shared between the library and back-end components
- benchmarks - Distributed Collections benchmarks
Project is being developed by IntelliJ IDEA 10.x and following parts of the project tree are for IDEA convenience purposes only:
- .idea - project folder
Using the library is very similar to using Scala collections and parallel collections with several key differences:
- Collections are instantiated from the URI that points to the data
- Many operations use Int data type which can easily overflow in the domain so long versions should be used instead
In the communication between the cluster nodes every piece of data is serialized and deserialized and the user must be careful not to accidentally transfer large parts of object graph
For more detailed description see the wiki page