ECL (data-centric programming language)
Template:New unreviewed article
| Paradigm | declarative structured, data-centric |
|---|---|
| First appeared | 2000 |
| Typing discipline | static, strong, safe |
| Website | hpccsystems |
| Dialects | |
| UCSD, Borland, Turbo | |
| Influenced by | |
| Prolog, Pascal, SQL, Snobol4, C++, Clarion | |
| Influenced | |
| Big Data | |
ECL is a declarative, data centric programming language designed in 2000 to allow a team of programmers to process Big Data across a high performance computing cluster without the programmer being involved in many of the lower level, imperative decisions.[1]
History
ECL was initially designed and developed in 2000 as an in-house productivity tool within Seisint Inc and was considered to be ‘secret weapon’ that allowed Seisint to gain market share in its data business. The technology was cited as a driving force behind the acquisition of Seisint by LexisNexis and then again as a major source of synergies when LexisNexis acquired ChoicePoint Inc.
Implementations
The first implementation of ECL in June 2000 translated the input ECL into a variant of SQL, to run on a (now retired) in-memory query engine known as hOle. Later in 2000 a second implementation of an ECL execution engine was created (known as Thor), which ran on a cluster of Windows 2000 servers, and the ECL compiler was extended to generate C++ code, which was then compiled using MSVC to create executable DLLs that the execution engine would load and run. In 2002 the engines were ported to Linux and the ECL compiler extended to support generation of Gnu g++ code. A third execution engine, designed for rapid repeated execution of similar queries (known as Roxie) was also developed around this time, using the same ECL compiler, language, and generated DLL technology.
Language Constructs
ECL, at least in its purest form, is a declarative, data centric language. Programs, in the strictest sense, do not exist. Rather an ECL application will specify a number of core datasets (or data values) and then the operations which are to be performed on those values.
Hello world
ECL is to have succinct solutions to problems and sensible defaults. The ‘Hello World’ program is characteristically short: ‘Hello World’ Perhaps a more flavorful example would take a list of strings, sort them into order, and then return that as a result instead.
// First declare a dataset with one column containing a list of strings
// Datasets can also be binary, csv, xml or externally defined structures
D := DATASET([{'ECL'},{'Declarative'},{'Data'},{'Centric'},{'Programming'},{'Language'}],{STRING Value;});
SD := SORT(D,Value);
output(SD)
The statements containing a := are defined in ECL as attribute definitions. They do not denote an action; rather a definition of a term. Thus, logically, an ECL program can be read: “bottom to top”
OUTPUT(SD)
What is an SD?
SD := SORT(D,Value);
SD is a D that has been sorted by ‘Value’
What is a D?
D := DATASET([{'ECL'},{'Declarative'},{'Data'},{'Centric'},{'Programming'},{'Language'}],{STRING Value;});
D is a dataset with one column labeled ‘Value’ and containing the following list of data.
ECL Primitives
ECL primitives that act upon datasets include: SORT, ROLLUP, DEDUP, ITERATE, PROJECT, JOIN, NORMALIZE, DENORMALIZE, PARSE, CHOOSEN, ENTH, TOPN, DISTRIBUTE
ECL Encapsulation
Whilst ECL is terse and LexisNexis claims that 1 line of ECL is roughly equivalent to 120 lines of C++ it still has significant support for large scale programming including data encapsulation and code re-use. The constructs available include: MODULE, FUNCTION, INTERFACE, MACRO, EXPORT, SHARED
Support for Parallelism in ECL
In the HPCC implementation, by default, most ECL constructs will execute in parallel across the hardware being used. Many of the primitives also have a LOCAL option to specify that the operation is to occur locally on each node.
Comparison to Map-Reduce
The Hadoop Map-Reduce paradigm actually consists of three phases which correlate to ECL primitives as follows.
| Hadoop Name/Term | ECL equivalent | Comments |
|---|---|---|
| MAPing within the MAPper | PROJECT/TRANSFORM | Takes a record and coverts to a different format; in the Hadoop case the conversion is into a key-value pair |
| SHUFFLE (Phase 1) | DISTRIBUTE(,HASH(KeyValue)) | The records from the mapper are distributed dependent upon the KEY value |
| SHUFFLE (Phase 2) | SORT(,LOCAL) | The records arriving at a particular reducer are sorted into KEY order |
| REDUCE | ROLLUP(,Key,LOCAL) | The records for a particular KEY value are now combined together |