For too long S and similar data analysis environments have lacked good interfaces to relational database systems (RDBMS). For the last twenty years or so these RDBMS have evolved into highly optimized client-server systems for data storage and manipulation, and currently they serve as repositories for most of the business, industrial, and research “raw” data that analysts work with. Other analysis packages, such as SAS, have traditionally provided good data connectivity, but S and GNU R have relied on intermediate text files as means of importing data (but see R Data Import/Export (2001) and Using Relational Database Systems with R (2000).) Although this simple approach works well for relatively modest amounts of mostly static data, it does not scale up to larger amounts of data distributed over machines and locations, nor does it scale up to data that is highly dynamic – situations that are becoming increasingly common.
We want to propose a common interface between R/S and RDBMS that would allow users to access data stored on database servers in a uniform and predictable manner irrespective of the database engine. The interface defines a small set of classes and methods similar in spirit to Python’s DB-API, Java’s JDBC, Microsoft’s ODBC, Perl’s DBI, etc., but it conforms to the “whole-object” philosophy so natural in S and R.
As data analysts, we are increasingly faced with the challenge of larger data sources distributed over machines and locations; most of these data sources reside in relational database management systems (RDBMS). These relational databases represent a mature client-server distributed technology that we as analysts could be exploiting more that we’ve done in the past. The relational technology provides a well-defined standard, the ANSI SQL-92 X/Open CAE Specification: SQL and RDA (1994), both for defining and manipulating data in a highly optimized fashion from virtually any application.
In contrast, S and Splus have provided somewhat limited tools for coping with the challenges of larger and distributed data sets (Splus does provide an
import function to import from databases, but it is quite limited in terms of SQL facilities). The R community has been more resourceful and has developed a number of good libraries for connecting to mSQL, MySQL, PostgreSQL, and ODBC; each library, however, has defined its own interface to each database engine a bit differently. We think it would be to everybody’s advantage to coordinate the definition of a common interface, an effort not unlike those taken in the Python and Perl communities.
The goal of a common, seamless access to distributed data is a modest one in our evolution towards a fully distributed computing environment. We recognize the greater goal of distributed computing as the means to fully integrate diverse systems – not just databases – into a truly flexible analysis environment. Good connectivity to databases, however, is of immediate necessity both in practical terms and as a means to help us transition from monolithic, self-contained systems to those in which computations, not only the data, can be carried out in parallel over a wide number of computers and/or systems Temple Lang (2000). Issues of reliability, security, location transparency, persistence, etc., will be new to most of us and working with distributed data may provide a more gradual change to ease in the ultimate goal of full distributed computing.
We believe that a common interface to databases can help users easily access data stored in RDBMS. A common interface would describe, in a uniform way, how to connect to RDBMS, extract meta-data (such as list of available databases, tables, etc.) as well as a uniform way to execute SQL statements and import their output into R and S. The current emphasis is on querying databases and not so much in a full low-level interface for database development as in JDBC or ODBC, but unlike these, we want to approach the interface from the “whole-object” perspective J. M. Chambers (1998) so natural to R/S and Python – for instance, by fetching all fields and records simultaneously into a single object.
The basic idea is to split the interface into a front-end consisting of a few classes and generic functions that users invoke and a back-end set of database-specific classes and methods that implement the actual communication. (This is a very well-known pattern in software engineering, and another good verbatim is the device-independent graphics in R/S where graphics functions produce similar output on a variety of different devices, such X displays, Postscript, etc.)
The following verbatim shows the front-end:
> mgr <- dbManager("Oracle") > con <- dbConnect(mgr, user = "user", passwd = "passwd") > rs <- dbExecStatement(con, "select fld1, fld2, fld3 from MY_TABLE") > tbls <- fetch(rs, n = 100) > hasCompleted(rs)  T > close(rs) > rs <- dbExecStatement(con, "select id_name, q25, q50 from liv2") > res <- fetch(rs) > getRowCount(rs)  73 > close(con)
Such scripts should work with other RDBMS (say, MySQL) by replacing the first line with
> mgr <- dbManager("MySQL")
The following are the main RS-DBI classes. They need to be extended by individual database back-ends (MySQL, Oracle, etc.)
Virtual class1 extended by actual database managers, e.g., Oracle, MySQL, Informix.
Virtual class that captures a connection to a database instance2.
Virtual class that describes the result of an SQL statement.
Virtual class, extends
dbResult to fully describe the output of those statements that produce output records, i.e.,
SELECT-like) SQL statement.
All these classes should implement the methods
prints a short summary of the meta-data of the specified object (like
summary in R/S).
takes an object of one of the above classes and a string specifying a meta-data item, and it returns the corresponding information (
NULL if unavailable).
> mgr <- dbManager("MySQL") > getInfo(mgr, "version") > con <- dbConnect(mgr, ...) > getInfo(con, "type")
The reason we implement the meta-data through
getInfo in this way is to simplify the writing of database back-ends. We don’t want to overwhelm the developers of drivers (ourselves, most likely) with hundreds of methods as in the case of JDBC.
In addition, the following methods should also be implemented:
lists all available databases known to the
lists tables in a database.
lists the fields in a table in a database.
lists the indices defined for a table in a database.
These methods may be implemented using the appropriate
getInfo method above.
In the next few sections we describe in detail each of these classes and their methods.
This class identifies the relational database management system. It needs to be extended by individual back-ends (Oracle, PostgreSQL, etc.) The
dbManager class defines the following methods:
initializes the driver code. We suggest having the generator,
dbManager(driver), automatically load the driver.
releases whatever resources the driver is using.
returns the version of the RS-DBI currently implemented, plus any other relevant information about the implementation itself and the RDBMS being used.
This virtual class captures a connection to a RDBMS, and it provides access to dynamic SQL, result sets, RDBMS session management (transactions), etc. Note that the
dbManager may or may not allow multiple simultaneous connections. The methods it defines include:
opens a connection to the database
dbname. Other likely arguments include
password. It returns an object that extends
dbConnection in a driver-specific manner (e.g., the MySQL implementation creates a connection of class
MySQLConnection that extends
dbConnection). Note that we could separate the steps of connecting to a RDBMS and opening a database there (i.e., opening an instance). For simplicity we do the 2 steps in this method. If the user needs to open another instance in the same RDBMS, just open a new connection.
closes the connection and discards all pending work.
submits one SQL statement. It returns a
dbResult object, and in the case of a
SELECT statement, the object also inherits from
dbResultSet object is needed for fetching the output rows of
SELECT statements. The result of a non-
SELECT statement (e.g.,
UPDATE, DELETE, CREATE, ALTER, …) is defined as the number of rows affected (this seems to be common among RDBMS).
commits pending transaction (optional).
undoes current transaction (optional).
invokes a stored procedure in the RDBMS (tentative). Stored procedures are not part of the ANSI SQL-92 standard and possibly vary substantially from one RDBMS to another. For instance, Oracle seems to have a fairly decent implementation of stored procedures, but MySQL currently does not support them.
submit an SQL “script” (multiple statements). May be implemented by looping with
When running SQL scripts (multiple statements), it closes the current result set in the
dbConnection, executes the next statement and returns its result set.
This virtual class describes the result of an SQL statement (any statement) and the state of the operation. Non-query statements (e.g.,
DELETE) set the “completed” state to 1, while
SELECT statements to 0. Error conditions set this slot to a negative number. The
dbResult class defines the following methods:
returns the SQL statement associated with the result set.
dbConnection associated with the result set.
returns the number of rows affected by the operation.
was the operation completed?
SELECT’s, for instance, are not completed until their output rows are all fetched.
returns the status of the last SQL statement on a given connection as a list with two members, status code and status description.
This virtual class extends
dbResult, and it describes additional information from the result of a
SELECT statement and the state of the operation. The
completed state is set to 0 so long as there are pending rows to fetch. The
dbResultSet class defines the following additional methods:
returns the number of rows fetched so far.
returns a logical vector with as many elements as there are fields in the result set, each element describing whether the corresponding field accepts
SELECTed fields. The description includes field names, RDBMS internal data types, internal length, internal precision and scale, null flag (i.e., column allows
NULL’s), and corresponding S classes (which can be over-ridden with user-provided classes). The current MySQL and Oracle implementations define a
dbResultSet as a named list with the following elements:
the connection object associated with this result set;
a string with the SQL statement being processed;
a field description
data.frame with as many rows as there are fields in the
SELECT output, and columns specifying the
Sclass of the corresponding output field.
the number of rows that were affected;
the number of rows so far fetched;
a logical value describing whether the operation has completed or not.
a logical vector specifying whether the corresponding column may take NULL values.
The methods above are implemented as accessor functions to this list in the obvious way.
defines a conversion between internal RDBMS data types and R/S classes. We expect the default mappings to be by far the most common ones, but users that need more control may specify a class generator for individual fields in the result set. (See Section [sec:mappings] for details.)
closes the result set and frees resources both in R/S and the RDBMS.
extracts the next
max.rec records (-1 means all).
The data types supported by databases are slightly different than the data types in R and S, but the mapping between them is straightforward: Any of the many fixed and varying length character types are mapped to R/S
character. Fixed-precision (non-IEEE) numbers are mapped into either doubles (
numeric) or long (
integer). Dates are mapped to character using the appropriate
TO_CHAR function in the RDBMS (which should take care of any locale information). Some RDBMS support the type
MONEY which should be mapped to
numeric. Large objects (character, binary, file, etc.) also need to be mapped. User-defined functions may be specified to do the actual conversion as follows:
run the query (either with
> rs <- dbExecStatement(con, "select whatever-You-need")
extract the output field definitions
> flds <- getFields(rs)
replace the class generator in the, say 3rd field, by the user own generator:
> flds[3, "Sclass"] # default mapping  "character"
> flds[3, "Sclass"] <- "myOwnGeneratorFunction"
set the new data mapping prior to fetching
> setDataMappings(resutlSet, flds)
fetch the rows and store in a
> data <- fetch(resultSet)
We may need to provide some additional utilities, for instance to convert dates, to escape characters such as quotes and slashes in query strings, to strip excessive blanks from some character fields, etc. We need to decide whether we provide hooks so these conversions are done at the C level, or do all the post-processing in R or S.
Another issue is what kind of data object is the output of an SQL query. Currently the MySQL and Oracle implementations return data as a
data.frame; data frames have the slight inconvenience that they automatically re-label the fields according to R/S syntax, changing the actual RDBMS labels of the variables; the issue of non-numeric data being coerced into factors automatically “at the drop of a hat” (as someone in s-news wrote) is also annoying.
The execution of SQL scripts is not fully described. The method that executes scripts could run individual statements without returning until it encounters a query (
SELECT-like) statement. At that point it could return that one result set. The application is then responsible for fetching these rows, and then for invoking
dbNextResultSet on the opened
dbConnection object to repeat the
fetch loop until it encounters the next
dbResultSet. And so on. Another (potentially very expensive) alternative would be to run all statements sequentially and return a list of
data.frames, each element of the list storing the result of each statement.
Binary objects and large objects present some challenges both to R and S. It is becoming more common to store images, sounds, and other data types as binary objects in RDBMS, some of which can be in principle quite large. The SQL-92 ANSI standard allows up to 2 gigabytes for some of these objects. We need to carefully plan how to deal with binary objects – perhaps tentatively not in full generality. Large objects could be fetched by repeatedly invoking a specified R/S function that takes as argument chunks of a specified number of raw bytes. In the case of S4 (and Splus5.x) the RS-DBI implementation can write into an opened connection for which the user has defined a reader (but can we guarantee that we won’t overflow the connection?). In the case of R it is not clear what data type binary large objects (BLOB) should be mapped into.
These are some of the limitations of the current interface definition:
we only allow one SQL statement at a time, forcing users to split SQL scripts into individual statements;
transaction management is not fully described;
the interface is heavily biased towards queries, as opposed to general purpose database development. In particular we made no attempt to define “bind variables”; this is a mechanism by which the contents of S objects are implicitly moved to the database during SQL execution. For instance, the following embedded SQL statement
/* SQL */ SELECT * from emp_table where emp_id = :sampleEmployee
would take the vector
sampleEmployee and iterate over each of its elements to get the result. Perhaps RS-DBI could at some point in the future implement this feature.
The high-level, front-end description of RS-DBI is the more critical aspect of the interface. Details on how to actually implement this interface may change over time. The approach described in this document based on one back-end driver per RDBMS is reasonable, but not the only approach – we simply felt that a simpler approach based on well-understood and self-contained tools (R, S, and C API’s) would be a better start. Nevertheless we want to briefly mention a few alternatives that we considered and tentatively decided against, but may quite possibly re-visit in the near future.
The ODBC protocol was developed by Microsoft to allow connectivity among C/C++ applications and RDBMS. As you would expect, originally implementations of the ODBC were only available under Windows environments. There are various effort to create a Unix implementation (see the Unix ODBC web-site and Harvey (1999)). This approach looks promising because it allows us to write only one back-end, instead of one per RDBMS. Since most RDBMS already provide ODBC drivers, this could greatly simplify development. Unfortunately, the Unix implementation of ODBC was not mature enough at the time we looked at it, a situation we expect will change in the next year or so. At that point we will need to re-evaluate it to make sure that such an ODBC interface does not penalize the interface in terms of performance, ease of use, portability among the various Unix versions, etc.
Another protocol, the Java database connectivity, is very well-done and supported by just about every RDBMS. The issue with JDBC is that as of today neither S nor R (which are written in C) interfaces cleanly with Java. There are several efforts (some in a quite fairly advanced state) to allow S and R to invoke Java methods. Once this interface is widely available in Splus5x and R we will need to re-visit this issue again and study the performance, usability, etc., of JDBC as a common back-end to the RS-DBI.
Yet another approach is to move the interface to RDBMS out of R and S altogether into a separate system or server that would serve as a proxy between R/S and databases. The communication to this middle-layer proxy could be done through CORBA Siegel (1996), Java’s RMI, or some other similar technology. Such a design could be very flexible, but the CORBA facilities both in R and S are not widely available yet, and we do not know whether this will be made available to Splus5 users from MathSoft. Also, my experience with this technology is rather limited.
On the other hand, this 3-tier architecture seem to offer the most flexibility to cope with very large distributed databases, not necessarily relational.
The latest documentation and software on the RS-DBI was available at www.omegahat.net (link dead now:
https://www.omegahat.net/contrib/RS-DBI/index.html). The R community has developed interfaces to some databases: RmSQL is an interface to the mSQL database written by Torsten Hothorn; RPgSQL is an interface to PostgreSQL and was written by Timothy H. Keitt; RODBC is an interface to ODBC, and it was written by Michael Lapsley. (For more details on all these see R Data Import/Export (2001).)
The are R and S-Plus interfaces to MySQL that follow the propose RS-DBI API described here; also, there’s an S-Plus interface SOracle James (In preparation) to Oracle (we expect to have an R implementation soon.)
The idea of a common interface to databases has been successfully implemented in Java’s Database Connectivity (JDBC) (www.javasoft.com), in C through the Open Database Connectivity (ODBC) (www.unixodbc.org), in Python’s Database Application Programming Interface (www.python.org), and in Perl’s Database Interface (www.cpan.org).
The R/S database interface came about from suggestions, comments, and discussions with John M. Chambers and Duncan Temple Lang in the context of the Omega Project for Statistical Computing. Doug Bates and Saikat DebRoy ported (and greatly improved) the first MySQL implementation to R.
The following code is meant to serve as a detailed description of the R/S to database interface. We decided to use S4 (instead of R or S version 3) because its clean syntax help us to describe easily the classes and methods that form the RS-DBI, and also to convey the inter-class relationships.
## Define all the classes and methods to be used by an ## implementation of the RS-DataBase Interface. Mostly, ## these classes are virtual and each driver should extend ## them to provide the actual implementation. ## Class: dbManager ## This class identifies the DataBase Management System ## (Oracle, MySQL, Informix, PostgreSQL, etc.) setClass("dbManager", VIRTUAL) setGeneric("load", def = function(dbMgr,...) standardGeneric("load") )setGeneric("unload", def = function(dbMgr,...) standardGeneric("unload") )setGeneric("getVersion", def = function(dbMgr,...) standardGeneric("getVersion") ) ## Class: dbConnections ## This class captures a connection to a database instance. setClass("dbConnection", VIRTUAL) setGeneric("dbConnection", def = function(dbMgr, ...) standardGeneric("dbConnection") )setGeneric("dbConnect", def = function(dbMgr, ...) standardGeneric("dbConnect") )setGeneric("dbExecStatement", def = function(con, statement, ...) standardGeneric("dbExecStatement") )setGeneric("dbExec", def = function(con, statement, ...) standardGeneric("dbExec") )setGeneric("getResultSet", def = function(con, ..) standardGeneric("getResultSet") )setGeneric("commit", def = function(con, ...) standardGeneric("commit") )setGeneric("rollback", def = function(con, ...) standardGeneric("rollback") )setGeneric("callProc", def = function(con, ...) standardGeneric("callProc") )setMethod("close", signature = list(con="dbConnection", type="missing"), def = function(con, type) NULL ) ## Class: dbResult ## This is a base class for arbitrary results from the RDBMS ## (INSERT, UPDATE, DELETE). SELECTs (and SELECT-like) ## statements produce "dbResultSet" objects, which extend ## dbResult. setClass("dbResult", VIRTUAL) setMethod("close", signature = list(con="dbResult", type="missing"), def = function(con, type) NULL ) ## Class: dbResultSet ## Note that we define a resultSet as the result of a ## SELECT SQL statement. setClass("dbResultSet", "dbResult") setGeneric("fetch", def = function(resultSet,n,...) standardGeneric("fetch") )setGeneric("hasCompleted", def = function(object, ...) standardGeneric("hasCompleted") )setGeneric("getException", def = function(object, ...) standardGeneric("getException") )setGeneric("getDBconnection", def = function(object, ...) standardGeneric("getDBconnection") )setGeneric("setDataMappings", def = function(resultSet, ...) standardGeneric("setDataMappings") )setGeneric("getFields", def = function(object, table, dbname, ...) standardGeneric("getFields") )setGeneric("getStatement", def = function(object, ...) standardGeneric("getStatement") )setGeneric("getRowsAffected", def = function(object, ...) standardGeneric("getRowsAffected") )setGeneric("getRowCount", def = function(object, ...) standardGeneric("getRowCount") )setGeneric("getNullOk", def = function(object, ...) standardGeneric("getNullOk") ) ## Meta-data: setGeneric("getInfo", def = function(object, ...) standardGeneric("getInfo") )setGeneric("describe", def = function(object, verbose=F, ...) standardGeneric("describe") )setGeneric("getCurrentDatabase", def = function(object, ...) standardGeneric("getCurrentDatabase") )setGeneric("getDatabases", def = function(object, ...) standardGeneric("getDatabases") )setGeneric("getTables", def = function(object, dbname, ...) standardGeneric("getTables") )setGeneric("getTableFields", def = function(object, table, dbname, ...) standardGeneric("getTableFields") )setGeneric("getTableIndices", def = function(object, table, dbname, ...) standardGeneric("getTableIndices") )
A virtual class allows us to group classes that share some common functionality, e.g., the virtual class “
dbConnection” groups all the connection implementations by Informix, Ingres, DB/2, Oracle, etc. Although the details will vary from one RDBMS to another, the defining characteristic of these objects is what a virtual class captures. R and S version 3 do not explicitly define virtual classes, but they can easily implement the idea through inheritance.↩︎
The term “database” is sometimes (confusingly) used both to denote the RDBMS, such as Oracle, MySQL, and also to denote a particular database instance under a RDBMS, such as “opto” or “sales” databases under the same RDBMS.↩︎