SQL escaping/quoting proposal

Hi all,

The approach that DBI takes to escaping is sub-optimal: it tries to
figure out if an R variable name matches an SQL reserved word, and if
so munge it so that there's no longer a conflict. This creates a
situation where there are some identifiers that are valid in R, and
some that are valid in SQL and we have a complicated and bug prone
approach to converting between them.

Instead, I recommend taking an approach where identifiers (i.e. table
and field names) are always quoted using the appropriate database
syntax. This not only avoids any problems with SQL reserved words, but
it also ensures that every field name in R (even those containing
spaces and other special characters) can be used in SQL.

To achieve this change, I think we should to:

* deprecate `make.db.names()`, `isSQLKeyword()`, and `SQLKeywords()`
* add new generics `sqlQuoteString()` and `sqlQuoteIdentifier()`.

The new generics would be defined on the driver object, and would come
with default methods as follows:

```
setGeneric("sqlQuoteString", function(drv, x, ...) {
   standardGeneric("sqlQuoteString")
})
setMethod("sqlQuoteString", "DBIDriver", function(drv, x, ...) {
   x <- gsub('"', '""', x, fixed = TRUE)
   paste('"', x, '"', sep = "")
})

setGeneric("sqlQuoteIdentifer", function(drv, x, ...) {
   standardGeneric("sqlQuoteIdentifer")
})
setMethod("sqlQuoteString", "DBIDriver", function(drv, x, ...) {
   x <- gsub("'", "''", x, fixed = TRUE)
   paste("'", x, "'", sep = "")
})
```

Individual implementations would be encouraged to provide methods that
use the quoting functions provided by the client library, where
available.

Does anyone see any problems with this approach?

Hadley

Regarding (b), my reasoning is that when working with Postgres, I would want
to allow Postgres to do its normal lower casing of unquoted identifiers.
That is, `thisField` is a valid identifier, but passed in quoted case will
be preserved, which means always having to quote the identifier in the
future. Passed in quoted, Postgres will force it to `thisfield`, and if a
client requests `thisField` unquoted, the correct field will be returned.

This not only avoids any problems with SQL reserved words, but
it also ensures that every field name in R (even those containing
spaces and other special characters) can be used in SQL.

Hi all,

The approach that DBI takes to escaping is sub-optimal: it tries to
figure out if an R variable name matches an SQL reserved word, and if
so munge it so that there's no longer a conflict. This creates a
situation where there are some identifiers that are valid in R, and
some that are valid in SQL and we have a complicated and bug prone
approach to converting between them.

Instead, I recommend taking an approach where identifiers (i.e. table
and field names) are always quoted using the appropriate database
syntax.

it also ensures that every field name in R (even those containing
spaces and other special characters) can be used in SQL.

To achieve this change, I think we should to:

* deprecate `make.db.names()`, `isSQLKeyword()`, and `SQLKeywords()`
* add new generics `sqlQuoteString()` and `sqlQuoteIdentifier()`.

The new generics would be defined on the driver object, and would come
with default methods as follows:

```
setGeneric("sqlQuoteString", function(drv, x, ...) {
   standardGeneric("sqlQuoteString")
})
setMethod("sqlQuoteString", "DBIDriver", function(drv, x, ...) {
   x <- gsub('"', '""', x, fixed = TRUE)
   paste('"', x, '"', sep = "")
})

setGeneric("sqlQuoteIdentifer", function(drv, x, ...) {
   standardGeneric("sqlQuoteIdentifer")
})
setMethod("sqlQuoteString", "DBIDriver", function(drv, x, ...) {
   x <- gsub("'", "''", x, fixed = TRUE)
   paste("'", x, "'", sep = "")
})
```

Individual implementations would be encouraged to provide methods that
use the quoting functions provided by the client library, where
available.

Does anyone see any problems with this approach?

Hadley

Could you provide a bit more clarification on what the sqlQuoteIdentifier should do?
How shall we deal when a vector of strings is passed?
  Is it right to assume that sqlQuoteIdentifier(drv, c('a', 'b')) should
  return a vector consisting of quoted results of individual element of the vector?

How do we construct a reference to table with schema, or column with table?
eg schema.table or table.column?
  More specifically, is it right to assume that sqlQuoteIdentifier is used for constructing
  individual part of the composite identifier?

You had a minor mistake in showing the default method (The name is both "sqlQuoteString",
and I am not sure which is intended for sqlQuoteIdentifier).

Another consideration is for the name of the function. Whether we should
use sql prefix or use db prefix.  I would like to know what others think for this point.

Here, I think we can avoid problem in most cases, but there are still a bit
cases where the encoding does not allow proper conversion.
That's the problem of the database capability, there is not much things that
the driver can do, though.

I've cleaned up the examples and made them work and put the results in
https://gist.github.com/hadley/7057387 - that should make discussion a
bit more concrete.