Managing Data Sources with srcr

Charles Bailey baileyc@email.chop.edu

2020-12-13

Setting up connections

The dplyr package abstracts away many aspects of database interaction. However, code that uses dbplyr must construct an appropriate database connection, and use it to create any tbl instances that are used to access data. This creates two problems for code that one might want to distribute. First, the connection-building dbConnect or src_whatever statement typically contains credentials used to authenticate to the database server. This creates a security risk, as sending around personal credentials isn’t a good idea for a variety of reasons, and using the same “service credentials” for many users makes tracking usage more difficult. It also makes it harder to reuse code, since it has to be edited to accommodate different users or different databases. Second, the connection info will often include other parameters, such as schema, data transfer settings, or the like. Users have adopted a number of solutions to these problems, such as reading environment variables or sourcing a file in a known location. Each works well in some circumstances, but runs into problems with scope or logistics in others.

The package provides an alternative in this area that aims to be easy enough to use for straightforward cases, and flexible enough to adapt to complex projects and environments. Its workhorse is the srcr() function. As its documentation explains, srcr() is an adapter that lets you create a data source of a type known to DBI or old-style dplyr connection functions, using configuration data passed to srcr(), or, more often, supplied in a configuration file.

Configuration structure

Establishing the connection

Configuration files provide a simple way to represent in JSON the information needed to construct a data source. The JSON must define a single object (or hash), which is translated into a list structure within R. Two keys from the object define the database connection.

src_name
either the name of a driver function compatible with the DBI specification, such as SQLite or PostgreSQL (N.B. the initial R in the package name is not included), or the name of an old-style dplyr function used to construct a data source, typically a database connection, such as src_postgres or src_mysql.
src_args
a list, where the keys are names of arguments to the constructor, and the corresponding elements are the argument values.

Here’s a typical example:

{
    "src_name" : "Postgres",
    "src_args" : {
          "host"     : "my.database.server",
            "port"     : 5432,
            "dbname"   : "project_db",
            "username" : "my_credential",
            "password" : "DontLook",
            "options"  : "-c search_path=schema0,schema1,schema2"
     }
}

If you’re deriving the configuration information programatically, you can pass it directly to srcr() via the config argument, but it’s perhaps more common that configuration remains the same across different connections for a given project. For these cases, srcr encourages separation of configuration from code.

Post-connection setup

Once the connection is established, there may be additional work to do. For example, session settings may need to be altered, or schemas to search specified, or authorization roles changed. There are two options that srcr provides to address this:

post_connect_sql
This option allows you to pass a series of SQL statements to the newly-established database session for execution. In this way, you can change database session settings to match the intended use of the connection. While these statements can make any changes the database server will permit, they cannot alter the R environment directly.
post_connect_fun
This option gives you the most freedom to make changes. It lets you write an R function that takes the newly-established database connection as its single parameter. It may perform computation in the database, change connection settings, or even replace the connection. The value it returns will be passed back as the return value of srcr().

In both cases, this makes it possible to execute additional code specified in the configuration file. Since this creates the possibility that unknown code may be executed and produce unwanted effects, you need to opt in to each, using the allow_post_connect parameter to srcr().

Finding configuration data

The srcr package tries to provide you with a lot of flexibility in the way you deploy your code and configuration, by letting you get the latter to srcr() in a variety of ways. The first option that returns valid JSON is used, and later options aren’t checked; srcr() does not try to merge data from more than one source.

Telling srcr() what to read directly

If you know where your configuration data lives, you can point srcr() directly to it, using the paths argument. This is a vector of paths to check, so you can provide a series of places to look, and srcr() will use the first one it finds. Each place can be a path to a local file, or a URL that returns JSON. (As an implementation detail, since srcr() uses jsonlite::fromJSON under the hood, paths can also contain a JSON string rather than a pointer to outside configuration data. We don’t make any promises about this, as jsonlite::fromJSON might change someday, but it can be a handy way to provide fallback configuration information after having srcr() check for an outside resource.)

Searching for configuration files

Through the find_config_files() function (implicitly called by srcr() if there’s no paths or config provided), tries to support a number of common deployment styles through its use of default search locations for configuration files. For those who prefer per-user config files, it will look in your home directory. If you prefer to deploy configuration data with your application, you can put the configuration file in the same directory as your main application program. Finally, you can put the configuration file in the same directory as library code that calls srcr() either directly or through one intermediate call. Note that the current working directory isn’t in this search path by default; this is a measure intended to minimize the harm someone can cause by invoking your program in a working directory that might subvert the intended connection, either intentionally or through accidental use of a common configuration file name.

Similarly, srcr() will try by default to find files with the same basename as your application, or as the library file(s) making the call to srcr(). Optionally, the file can have a “type” (i.e. suffix) of .json or .conf, or none at all.

If these options don’t suit your deployment strategy, you can provide explicit hints to srcr() using the dirs, basenames, and suffices arguments.

Finally, to accommodate convention on Unix-like systems, find_config_files() first checks for a “hidden” file with a leading . before checking for the plain basename.

It’s worth noting that srcr() expects any configuration files it finds to contain JSON, regardless of the file suffix. This was done to minimize the number of other packages that are prerequisites for . If you prefer another format, such as YAML or Apache, you can get there pretty easily (example simplfied a bit):

my_srcr <- function(...) {
    files <- find_config_files(...)
    mydata <- read_my_config_format(files)
    srcr(config = mydata)
}

Using environment variables

If you need to specify where to look at runtime, you can use the environment variable BASENAME_CONFIG to point to a configuration file, where BASENAME is one of the basenames find_config_files() would usually check (see below). One note: srcr() will only pay attention to this environment variable if it points to an actual file, not a URL or JSON string. This is construed as a feature, in that it may limit the damage someone can inflict by fiddling with the environment. If you trust the environment, you can be more permissive by writing something like

my.paths <- c()
for (bn in my.basenames) {
    my.info <- Sys.getenv(paste0(toupper(bn), '_CONFIG'))
    if (my.info != '') my.paths <- c(my.paths, my.info)
}
src <- if (length(paths) > 0) srcr(paths = my.paths) else srcr(other.args)

Other uses

The srcr() function is intended to help manage database connections, and contains a few optimizations for this (such as turning src_name into a driver argument if it doesn’t start with ‘src_’). But at its root, it’s a function that takes a name and arguments from configuration data, executes them (and maybe some post-processing), and returns the result. You may find this pattern useful in other contexts. If so, you can implement your own function named “src_whatever”, and use srcr() to allow it to access configuration.

Enjoy!