track.design: Design of a tracking environment

Description

This document describes the layout of a tracking environment. Object tracking works by replacing a variable with an active binding, and keeping the actual value of the variable on disk and/or in another environment. Tracked objects are automatically resaved to disk when they are changed. Basic characteristics, such as class, size, extent, and creation and modification times are recorded in a summary of all tracked objects.

Arguments

Terminology

One could describe a tracking environment as "attached" to the tracked environment, but that using that term would risk confusion with the role of the attach() function and search path in R. So, instead the track package says that a tracking environment is "linked" to the tracked environment.

Untrackable variables -- reserved names

Only ordinary variables can be tracked -- variables that are active bindings cannot be tracked. Several variable names are reserved and cannot be tracked: .trackingEnv, .trackingFileMap, .trackingUnsaved, .trackingSummary, .trackingSummaryChanged, .trackingOptions. Additionally, any variable with a newline character ("\n") as part of its name cannot be tracked (the main reason for this is that the mapping from object names to file names is stored in a text file, and newline character delimits the name).

The file map

The mapping from object names to file names is stored in the file fileMap.txt. This data is stored as ordinary text file to make it easy for users to see the object-file mappings outside of R.

Implementation considerations

The reason that objects must be explicitly registered for tracking is that there is currently no way of setting up a function to be called when a new object is created, so new objects are always created as ordinary R objects. Similarly, the R remove() functions does not have any hooks, so if remove() is called on a tracked variable, it will just remove the active binding in the visible environment, but will not disturb the underlying tracking environment. The track.remove() function will completely remove a tracked variable from the visible environment and the underlying tracking environment (including deleting an associated disk file.) Object tracking was intended to be used in situations where large numbers of large objects must be manipulated. Consequently, there is a good chance of exhausting resources while using the track package. The track code tries to check return codes when creating objects or writing files, and in cases where it is unable to complete an operation it tries leave the tracking environment in a state from which objects can be salvaged. The functions track.rebuild() and track.flush() are provided to help recover from situations where resource limitations prevented successful operation. Note that files are generally written in a "unsafe" manner (i.e., existing files can be overwritten with partial new files), but in these cases data is retained in the memory and can be rewritten after resolving file system problems. The R functions exists() should be used with care on tracked objects, because it will actually fetch the object, possibly needing to read it from disk. In the track code, the exists("x") function is not used to check existence of a possibly tracked object x, instead an idiom like is.element("x", objects(all=TRUE)) is used. These statements about the available facilities in R were true as of R-2.4.1 (released Dec 2006). The rules for how variable names are mapped to file names are based on trying to use filenames that will work properly on all three operating systems R works on (Linux, Windows, and Mac OS X). A somewhat obscure point that must be taken into account is the case-insensitivity of Mac OS X and Windows. Even though modern versions of the OS's seem to use case in their file names, this is because they are case preserving, but they are in fact still case insensitive. This means that a file created with the name "X.rda" is the same file as the "x.rda". Here is a short shell transcript showing this behavior in a bash shell running under Windows and Mac OS X (it's the same in both).

    $ echo 123 > X
    $ cat x
    123
    $ echo 456 > x
    $ cat x
    456
    $ cat X
    456

Thus, in order to work on OS's, file mapping must be used to create different filenames for the R objects "x" and "X" (which are in fact different in R.)

Portability

Tracking directories are intended to be operating-system independent and completely portable across different operating systems.

Compression

Saved R objects are compressed by default in R and by the track package. Decompression speed is very important for interactive response when using track, because each time an object is accessed, it is read from its file (unless the object is cached). Of the compression algorithms available as of R-2.12.0, which are gzip, bzip2, and xz, gzip is the winner in terms of speed. The default compression level in R for gzip is 6, but level 1 gives faster compression with slightly larger files (though decompression is not faster). The lzop compression algorithm http://www.lzop.org is still faster but it is not yet available in R. Here are some comparisons and benchmarks of various compression programs:

Compression/decompression is nicely handled in R: only the call to save() has arguments for compression. Decompression in load() is handled automatically using a standard code (magic) at the start of the saved file. Saved files can also be compressed or decompressed outside of R, and load() will still handle them correctly, provided the compression used is one of the types that R knows about.

Details

Object tracking works by replacing a variable with an active binding, and keeping the actual value of the variable on disk and/or in another environment. Whenever the variable is fetched or assigned, the active binding is called, and it writes the object to disk if necessary, and records basic characteristics of the objects in a summary of all objects, including creation, modification and access times.

A tracking environment can be linked to one environment on the search path, but the tracking environment is not on the search path itself. An environment on the search path can only have one tracking environment linked to it. In standard use, variables are tracked automatically by a task callback function. Alternatively, variables to track can be registered with the tracking environment using the function track().

Any user-created environment on the search path, or the global environment, can be tracked.

The format used to store R objects in files is the one used by save()/load() -- the objects in those files can be read using load() if desired.

The various variables and files involved in tracking are as follows (assuming the RData suffix being used is "rda"). Note that the default tracked visible environment is the global environment.

Tracked Visible Environment (on search list) Tracking Environment (not on search list) +-----------------+ +-------------------------+ | .trackingEnv|-->| .trackingDir |---> Tracking Directory (files) | | | | | | x (*) | | x (@) | +- x.rda | abc (*) | | abc (@) | +- abc.rda | Y (*) | | Y (@) | +- _1.rda | x1 | | | | | x2 | | | | | | | .trackingFileMap | +- filemap.txt | | | .trackingSummary | +- .trackingSummary.rda | | | .trackingUnsaved | | | | .trackingSummaryChanged | | | | .trackingOptions | +-----------------+ +-------------------------+

variables marked (*) are tracked and are actually an active binding that refers to the corresponding variable in the tracking environment. There can also be untracked variables in the visible tracked environment, but in the standard mode of operation these are detected by the end-of-task callback function and are immediately converted to tracked variables (except for variables with reserved names like .trackingSummary, and variables matching exclude patterns, see options autoTrackExcludePattern and autoTrackExcludeClass in track.options.
variables marked (@) may or may not exist -- if they do not exist in the tracking environment, they will be automatically read from file when the corresponding tracked object is accessed.
The "trackingEnv" attribute on the tracked environment is the tracking environment. This is implemented as an attribute on the tracked environment rather than as a variable in the tracked environment so that save.image() on the tracked environment will ignore the tracking environment. If the tracking environment were stored as a variable in the tracked environment, save.image() could end up storing two copies of every tracked variable: one when it accessed the active binding (it stores a copy of the object: save() doesn't know it's an active binding); and another if the object is cached in the tracking environment.
The "trackingDir" attribute on the tracking environment specifies the absolute pathname of the directory under which tracked objects are stored on file. It uses the absolute pathname because the current directory of the R session can be changed using setwd(), which would result in losing a relative pathname.
.trackingFileMap stores the base part of the file name corresponding to each tracked object as a named character vector (the names on the vector are the object names). Objects that do not have simple names have an associated file name like "\_NNN" where "NNN" is a number. For example, the .trackingFileMap for the above configuration could be c(abc="abc", x="x", Y="_1"). Simple object names are those conforming to the following rules:
- less than 55 characters
- are comprised of only lower-case letters, digits 0 through 9, "." and "\_"
- begin with a lower-case letter
- are not one of the following: con, prn, aux, nul, com1 through com9, lpt1 through lpt9, and do not begin with one of these names followed by a period (i.e., prn.foo and prn.foo.bar are both not simple names) (these are special file names under Microsoft Windows - see http://en.wikipedia.org/wiki/Filename and search the web on the keywords "windows short file names rules prn com" to find an official Microsoft site.)

The object .trackingFileMap is always kept in memory and is always saved to disk (as text in the file filemap.txt) whenever it is changed.

.trackingSummary is a data frame recording various basic characteristics of the tracked objects, such as class, size and extent, and also times of creation, and most recent modification and access. The tracking summary should be accessed using the function track.summary().

.trackingSummaryChanged A logical flag indicating whether or not the tracking summary copy on disk is in sync with version in memory. To reduce overhead on accessing objects, there is an option to not resave the tracking summary when it is changed on accessing an object -- this variable indicates if it has been changed.

.trackingUnsaved: If the tracking options are set up so that objects are not automatically written to files on assignment, this variable contains a vector of names of all objects that have not been saved.

.trackingOptions: are accessed and changed by the track.options() function. They are kept in memory, and also written to disk whenever they are changed. The tracking directory is organized as an R package. It's layout is as follows (saying, for example, that attr(trackingEnv, "trackingDir") is /tmp/trackdir1, and .trackingFileMap is c(abc="abc", x="x", Y="_1")):

/tmp/trackdir1
      |
      +- filemap.txt
      +- .trackingSummary.rda
      +- x.rda
      +- abc.rda
      +- _1.rda

References

Roger D. Peng. Interacting with data using the filehash package. R News, 6(4):19-24, October 2006. http://cran.r-project.org/doc/Rnews.

David E. Brahm. Delayed data packages. R News, 2(3):11-12, December 2002. http://cran.r-project.org/doc/Rnews