basiliskStart: Start and stop 'basilisk'-related processes

In LTLA/jormungandR: Freezing Python Dependencies Inside Bioconductor Packages

Start and stop basilisk-related processes

Description

Creates a basilisk process in which Python operations (via reticulate) can be safely performed with the correct versions of Python packages.

Usage

basiliskStart(
  env,
  full.activation = NA,
  fork = getBasiliskFork(),
  shared = getBasiliskShared(),
  testload = NULL
)

basiliskStop(proc)

basiliskRun(
  proc = NULL,
  fun,
  ...,
  env,
  full.activation = NA,
  persist = FALSE,
  fork = getBasiliskFork(),
  shared = getBasiliskShared(),
  testload = NULL
)

Arguments

env	A BasiliskEnvironment object specifying the basilisk environment to use. Alternatively, a string specifying the path to an environment, though this should only be used for testing purposes. Alternatively, NULL to indicate that the base Conda installation should be used as the environment.
full.activation	Logical scalar, see activateEnvironment for more details.
fork	Logical scalar indicating whether forking should be performed on non-Windows systems, see getBasiliskFork. If FALSE, a new worker process is created using communication over sockets.
shared	Logical scalar indicating whether basiliskStart is allowed to load a shared Python instance into the current R process, see getBasiliskShared.
testload	Character vector specifying the Python packages to load into the process during set-up. This is used to check that packages can be correctly loaded, switching to a fallback on GLIBCXX dynamic linking failures.
proc	A process object generated by basiliskStart.
fun	A function to be executed in the basilisk process. This should return a “pure R” object, see details.
...	Further arguments to be passed to fun.
persist	Logical scalar indicating whether to pass a persistent store to fun. If TRUE, fun should accept a store argument.

Details

These functions ensure that any Python operations in fun will use the environment specified by envname. This avoids version conflicts in the presence of other Python instances or environments loaded by other packages or by the user. Thus, basilisk clients are not affected by (and if shared=FALSE, do not affect) the activity of other R packages.

It is good practice to call basiliskStop once computation is finished to terminate the process. Any Python-related operations between basiliskStart and basiliskStop should only occur via basiliskRun. Calling reticulate functions directly will have unpredictable consequences, Similarly, it would be unwise to interact with proc via any function other than the ones listed here.

If proc=NULL in basiliskRun, a process will be created and closed automatically. This may be convenient in functions where persistence is not required. Note that doing so requires specification of pkgname and envname.

If the base Conda installation provided with basilisk satisfies the requirements of the client package, developers can set env=NULL in this function to use that base installation rather than constructing a separate environment.

Value

basiliskStart returns a process object, the exact nature of which depends on fork and shared. This object should only be used in basiliskRun and basiliskStop.

basiliskRun returns the output of fun(...), possibly executed inside the separate process.

basiliskStop stops the process in proc.

Choice of process type

• If shared=TRUE and no Python version has already been loaded, basiliskStart will load Python directly into the R session from the specified environment. Similarly, if the existing environment is the same as the requested environment, basiliskStart will use that directly. This mode is most efficient as it avoids creating any new processes, but the use of a shared Python configuration may prevent non-basilisk packages from working correctly in the same session.

• If fork=TRUE, no Python version has already been loaded and we are not on Windows, basiliskStart will create a new process by forking. In the forked process, basiliskStart will load the specified environment for operations in Python. This is less efficient as it needs to create a new process but it avoids forcing a Python configuration on other packages in the same R session.

• Otherwise, basiliskStart will create a parallel socket process containing a separate R session. In the new process, basiliskStart will load the specified environment for Python operations. This is the least efficient as it needs to transfer data over sockets but is guaranteed to work.

Developers can control these choices directly by explicitly specifying shared and fork, while users can control them indirectly with setBasiliskFork and related functions.

Testing package loads

If testload is provided, basiliskStart will attempt to load those Python packages into the newly created process. This is used to detect loading failures due to differences in the versions of the shared libraries. Most typically, a conda-supplied Python package (often scipy submodules) will have been compiled against a certain version of libstdc++ but R is compiled against an older version. R's version takes precedence when reticulate attempts to load the Python package, causing cryptic “GLIBCXX version not found” errors.

By checking the specified testload, basiliskStart can check for loading failures in potentially problematic packages. Upon any failure, basiliskStart will fall back to a separate socket process running a conda-supplied R installation. The idea is that, if both Python and R are sourced from conda, they will be using the same version of libstdc++ and other libraries. This avoids loading errors and/or segmentation faults due to version mismatches.

Use of this "last resort fallback" overrides any choice of process type from fork and shared. If no failures are encountered, a process will be created using the current R installation.

Note that the fallback R installation is very minimalistic; only reticulate is guaranteed to be available. This places some limitations on the code that can be executed inside fun for basilisk environments that might trigger use of the fallback.

Constraints on user-defined functions

In basiliskRun, there is no guarantee that fun has access to basiliskRun's calling environment. This has several consequences for code in the body of fun:

• Variables used inside fun should be explicitly passed as an argument to fun. Developers should not rely on closures to capture variables in the calling environment of basiliskRun.

• Developers should not attempt to pass complex objects to memory in or out of fun. This mostly refers to objects that contain custom pointers to memory, e.g., file handles, pointers to reticulate objects. Both the arguments and return values of fun should be pure R objects.

• Functions or variables from non-base R packages should be prefixed with the package name via ::, or those packages should be reloaded inside fun. However, if fun loads Python packages that might trigger the last resort fallback, no functions or variables should be used from non-base R packages.

Developers can test that their function behaves correctly in basiliskRun by setting setBasiliskShared and setBasiliskFork to FALSE. This forces the execution of fun in a new process; any incorrect assumption of shared environments will cause errors. If fun involves fallback-inducing Python packages, developers can further set setBasiliskForceFallback before running basiliskRun. This tests that fun works with the minimal conda-supplied R installation.

Persisting objects across calls

Objects created inside fun can be persisted across calls to basiliskRun by setting persist=TRUE. This will instruct basiliskRun to pass a store argument to fun that can be used to store arbitrary objects. Those same objects can be retrieved from store in later calls to basiliskRun using the same proc. Any object can be stored in .basilisk.store but will remain strictly internal to proc.

This capability is primarily useful when a Python workflow is split across multiple basiliskRun calls. Each subsequent call can pick up from temporary intermediate objects generated by the previous call. In this manner, basilisk enables modular function design where developers can easily mix and match different basiliskRun invocations. See Examples for a working demonstration.

Use of lazy installation

If the specified basilisk environment is not present and env is a BasiliskEnvironment object, the environment will be created upon first use of basiliskStart. If the base Conda installation is not present, it will also be installed upon first use of basiliskStart. We do not provide Conda with the basilisk package binaries to avoid portability problems with hard-coded paths (as well as potential licensing issues from redistribution).

By default, both the base conda installation and the environments will be placed in an external user-writable directory defined by rappdirs via getExternalDir. The location of this directory can be changed by setting the BASILISK_EXTERNAL_DIR environment variable to the desired path. This may occasionally be necessary if the file path to the default location is too long for Windows, or if the default path has spaces that break the Miniconda/Anaconda installer.

Advanced users may consider setting the environment variable BASILISK_USE_SYSTEM_DIR to 1 when installing basilisk and its client packages from source. This will place both the base installation and the environments in the R system directory, which simplifies permission management and avoids duplication in enterprise settings.

Persistence of environment variables

When shared=TRUE and if no Python instance has already been loaded into the current R session, a side-effect of basiliskStart is that it will modify a number of environment variables. This is done to mimic activation of the Conda environment located at env. Importantly, old values for these variables will not be restored upon basiliskStop.

This behavior is intentional as (i) the correct use of the Conda-derived Python depends on activation and (ii) the loaded Python persists for the entire R session. It may not be safe to reset the environment variables and “deactivate” the environment while the Conda-derived Python instance is effectively still in use. (In practice, lack of activation is most problematic on Windows due to its dependence on correct PATH specification for dynamic linking.)

If persistence is not desirable, users should set shared=FALSE via setBasiliskShared. This will limit any modifications to the environment variables to a separate R process.

Author(s)

Aaron Lun

See Also

setupBasiliskEnv, to set up the conda environments.

activateEnvironment in the basilisk.utils package.

getBasiliskFork and getBasiliskShared, to control various global options.

Examples

# Creating an environment (note, this is not necessary
# when supplying a BasiliskEnvironment to basiliskStart):
tmploc <- file.path(tempdir(), "my_package_A")
if (!file.exists(tmploc)) {
    setupBasiliskEnv(tmploc, c(pandas_spec("2.1.4")))
}

# Pulling out the pandas version, as a demonstration:
cl <- basiliskStart(tmploc, testload="pandas")
basiliskRun(proc=cl, function() { 
    X <- reticulate::import("pandas"); X$`__version__` 
})
basiliskStop(cl)

# This happily co-exists with our other environment:
tmploc2 <- file.path(tempdir(), "my_package_B")
if (!file.exists(tmploc2)) {
    setupBasiliskEnv(tmploc2, c(pandas_spec("2.1.3")))
}

cl2 <- basiliskStart(tmploc2, testload="pandas")
basiliskRun(proc=cl2, function() { 
    X <- reticulate::import("pandas"); X$`__version__` 
})
basiliskStop(cl2)

# Persistence of variables is possible within a Start/Stop pair.
cl <- basiliskStart(tmploc)
basiliskRun(proc=cl, function(store) {
    store$snake.in.my.shoes <- 1
    invisible(NULL)
}, persist=TRUE)
basiliskRun(proc=cl, function(store) {
    return(store$snake.in.my.shoes)
}, persist=TRUE)
basiliskStop(cl)

LTLA/jormungandR documentation built on Feb. 6, 2024, 2:29 p.m.