Changes in version 1.8.0
o Added createDedupSession(), addObjectToDedupSession() and
checkObjectInDedupSession() utilities, that allow saveObject()
method developers to check for duplicated objects and avoid
re-saving them. Most of this functionality was migrated and
generalized from the alabaster.matrix package.
o Added the cloneDirectory() function to easily clone the
contents of an existing directory, either by copying or linking
to the files. This should be used for creating lightweight
on-disk representations of duplicated objects.
o Added the absolutizePath() function to obtain an absolute file
path. This is occasionally necessary to ensure that references
to deduplicated resources are robust to changes to the working
directory.
o Coerce numeric_version instances to strings or character
vectors before saving them in saveObject(). This includ such
objects encountered inside lists or DataFrames.
o Added getSaveEnvironment() and related functions to keep track
of the R environment used to save each object. This facilitates
debugging and enables corrective actions to be taken by loading
functions when encountering files created by buggy versions of
any package in the alabaster stack.
o Added preliminary support for custom variable length string
(VLS) arrays, which compress the heap for more efficient
storage than HDF5's VLS implementation. This is enabled in the
saveObject() methods for standalone character vectors as well
as character vectors in DataFrames and lists via the relevant
character.vls= option. Setting character.vls=NULL will
automatically switch between fixed-length string datasets and
VLS arrays based on the estimated storage. (Defaults to FALSE
for back-compatibility.)
Changes in version 1.6.0
o Distinguish between scalars and length-1 vectors when
saving/loading lists. This effectively unboxes all length-1
vectors in a list, by default; this is probably the more
reasonable expectation for other languages that have a concept
of scalars. Users can override this by calling I() on elements
that they want to keep as length-1 vectors, in the same manner
as jsonlite.
o Streamlined the definition of the Rfc3339 class so that it
behaves better with I().
o Normalize paths to resolve ~ prior to calling C++ code.
o Open HDF5 files in read-only mode to avoid permission-related
problems for readObject()-dispatched functions.
o Store numbers at maximum precision when saving lists in the
JSON format via saveObject().
o Added registerValidateObjectSatisfiesInterface() and
registerValidateObjectDerivedFrom(), to allow developers to
declare that custom subclasses satisfy an interface or have an
inheritance relationship, respectively.
o Updated validateDirectory() so that it works with a directory
of objects saved via saveObject(). Objects saved under the old
regime (i.e., stageObject()) are auto-detected but can also be
explicitly validated by setting legacy=FALSE.
o Added a data.frame method for saveObject(), to avoid fallback
to the list method.
Changes in version 1.4.0
o Added the saveObject and readObject functions, which comply
with the newly formalized takane specifications for each
object's on-disk representation. This enables greater
consistency across language frameworks (e.g., Python) by using
a common library for enforcing the specification. It also
eliminates the need for a separate metadata file, simplifying
the user experience and the expectations on extension
developers.
o Added a dedicated validateObject function, which uses the
takane libary to validate each object's files. This allows
users to easily check if a set of files is compliant with the
specification, e.g., when modifying parts of the object outside
of saveObject.
o Soft-deprecation of the stageObject and loadObject functions.
These use legacy representations that do not comply with the
tightened takane specifications.