The purpose of this package is to generate verification trees and evaluations of user supplied statements. Trees are made by parsing a statement into a data structure composed of lists. Safe statement evaluations are done by executing the verification trees.
Any statement can be represented by a tree data structure. Here is a quote from Wikipedia to explain the concept:
In computer science, a tree is a widely used abstract data type that represents a hierarchical tree structure with a set of connected nodes. Each node in the tree can be connected to many children (depending on the type of tree), but must be connected to exactly one parent, except for the root node, which has no parent. These constraints mean there are no cycles or “loops” (no node can be its own ancestor), and also that each child can be treated like the root node of its own subtree, making recursion a useful technique for tree traversal. In contrast to linear data structures, many trees cannot be represented by relationships between neighboring nodes in a single straight line.
By default the package will know how to parse a R statement into a tree. In theory, you could supply your own tokens and use the package to parse any language that follows similar grammar.
These verification trees have been used to port R statements into other languages. For example, one use case is to write out formulas in excel that replicate the calculations done in R.
Writing code that executes unverified code can be both powerful and dangerous. A common approach to doing this is with the following pattern:
eval(parse(text="unverified_code"))
The power comes from the flexibility this pattern gives us. It usually comes with a significant performance cost, but CPU time is cheaper than our developers.
The danger comes from unknown unknowns. Your own input on your personal computer may not pose much of a risk. The same does not hold for input from nefarious/clever users on a server.
Execution of a verification tree generated on unverified code is a different story. The risk is limited to what is deemed acceptable based on the supplied tokens. It comes with a greater performance cost, but removes the danger from unverified code.
The focus of this package is limited to just creating and evaluating trees. The next sections cover the main functions.
The create_tree
function takes a string and generates a
tree. For example, we can parse the simple expression
2+3
:
<- evalR::create_tree("2+3")
x str(x)
#> List of 2
#> $ pval: list()
#> $ eval:List of 3
#> ..$ : chr "+"
#> ..$ :List of 2
#> .. ..$ : chr "atomic"
#> .. ..$ : chr "2"
#> ..$ :List of 2
#> .. ..$ : chr "atomic"
#> .. ..$ : chr "3"
We can see the structure is a list of lists.
You don’t need to understand the structure of the tree to use it. Just like you can drive a car without knowing how an engine works. This section will help reveal how the trees are formed.
First lets confirm that we can replicate the tree structure:
<- list(
tree pval = list(),
eval = list("+", list("atomic", "2"), list("atomic", "3"))
)::expect_equal(x,tree) testthat
This test passes with zero error.
A full tree is made up of two main branches:
pval
- stands for parenthesis valueseval
- stands for verification valuesThe first thing the function does is find all parenthesis blocks and
treats them as sub statements. Each of these sub statements becomes an
element of the pval
element. The eval
tree
will have references to these pval
entries.
For example, lets tweak the 2+3
to
(2)+(3)
:
<- evalR::create_tree("(2)+(3)")
x str(x)
#> List of 2
#> $ pval:List of 2
#> ..$ \0:List of 2
#> .. ..$ pval: list()
#> .. ..$ eval:List of 2
#> .. .. ..$ : chr "atomic"
#> .. .. ..$ : chr "2"
#> ..$ \1:List of 2
#> .. ..$ pval: list()
#> .. ..$ eval:List of 2
#> .. .. ..$ : chr "atomic"
#> .. .. ..$ : chr "3"
#> $ eval:List of 3
#> ..$ : chr "+"
#> ..$ :List of 2
#> .. ..$ : chr "atomic"
#> .. ..$ : chr "\\0"
#> ..$ :List of 2
#> .. ..$ : chr "atomic"
#> .. ..$ : chr "\\1"
To replicate the structure:
<- list(
pval_list "\\0"=list(
pval = list(),
eval = list("atomic", "2")
),"\\1"=list(
pval = list(),
eval = list("atomic", "3")
)
)<- list(
tree pval = pval_list,
eval = list("+", list("atomic", "\\0"), list("atomic", "\\1"))
)::expect_equal(x,tree) testthat
This test passes with zero error.
Now the pval
list is not empty. We have an entry for
(2)
and (3)
. Each entry in pval
is a new tree unto itself and contains pval
and
eval
branches.
The eval
branch splits the statement by operators into
“atomic” elements.
For example, if we just parse 2
:
<- evalR::create_tree("2")
x str(x)
#> List of 2
#> $ pval: list()
#> $ eval:List of 2
#> ..$ : chr "atomic"
#> ..$ : chr "2"
The eval
is one level deep and the first element is the
string atomic
. This signifies that this is an end node of
the tree.
Lets expand this just a little bit:
<- evalR::create_tree("-2")
x str(x)
#> List of 2
#> $ pval: list()
#> $ eval:List of 2
#> ..$ : chr "-"
#> ..$ :List of 2
#> .. ..$ : chr "atomic"
#> .. ..$ : chr "2"
Now eval
is two levels deep. The first element states
the operator -
and the second element is another branch
that looks exactly like the eval
branch in the previous
example.
If a parenthesis block is found, then the atomic element will be a
reference to the which pval
element:
<- evalR::create_tree("-(2)")
x str(x)
#> List of 2
#> $ pval:List of 1
#> ..$ \0:List of 2
#> .. ..$ pval: list()
#> .. ..$ eval:List of 2
#> .. .. ..$ : chr "atomic"
#> .. .. ..$ : chr "2"
#> $ eval:List of 2
#> ..$ : chr "-"
#> ..$ :List of 2
#> .. ..$ : chr "atomic"
#> .. ..$ : chr "\\0"
In this example, the eval
second level atomic element is
\0
. This is a reference to the \0
named
element of the pval
branch.
Given a tree, we can execute it with the function
eval_tree
. Here is a basic example:
<- evalR::create_tree("2+3")
x <- evalR::eval_tree(x)
y print(y)
#> [1] 5
There is a convenience function that contains the tree creation
stage. The eval_text
takes text as an input:
<- evalR::eval_text("2+3")
y print(y)
#> [1] 5
The eval_tree
and eval_text
share the
following parameter:
This parameter limits the scope of the execution environment (not in a strictly technical sense). In other words, they limit what values can be reference.
Here is a basic concert example:
<- list("#" = data.frame(x = 1:5, y = 5:1))
map_obj <- evalR::eval_text("log(#x#)", map=map_obj)
y print(y)
#> [1] 0.0000000 0.6931472 1.0986123 1.3862944 1.6094379
Here is a more complex example:
<- list("#" = data.frame(x = 1:5, y = 5:1),"$" = list(z = -(1:5)))
map_obj <- evalR::eval_text("#x# + $z$", map=map_obj)
y print(y)
#> [1] 0 0 0 0 0
To get a sense of the performance. Lets compare different ways we can
run log(1+2)
:
<- "log(1+3)"
text <- evalR::create_tree(text)
tree ::microbenchmark(
microbenchmarklog(1+2)},
{eval(parse(text=text))},
{::eval_tree(tree)},
{evalR::eval_text(text)},
{evalR::create_tree(text)}, n=1000)
{evalR#> Warning in microbenchmark::microbenchmark({: Could not measure a positive
#> execution time for 74 evaluations.
#> Unit: nanoseconds
#> expr min lq mean median uq max neval
#> { log(1 + 2) } 100 100 341 200 200 13100 100
#> { eval(parse(text = text)) } 4100 5000 6300 6100 6600 22700 100
#> { evalR::eval_tree(tree) } 18600 19700 21809 20500 21600 50000 100
#> { evalR::eval_text(text) } 72900 74700 78777 75800 77250 116900 100
#> { evalR::create_tree(text) } 50400 52350 57119 53550 55100 112100 100
#> n 0 0 0 0 0 0 100
The pure R evaluation is much faster than any of the other methods. This is no surprise.
The evalR::eval_tree
block takes a couple times longer
than the eval(parse(text))
execution. This is a trade off
considering the reduced risk. The eval(parse(text))
execution is sometimes much slower when ran outside R Markdown.
We also see that evalR::eval_text
takes much longer than
evalR::eval_tree
. The majority of the
evalR::eval_text
time comes from the internal call to
evalR::create_tree
. This is where the no free lunch
principle comes into play. The cost of the reduced risk is spent in
creating the tree.
If you’re lucky enough to have a set of user input that will be
evaluated multiple times, then the design pattern of generating the tree
once and using evalR::eval_tree
will give similar
performance to a straight eval
call.
This package should be viewed as a building block and not as an end
unto itself. You can consider using it anytime you’re tempted to write
an eval(parse(text))
statement.
Package logo was created using the game-icons.net website. Thank you for such a great product and for people willing to share their talents with others.