Frequently Asked Questions

No.

Since many people are familiar with the syntax of other dynamic languages, and lots of code has already been written in those languages, it is natural to wonder why we didn’t just plug a Matlab or Python front-end into a Julia back-end (or “transpile” code to Julia) in order to get all the performance benefits of Julia without requiring programmers to learn a new language. Simple, right?

The basic issue is that there is nothing special about Julia’s compiler: we use a commonplace compiler (LLVM) with no “secret sauce” that other language developers don’t know about. Indeed, Julia’s compiler is in many ways much simpler than those of other dynamic languages (e.g. PyPy or LuaJIT). Julia’s performance advantage derives almost entirely from its front-end: its language semantics allow a well-written Julia program to give more opportunities to the compiler to generate efficient code and memory layouts. If you tried to compile Matlab or Python code to Julia, our compiler would be limited by the semantics of Matlab or Python to producing code no better than that of existing compilers for those languages (and probably worse). The key role of semantics is also why several existing Python compilers (like Numba and Pythran) only attempt to optimize a small subset of the language (e.g. operations on Numpy arrays and scalars), and for this subset they are already doing at least as well as we could for the same semantics. The people working on those projects are incredibly smart and have accomplished amazing things, but retrofitting a compiler onto a language that was designed to be interpreted is a very difficult problem.

Julia’s advantage is that good performance is not limited to a small subset of “built-in” types and operations, and one can write high-level type-generic code that works on arbitrary user-defined types while remaining fast and memory-efficient. Types in languages like Python simply don’t provide enough information to the compiler for similar capabilities, so as soon as you used those languages as a Julia front-end you would be stuck.

For similar reasons, automated translation to Julia would also typically generate unreadable, slow, non-idiomatic code that would not be a good starting point for a native Julia port from another language.

On the other hand, language interoperability is extremely useful: we want to exploit existing high-quality code in other languages from Julia (and vice versa)! The best way to enable this is not a transpiler, but rather via easy inter-language calling facilities. We have worked hard on this, from the built-in ccall intrinsic (to call C and Fortran libraries) to JuliaInterop packages that connect Julia to Python, Matlab, C++, and more.

The only interfaces that are stable with respect to SemVer of julia version are the Julia Base and standard libraries interfaces described in the documentation and not marked as unstable (e.g., experimental and internal). Functions, types, and constants are not part of the public API if they are not included in the documentation, even if they have docstrings.

Updating Julia may break your code if you use non-public API. If the code is self-contained, it may be a good idea to copy it into your project. If you want to rely on a complex non-public API, especially when using it from a stable package, it is a good idea to open an issue or pull request to start a discussion for turning it into a public API. However, we do not discourage the attempt to create packages that expose stable public interfaces while relying on non-public implementation details of julia and buffering the differences across different julia versions.

Please open an issue or pull request to start a discussion for turning the existing behavior into a public API.

Julia does not have an analog of MATLAB’s clear function; once a name is defined in a Julia session (technically, in module Main), it is always present.

If memory usage is your concern, you can always replace objects with ones that consume less memory. For example, if A is a gigabyte-sized array that you no longer need, you can free the memory with A = nothing. The memory will be released the next time the garbage collector runs; you can force this to happen with GC.gc(). Moreover, an attempt to use A will likely result in an error, because most methods are not defined on type Nothing.

Perhaps you’ve defined a type and then realize you need to add a new field. If you try this at the REPL, you get the error:

Types in module Main cannot be redefined.

While this can be inconvenient when you are developing new code, there’s an excellent workaround. Modules can be replaced by redefining them, and so if you wrap all your new code inside a module you can redefine types and constants. You can’t import the type names into Main and then expect to be able to redefine them there, but you can use the module name to resolve the scope. In other words, while developing you might use a workflow something like this:

When a file is run as the main script using julia file.jl one might want to activate extra functionality like command line argument handling. A way to determine that a file is run in this fashion is to check if abspath(PROGRAM_FILE) == @__FILE__ is true.

Running a Julia script using julia file.jl does not throw InterruptException when you try to terminate it with CTRL-C (SIGINT). To run a certain code before terminating a Julia script, which may or may not be caused by CTRL-C, use atexit. Alternatively, you can use julia -e 'include(popfirst!(ARGS))' file.jl to execute a script while being able to catch InterruptException in the try block. Note that with this strategy PROGRAM_FILE will not be set.

Passing options to julia in a so-called shebang line, as in #!/usr/bin/env julia --startup-file=no, will not work on many platforms (BSD, macOS, Linux) where the kernel, unlike the shell, does not split arguments at space characters. The option env -S, which splits a single argument string into multiple arguments at spaces, similar to a shell, offers a simple workaround:

Julia’s run function launches external programs directly, without invoking an operating-system shell (unlike the system("...") function in other languages like Python, R, or C). That means that run does not perform wildcard expansion of * ("globbing"), nor does it interpret shell pipelines like | or >.

You can still do globbing and pipelines using Julia features, however. For example, the built-in pipeline function allows you to chain external programs and files, similar to shell pipes, and the Glob.jl package implements POSIX-compatible globbing.

You can, of course, run programs through the shell by explicitly passing a shell and a command string to run, e.g. run(`sh -c ls  files.txt)` to use the Unix Bourne shell, but you should generally prefer pure-Julia scripting like run(pipeline(`ls, "files.txt"))`. The reason why we avoid the shell by default is that shelling out sucks: launching processes via the shell is slow, fragile to quoting of special characters, has poor error handling, and is problematic for portability. (The Python developers came to a similar conclusion.)

You might have something like:

and notice that it works fine in an interactive environment (like the Julia REPL), but gives UndefVarError: `x not defined` when you try to run it in script or other file. What is going on is that Julia generally requires you to be explicit about assigning to global variables in a local scope.

Here, x is a global variable, while defines a local scope, and x += 1 is an assignment to a global in that local scope.

As mentioned above, Julia (version 1.5 or later) allows you to omit the global keyword for code in the REPL (and many other interactive environments), to simplify exploration (e.g. copy-pasting code from a function to run interactively). However, once you move to code in files, Julia requires a more disciplined approach to global variables. You have least three options:

More explanation can be found in the manual section on soft scope.

Suppose you call a function like this:

In Julia, the binding of a variable x cannot be changed by passing x as an argument to a function. When calling change_value!(x) in the above example, y is a newly created variable, bound initially to the value of x, i.e. 10; then y is rebound to the constant 17, while the variable x of the outer scope is left untouched.

However, if x is bound to an object of type Array (or any other mutable type). From within the function, you cannot "unbind" x from this Array, but you can change its content. For example:

Here we created a function change_array!, that assigns 5 to the first element of the passed array (bound to x at the call site, and bound to A within the function). Notice that, after the function call, x is still bound to the same array, but the content of that array changed: the variables A and x were distinct bindings referring to the same mutable Array object.

No, you are not allowed to have a using or import statement inside a function. If you want to import a module but only use its symbols inside a specific function or set of functions, you have two options:

The operator = always returns the right-hand side, therefore:

and similarly:

It means that the type of the output is predictable from the types of the inputs. In particular, it means that the type of the output cannot vary depending on the values of the inputs. The following code is not type-stable:

It returns either an Int or a Float64 depending on the value of its argument. Since Julia can’t predict the return type of this function at compile-time, any computation that uses it must be able to cope with values of both types, which makes it hard to produce fast machine code.

Certain operations make mathematical sense but result in errors:

This behavior is an inconvenient consequence of the requirement for type-stability. In the case of sqrt, most users want sqrt(2.0) to give a real number, and would be unhappy if it produced the complex number 1.4142135623730951 + 0.0im. One could write the sqrt function to switch to a complex-valued output only when passed a negative number (which is what sqrt does in some other languages), but then the result would not be type-stable and the sqrt function would have poor performance.

In these and other cases, you can get the result you want by choosing an input type that conveys your willingness to accept an output type in which the result can be represented:

The parameters of a parametric type can hold either types or bits values, and the type itself chooses how it makes use of these parameters. For example, Array{Float64, 2} is parameterized by the type Float64 to express its element type and the integer value 2 to express its number of dimensions. When defining your own parametric type, you can use subtype constraints to declare that a certain parameter must be a subtype (<:) of some abstract type or a previous type parameter. There is not, however, a dedicated syntax to declare that a parameter must be a value of a given type — that is, you cannot directly declare that a dimensionality-like parameter isa Int within the struct definition, for example. Similarly, you cannot do computations (including simple things like addition or subtraction) on type parameters. Instead, these sorts of constraints and relationships may be expressed through additional type parameters that are computed and enforced within the type’s constructors.

As an example, consider

where the user would like to enforce that the third type parameter is always the second plus one. This can be implemented with an explicit type parameter that is checked by an inner constructor method (where it can be combined with other checks):

This check is usually costless, as the compiler can elide the check for valid concrete types. If the second argument is also computed, it may be advantageous to provide an outer constructor method that performs this calculation:

Julia uses machine arithmetic for integer computations. This means that the range of Int values is bounded and wraps around at either end so that adding, subtracting and multiplying integers can overflow or underflow, leading to some results that can be unsettling at first:

Clearly, this is far from the way mathematical integers behave, and you might think it less than ideal for a high-level programming language to expose this to the user. For numerical work where efficiency and transparency are at a premium, however, the alternatives are worse.

One alternative to consider would be to check each integer operation for overflow and promote results to bigger integer types such as Int128 or BigInt in the case of overflow. Unfortunately, this introduces major overhead on every integer operation (think incrementing a loop counter) — it requires emitting code to perform run-time overflow checks after arithmetic instructions and branches to handle potential overflows. Worse still, this would cause every computation involving integers to be type-unstable. As we mentioned above, type-stability is crucial for effective generation of efficient code. If you can’t count on the results of integer operations being integers, it’s impossible to generate fast, simple code the way C and Fortran compilers do.

A variation on this approach, which avoids the appearance of type instability is to merge the Int and BigInt types into a single hybrid integer type, that internally changes representation when a result no longer fits into the size of a machine integer. While this superficially avoids type-instability at the level of Julia code, it just sweeps the problem under the rug by foisting all of the same difficulties onto the C code implementing this hybrid integer type. This approach can be made to work and can even be made quite fast in many cases, but has several drawbacks. One problem is that the in-memory representation of integers and arrays of integers no longer match the natural representation used by C, Fortran and other languages with native machine integers. Thus, to interoperate with those languages, we would ultimately need to introduce native integer types anyway. Any unbounded representation of integers cannot have a fixed number of bits, and thus cannot be stored inline in an array with fixed-size slots — large integer values will always require separate heap-allocated storage. And of course, no matter how clever a hybrid integer implementation one uses, there are always performance traps — situations where performance degrades unexpectedly. Complex representation, lack of interoperability with C and Fortran, the inability to represent integer arrays without additional heap storage, and unpredictable performance characteristics make even the cleverest hybrid integer implementations a poor choice for high-performance numerical work.

An alternative to using hybrid integers or promoting to BigInts is to use saturating integer arithmetic, where adding to the largest integer value leaves it unchanged and likewise for subtracting from the smallest integer value. This is precisely what Matlab™ does:

At first blush, this seems reasonable enough since 9223372036854775807 is much closer to 9223372036854775808 than -9223372036854775808 is and integers are still represented with a fixed size in a natural way that is compatible with C and Fortran. Saturated integer arithmetic, however, is deeply problematic. The first and most obvious issue is that this is not the way machine integer arithmetic works, so implementing saturated operations requires emitting instructions after each machine integer operation to check for underflow or overflow and replace the result with typemin(Int) or typemax(Int) as appropriate. This alone expands each integer operation from a single, fast instruction into half a dozen instructions, probably including branches. Ouch. But it gets worse — saturating integer arithmetic isn’t associative. Consider this Matlab computation:

This makes it hard to write many basic integer algorithms since a lot of common techniques depend on the fact that machine addition with overflow is associative. Consider finding the midpoint between integer values lo and hi in Julia using the expression (lo + hi) >>> 1:

See? No problem. That’s the correct midpoint between 2^62 and 2^63, despite the fact that n + 2n is -4611686018427387904. Now try it in Matlab:

Oops. Adding a >>> operator to Matlab wouldn’t help, because saturation that occurs when adding n and 2n has already destroyed the information necessary to compute the correct midpoint.

Not only is lack of associativity unfortunate for programmers who cannot rely it for techniques like this, but it also defeats almost anything compilers might want to do to optimize integer arithmetic. For example, since Julia integers use normal machine integer arithmetic, LLVM is free to aggressively optimize simple little functions like f(k) = 5k-1. The machine code for this function is just this:

The actual body of the function is a single leaq instruction, which computes the integer multiply and add at once. This is even more beneficial when f gets inlined into another function:

Since the call to f gets inlined, the loop body ends up being just a single leaq instruction. Next, consider what happens if we make the number of loop iterations fixed:

Because the compiler knows that integer addition and multiplication are associative and that multiplication distributes over addition — neither of which is true of saturating arithmetic — it can optimize the entire loop down to just a multiply and an add. Saturated arithmetic completely defeats this kind of optimization since associativity and distributivity can fail at each loop iteration, causing different outcomes depending on which iteration the failure occurs in. The compiler can unroll the loop, but it cannot algebraically reduce multiple operations into fewer equivalent operations.

The most reasonable alternative to having integer arithmetic silently overflow is to do checked arithmetic everywhere, raising errors when adds, subtracts, and multiplies overflow, producing values that are not value-correct. In this blog post, Dan Luu analyzes this and finds that rather than the trivial cost that this approach should in theory have, it ends up having a substantial cost due to compilers (LLVM and GCC) not gracefully optimizing around the added overflow checks. If this improves in the future, we could consider defaulting to checked integer arithmetic in Julia, but for now, we have to live with the possibility of overflow.

In the meantime, overflow-safe integer operations can be achieved through the use of external libraries such as SaferIntegers.jl. Note that, as stated previously, the use of these libraries significantly increases the execution time of code using the checked integer types. However, for limited usage, this is far less of an issue than if it were used for all integer operations. You can follow the status of the discussion here.

As the error states, an immediate cause of an UndefVarError on a remote node is that a binding by that name does not exist. Let us explore some of the possible causes.

The closure x->x carries a reference to Foo, and since Foo is unavailable on node 2, an UndefVarError is thrown.

Globals under modules other than Main are not serialized by value to the remote node. Only a reference is sent. Functions which create global bindings (except under Main) may cause an UndefVarError to be thrown later.

In the above example, @everywhere module Foo defined Foo on all nodes. However the call to Foo.foo() created a new global binding gvar on the local node, but this was not found on node 2 resulting in an UndefVarError error.

Note that this does not apply to globals created under module Main. Globals under module Main are serialized and new bindings created under Main on the remote node.

This does not apply to function or struct declarations. However, anonymous functions bound to global variables are serialized as can be seen below.

As you’ll see if you try this, the result is a MethodError:

This is because Vector{Real} is not a supertype of Vector{Int}! You can solve this problem with something like foo(bar::Vector{T}) where {T<:Real} (or the short form foo(bar::Vector{<:Real}) if the static parameter T is not needed in the body of the function). The T is a wild card: you first specify that it must be a subtype of Real, then specify the function takes a Vector of with elements of that type.

This same issue goes for any composite type Comp, not just Vector. If Comp has a parameter declared of type Y, then another type Comp2 with a parameter of type X<:Y is not a subtype of Comp. This is type-invariance (by contrast, Tuple is type-covariant in its parameters). See Parametric Composite Types for more explanation of these.

The main argument against + is that string concatenation is not commutative, while + is generally used as a commutative operator. While the Julia community recognizes that other languages use different operators and * may be unfamiliar for some users, it communicates certain algebraic properties.

Note that you can also use string(...) to concatenate strings (and other values converted to strings); similarly, repeat can be used instead of ^ to repeat strings. The interpolation syntax is also useful for constructing strings.

There is only one difference, and on the surface (syntax-wise) it may seem very minor. The difference between using and import is that with using you need to say function Foo.bar(.. to extend module Foo’s function bar with a new method, but with import Foo.bar, you only need to say function bar(... and it automatically extends module Foo’s function bar.

The reason this is important enough to have been given separate syntax is that you don’t want to accidentally extend a function that you didn’t know existed, because that could easily cause a bug. This is most likely to happen with a method that takes a common type like a string or integer, because both you and the other module could define a method to handle such a common type. If you use import, then you’ll replace the other module’s implementation of bar(s::AbstractString) with your new implementation, which could easily do something completely different (and break all/many future usages of the other functions in module Foo that depend on calling bar).

Unlike many languages (for example, C and Java), Julia objects cannot be "null" by default. When a reference (variable, object field, or array element) is uninitialized, accessing it will immediately throw an error. This situation can be detected using the isdefined or isassigned functions.

Some functions are used only for their side effects, and do not need to return a value. In these cases, the convention is to return the value nothing, which is just a singleton object of type Nothing. This is an ordinary type with no fields; there is nothing special about it except for this convention, and that the REPL does not print anything for it. Some language constructs that would not otherwise have a value also yield nothing, for example if false; end.

For situations where a value x of type T exists only sometimes, the Union{T, Nothing} type can be used for function arguments, object fields and array element types as the equivalent of Nullable, Option or Maybe in other languages. If the value itself can be nothing (notably, when T is Any), the Union{Some{T}, Nothing} type is more appropriate since x == nothing then indicates the absence of a value, and x == Some(nothing) indicates the presence of a value equal to nothing. The something function allows unwrapping Some objects and using a default value instead of nothing arguments. Note that the compiler is able to generate efficient code when working with Union{T, Nothing} arguments or fields.

To represent missing data in the statistical sense (NA in R or NULL in SQL), use the missing object. See the Missing Values section for more details.

In some languages, the empty tuple (()) is considered the canonical form of nothingness. However, in julia it is best thought of as just a regular tuple that happens to contain zero values.

The empty (or "bottom") type, written as Union{} (an empty union type), is a type with no values and no subtypes (except itself). You will generally not need to use this type.

In Julia, x += y gets replaced during lowering by x = x + y. For arrays, this has the consequence that, rather than storing the result in the same location in memory as x, it allocates a new array to store the result. If you prefer to mutate x, use x .+= y to update each element individually.

While this behavior might surprise some, the choice is deliberate. The main reason is the presence of immutable objects within Julia, which cannot change their value once created. Indeed, a number is an immutable object; the statements x = 5; x += 1 do not modify the meaning of 5, they modify the value bound to x. For an immutable, the only way to change the value is to reassign it.

To amplify a bit further, consider the following function:

After a call like x = 5; y = power_by_squaring(x, 4), you would get the expected result: x == 5 && y == 625. However, now suppose that *=, when used with matrices, instead mutated the left hand side. There would be two problems:

Because supporting generic programming is deemed more important than potential performance optimizations that can be achieved by other means (e.g., using broadcasting or explicit loops), operators like += and *= work by rebinding new values.

While the streaming I/O API is synchronous, the underlying implementation is fully asynchronous.

Consider the printed output from the following:

This is happening because, while the write call is synchronous, the writing of each argument yields to other tasks while waiting for that part of the I/O to complete.

print and println "lock" the stream during a call. Consequently changing write to println in the above example results in:

You can lock your writes with a ReentrantLock like this:

Zero-dimensional arrays are arrays of the form Array{T,0}. They behave similar to scalars, but there are important differences. They deserve a special mention because they are a special case which makes logical sense given the generic definition of arrays, but might be a bit unintuitive at first. The following line defines a zero-dimensional array:

In this example, A is a mutable container that contains one element, which can be set by A[] = 1.0 and retrieved with A[]. All zero-dimensional arrays have the same size (size(A) == ()), and length (length(A) == 1). In particular, zero-dimensional arrays are not empty. If you find this unintuitive, here are some ideas that might help to understand Julia’s definition.

It is also important to understand the differences to ordinary scalars. Scalars are not mutable containers (even though they are iterable and define things like length, getindex, e.g. 1[] == 1). In particular, if x = 0.0 is defined as a scalar, it is an error to attempt to change its value via x[] = 1.0. A scalar x can be converted into a zero-dimensional array containing it via fill(x), and conversely, a zero-dimensional array a can be converted to the contained scalar via a[]. Another difference is that a scalar can participate in linear algebra operations such as 2 * rand(2,2), but the analogous operation with a zero-dimensional array fill(2) * rand(2,2) is an error.

You may find that simple benchmarks of linear algebra building blocks like

can be different when compared to other languages like Matlab or R.

Since operations like this are very thin wrappers over the relevant BLAS functions, the reason for the discrepancy is very likely to be

Julia compiles and uses its own copy of OpenBLAS, with threads currently capped at 8 (or the number of your cores).

Modifying OpenBLAS settings or compiling Julia with a different BLAS library, eg Intel MKL, may provide performance improvements. You can use MKL.jl, a package that makes Julia’s linear algebra use Intel MKL BLAS and LAPACK instead of OpenBLAS, or search the discussion forum for suggestions on how to set this up manually. Note that Intel MKL cannot be bundled with Julia, as it is not open source.

When using julia in high-performance computing (HPC) facilities, invoking n julia processes simultaneously creates at most n temporary copies of precompilation cache files. If this is an issue (slow and/or small distributed file system), you may:

The Stable version of Julia is the latest released version of Julia, this is the version most people will want to run. It has the latest features, including improved performance. The Stable version of Julia is versioned according to SemVer as v1.x.y. A new minor release of Julia corresponding to a new Stable version is made approximately every 4-5 months after a few weeks of testing as a release candidate. Unlike the LTS version the a Stable version will not normally receive bugfixes after another Stable version of Julia has been released. However, upgrading to the next Stable release will always be possible as each release of Julia v1.x will continue to run code written for earlier versions.

You may prefer the LTS (Long Term Support) version of Julia if you are looking for a very stable code base. The current LTS version of Julia is versioned according to SemVer as v1.6.x; this branch will continue to receive bugfixes until a new LTS branch is chosen, at which point the v1.6.x series will no longer received regular bug fixes and all but the most conservative users will be advised to upgrade to the new LTS version series. As a package developer, you may prefer to develop for the LTS version, to maximize the number of users who can use your package. As per SemVer, code written for v1.0 will continue to work for all future LTS and Stable versions. In general, even if targeting the LTS, one can develop and run code in the latest Stable version, to take advantage of the improved performance; so long as one avoids using new features (such as added library functions or new methods).

You may prefer the nightly version of Julia if you want to take advantage of the latest updates to the language, and don’t mind if the version available today occasionally doesn’t actually work. As the name implies, releases to the nightly version are made roughly every night (depending on build infrastructure stability). In general nightly released are fairly safe to use—​your code will not catch on fire. However, they may be occasional regressions and or issues that will not be found until more thorough pre-release testing. You may wish to test against the nightly version to ensure that such regressions that affect your use case are caught before a release is made.

Finally, you may also consider building Julia from source for yourself. This option is mainly for those individuals who are comfortable at the command line, or interested in learning. If this describes you, you may also be interested in reading our guidelines for contributing.

Links to each of these download types can be found on the download page at https://julialang.org/downloads/. Note that not all versions of Julia are available for all platforms.

Each minor version of julia has its own default environment. As a result, upon installing a new minor version of Julia, the packages you added using the previous minor version will not be available by default. The environment for a given julia version is defined by the files Project.toml and Manifest.toml in a folder matching the version number in .julia/environments/, for instance, .julia/environments/v1.3.

If you install a new minor version of Julia, say 1.4, and want to use in its default environment the same packages as in a previous version (e.g. 1.3), you can copy the contents of the file Project.toml from the 1.3 folder to 1.4. Then, in a session of the new Julia version, enter the "package management mode" by typing the key ], and run the command instantiate.

This operation will resolve a set of feasible packages from the copied file that are compatible with the target Julia version, and will install or update them if suitable. If you want to reproduce not only the set of packages, but also the versions you were using in the previous Julia version, you should also copy the Manifest.toml file before running the Pkg command instantiate. However, note that packages may define compatibility constraints that may be affected by changing the version of Julia, so the exact set of versions you had in 1.3 may not work for 1.4.