Tools for managing namespaces in Clojure. Parse ns declarations from
source files, extract their dependencies, build a graph of namespace
dependencies within a project, update that graph as files change, and
reload files in the correct order.
This is only about namespace dependencies within a single project. It has nothing to do with Leiningen, Maven, JAR files, or repositories.
Latest stable release is 0.2.11
Leiningen stable dependency information:
[org.clojure/tools.namespace "0.2.11"]
Maven stable dependency information:
<dependency>
<groupId>org.clojure</groupId>
<artifactId>tools.namespace</artifactId>
<version>0.2.11</version>
</dependency>
Latest development release is 0.3.0-alpha4
Leiningen latest dependency information:
[org.clojure/tools.namespace "0.3.0-alpha4"]
Git master branch is at 0.3.0-SNAPSHOT
Leiningen dependency information for development snapshots:
:dependencies [[org.clojure/tools.namespace "0.3.0-SNAPSHOT"]]
:repositories [["sonatype-oss-public"
"https://oss.sonatype.org/content/groups/public/"]]
See also Maven Settings and Repositories on dev.clojure.org.
tools.namespace consists of several parts:
clojure.tools.namespace.parse: A parser for namespace declarations
in Clojure source files. Given a stream of characters from a Clojure
source file, it can find the ns declaration and parse the :require
and :use clauses to find the names of other namespaces that file
depends on. This is all syntactic analysis: it does not
evaluate any code.
clojure.tools.namespace.find: Utilities to search for Clojure namespaces on the filesystem, in directories or JAR files. Combined with java.classpath, it can search for namespaces on the Java classpath. This namespace contains most of the functions in clojure.tools.namespace version 0.1.x.
clojure.tools.namespace.repl: Utilities to load and reload code based on the namespace dependency graph. This takes some explaining, see below. c.t.n.repl is built out of smaller parts:
- c.t.n.dependency - generic dependency graph data structure
- c.t.n.track - namespace dependency tracker
- c.t.n.file - file-reader extension to tracker
- c.t.n.dir - directory-scanner extension to tracker
- c.t.n.reload - namespace-reloading extension to tracker
You can recombine these parts in other ways, but c.t.n.repl is the primary public entry-point to their functionality.
clojure.tools.namespace.move: Utilities to aid in moving and renaming Clojure namespaces. This code is still ALPHA, and it modifies your source files, so be careful.
New in version 0.3.0-alpha1
These namespaces are .cljc files usable from both Clojure(JVM) and
ClojureScript:
- c.t.n.dependency
- c.t.n.track
- c.t.n.parse
These namespaces are usable on Clojure(JVM) only but can analyze both Clojure(JVM) and ClojureScript source files:
- c.t.n.file
- c.t.n.dir
- c.t.n.find
Most functions now take an optional "platform" argument, which is one
of the constant values defined in c.t.n.find: clj or cljs. The
default is clj.
These namespaces are still Clojure(JVM) only:
- c.t.n.reload
- c.t.n.repl
- c.t.n.move.
c.t.n.repl is a smarter way to reload code.
The traditional way to reload Clojure code without restarting the JVM
is (require ... :reload) or :reload-all or an editor/IDE feature
that does the same thing. This has several problems:
-
If you modify two namespaces which depend on each other, you must remember to reload them in the correct order to avoid compilation errors.
-
If you remove definitions from a source file and then reload it, those definitions are still available in memory. If other code depends on those definitions, it will continue to work but will break the next time you restart the JVM.
-
If the reloaded namespace contains
defmulti, you must also reload all of the associateddefmethodexpressions. -
If the reloaded namespace contains
defprotocol, you must also reload any records or types implementing that protocol and replace any existing instances of those records/types with new instances. -
If the reloaded namespace contains macros, you must also reload any namespaces which use those macros.
-
If the running program contains functions which close over values in the reloaded namespace, those closed-over values are not updated. (This is common in web applications which construct the "handler stack" as a composition of functions.)
Often the only surefire way to reload Clojure code is to restart the JVM. A large Clojure application can take 20 seconds or more just to compile. I wrote tools.namespace to help speed up this development cycle.
For more detail on how I use tools.namespace in my development workflow, see the article My Clojure Workflow, Reloaded.
There's only one important function, refresh:
user=> (require '[clojure.tools.namespace.repl :refer [refresh]])
nil
user=> (refresh)
:reloading (com.example.util com.example.app com.example.app-test)
:ok
The refresh function will scan all the directories on the classpath
for Clojure source files, read their ns declarations, build a graph
of their dependencies, and load them in dependency order. (You can
change the directories it scans with set-refresh-dirs.)
Later on, after you have changed and saved a few files in your editor, run it again:
user=> (refresh)
:reloading (com.example.app com.example.app-test)
:ok
Based on file modification timestamps and the graph of dependencies,
the refresh function will reload only the namespaces that have
changed, in dependency order. But first, it will unload (remove) the
namespaces that changed to clear out any old definitions.
This is quite unlike (require ... :reload). Calling refresh will
blow away your old code. Sometimes this is helpful: it can catch
trivial mistakes like deleting a function that another piece of code
depends on. But sometimes it hurts when you have built-up application
state stored in a Var that got deleted by refresh.
This brings us to the next section:
Being able to safely destroy and reload namespaces without breaking your application requires some discipline and careful design. It won't "just work" on any Clojure project.
The first rule for making your application reload-safe is no global state. That means you should avoid things like this:
(def state-of-world (ref {}))
(def object-handle (atom nil))
c.t.n.repl/refresh will destroy those Vars when it reloads the
namespace (even if you used defonce).
Instead of storing your state in global Vars, store it locally in an object that represents the running state of your application. Then provide a constructor function to initialize that state:
(defn create-application []
{:state-of-world (ref {})
:object-handle (atom nil)})
You can choose what representation works best for your application: map, vector, record, or even just a single Ref by itself.
Typically you'll still need one global def somewhere, perhaps in the
REPL itself, to hold the current application instance. See the next
section.
The second rule for making your application reload-safe is to have a consistent way to start and stop the entire system.
The "start" function should:
-
Acquire stateful resources such as sockets, files, and database connections
-
Start threads or background processes
-
Initialize application state such as caches or counters
-
Return an object encapsulating the state of the application
The "stop" function should take the state returned by "start" as an argument and do the opposite:
-
Close or release stateful resourecs
-
Stop all background processes
-
Clear out application state
It might take a few tries to get it right, but once you have working start and stop functions you can have a workflow like this:
Step 1. Start up a REPL.
Step 2. Load the app:
user=> (require '[clojure.tools.namespace.repl :refer [refresh]])
user=> (refresh)
user=> (def my-app (start-my-app))
Step 3. Test it out.
Step 4. Modify some source files.
Step 5. Restart:
user=> (stop-my-app my-app)
user=> (refresh)
user=> (def my-app (start-my-app))
(You could also combine all those steps in a single utility function, but see warnings below.)
After that, you've got a squeaky-clean new instance of your app running, in a fraction of the time it takes to restart the JVM.
If an exception is thrown while loading a namespace, refresh stops,
prints the namespace that caused the exception, and returns the
exception. You can print the rest of the stacktrace with
clojure.repl/pst; the exception itself is bound to *e.
user=> (refresh)
:reloading (com.example.app com.example.app-test)
:error-while-loading com.example.app
#<IllegalArgumentException java.lang.IllegalArgumentException:
Parameter declaration cond should be a vector>
user=> (clojure.repl/pst)
IllegalArgumentException Parameter declaration cond should be a vector
clojure.core/assert-valid-fdecl (core.clj:6567)
clojure.core/sigs (core.clj:220)
clojure.core/defn (core.clj:294)
clojure.lang.Var.invoke (Var.java:427)
...
Remember that any namespaces which depend on the namespace that caused the exception do not exist at this point: they have been removed but not yet reloaded.
After you fix the problem, call refresh again and it will resume
reloading where it left off.
NOTE: If your current REPL namespace is one of those that has not yet been reloaded, then none of the functions you defined in that namespace will exist! Starting with version 0.2.8, tools.namespace will attempt to restore aliases to the namespaces which were successfully loaded.
So, for example, if your current REPL namespace is named dev and
contains this ns declaration:
(ns dev
(:require [com.example.foo :as foo]
[com.example.bar :as bar]
[clojure.tools.namespace.repl :as tns]))
And you get an error on refresh like this:
dev=> (tns/refresh)
:reloading (com.example.bar dev)
:error-while-loading com.example.bar
#<CompilerException ... compiling:(com/example/bar.clj:1:21)>
Then the functions in com.example.foo should still be available in
the dev namespace via the alias foo.
ns syntax: Clojure's ns macro is notoriously lax in what
syntax it accepts. tools.namespace.parse is somewhat liberal, but it
cannot handle every possible variation of syntax that ns does. Stick
to the docstrings of ns and require and everything should be fine.
AOT-compilation: Reloading code does not work in the presence of
AOT-compiled namespaces. If you are using AOT-compilation in your
project, make sure it is disabled and you have deleted any
AOT-compiled .class files before starting a REPL development
session. (In Leiningen, run lein clean.)
Note that the presence of :main in project.clj triggers
AOT-compilation in some versions of Leiningen.
Conflicts: Other libraries which also do code-reloading may conflict with tools.namespace. One known example is ring-devel (as of Ring version 1.1.6) which uses ns-tracker, which uses an older version of tools.namespace.
REPL namespace: Be careful when reloading the namespace in which
you run your REPL. Because namespaces are removed when reloading, all
your past definitions are lost. Either keep your REPL in a namespace
which has no file associated with it, such as user, or put all your
REPL definitions in a file so that they can be reloaded.
Fully-qualified names: Be careful when using fully-qualified
symbol names without namespace aliases (require with no :as). If
the namespace happens to be loaded already, it will not necessarily
cause an error if you forget to require it, but the dependency graph
of namespaces will be incorrect.
Old definitions: Beware of code which has references to old definitions, especially references to things you created in the REPL.
Rolling your own: If you create your own instance of the dependency tracker, do not store it in a namespace which gets reloaded.
Be careful defining a helper function in a namespace which calls
refresh if that namespace also could get reloaded. For example, you
might try to combine the stop-refresh-start code from the "Managed
Lifecycle" section into a single function:
(def my-app nil)
(defn restart []
(stop-my-app my-app)
(refresh)
(alter-var-root #'my-app (constantly (start-my-app))))
This won't work if the namespace containing restart could get
reloaded. After refresh, the namespace containing restart has been
dropped, but the function continues to run in the old namespace and
refer to old Vars.
If you want to run some code after refresh, you can pass an option
naming a function you want to run after a successful reload. The
value of this option must be a symbol, and it must be fully
namespace-qualified. The previous example could be correctly written
(assuming these functions are defined in the dev namespace):
(def my-app nil)
(defn start []
(alter-var-root #'my-app (constantly (start-my-app))))
(defn restart []
(stop-my-app my-app)
(refresh :after 'dev/start))
Namespace aliases created at the REPL will still refer to the old namespace after refresh. For example:
user=> (require '[com.example.foo :as foo])
user=> foo/bar
user=> (refresh)
:reloading (com.example.foo)
:ok
user=> foo/bar ; this is the *old* foo/bar
If you try to recreate the alias with the new namespace, you will get an error:
user=> (require '[com.example.foo :as foo])
IllegalStateException Alias foo already exists in
namespace user, aliasing com.example.foo
clojure.lang.Namespace.addAlias (Namespace.java:224)
The only way out is to remove the alias before recreating it:
user=> (ns-unalias *ns* 'foo)
nil
user=> (alias 'foo 'com.example.foo)
When reloading namespaces which contain protocols, be careful that you do not leave any old instances of records or types implementing those protocols.
For example, if you have a namespace like this:
(ns com.example.foo)
(defprotocol IFoo
(foo [this]))
(defrecord FooRecord []
IFoo (foo [this] nil))
And you do something like the following at the REPL:
user=> (def my-foo (->FooRecord))
user=> (clojure.tools.namespace.repl/refresh)
user=> (foo my-foo)
You will get a confusing error message like this:
IllegalArgumentException
No implementation of method: :foo
of protocol: #'com.example.foo/IFoo
found for class: com.example.foo.FooRecord
clojure.core/-cache-protocol-fn (core_deftype.clj:527)
That's because my-foo is an instance of the old version of
FooRecord, implementing the old version of IFoo. As far as the
JVM is concerned, the old IFoo and the new IFoo are completely
different classes.
To avoid this problem, always create new instances of records after a refresh.
Calling prefer-method is a global side-effect. If you modify a call
to prefer-method and reload the namespace containing it, Clojure may
throw "java.lang.IllegalStateException: Preference conflict in multimethod."
The workaround is to call remove-method before reloading.
tools.namespace cannot detect this situation automatically. See [TNS-23].
In rare cases, reloading a lot of code may lead to out-of-memory
errors from the JVM like java.lang.OutOfMemoryError: PermGen space.
You may be able to mitigate this by increasing the size of the "Permanent Generation" where the JVM stores compiled classes. To do this, add the following command-line argument to your JVM startup:
-XX:MaxPermSize=<N>
where <N> is a number with a suffix like m for megabytes.
To find the default MaxPermSize for your JDK, run
java -XX:+PrintFlagsFinal and search the results for "MaxPermSize".
Try doubling it.
The Permanent Generation was removed in JDK 1.8 ([JEP 122]) so this section no longer applies.
In some older JDKs (1.5) the default garbage collector did not collect
the Permanent Generation at all unless it was explicitly enabled with
-XX:+CMSPermGenSweepingEnabled.
Some projects have a "project REPL" or a "scratch" namespace where you
want keep state during development. You can use the functions
disable-unload! and disable-reload! in
clojure.tools.namespace.repl to prevent refresh from automatically
un/reloading those namespaces.
Use this feature sparingly: it exists as a development-time convenience, not a work-around for code that is not reload-safe. Also, see the warnings about aliases, below. Aliases to reloaded namespaces will break if the namespace containing the alias is not reloaded also.
After an error, refresh will not attempt to recover symbol
mappings and aliases for namespaces with disable-unload! or
disable-reload! set.
Copyright © 2012 Stuart Sierra All rights reserved. The use and distribution terms for this software are covered by the Eclipse Public License 1.0 which can be found in the file epl-v10.html at the root of this distribution. By using this software in any fashion, you are agreeing to be bound by the terms of this license. You must not remove this notice, or any other, from this software.