Clojure today exists at a balance point between performance and dynamic flexibility. Some users would be willing to sacrifice dynamic features to gain more performance. Other users would be willing to sacrifice performance to gain better development error messages or debugging. The problem we aim to solve is to alter the Clojure build such that we can produce alternate versions of Clojure that are useful to smaller audiences while retaining Clojure as it is today.
Example builds we may want to produce (examples, not finalized):
- "dev" build
- More detailed error messages that would affect performance if included in default build
- Skip locals clearing or other features to aid in debugging
- Other features that may aid in development but affect performance
- "dyn" build
- Similar to what you get now with Clojure 1.6
- "prod" build
- Direct linking (static vars are inlined to be static fn calls) to gain greater perf
- Trim metadata (omit docstrings for speed and space savings)
- "experimental" build
- A place to provide experimental compiler features that we want users to evaluate for the default build
- Build profile artifacts should be available in public Maven repositories for use with existing build systems (Leiningen, Maven)
- Different artifacts: [org.clojure/clojure-foo "1.6.0"]
- Different qualifiers: [org.clojure/clojure "1.6.0-foo"]
- Different classifiers: [org.clojure/clojure "1.6.0" :classifier "foo"]
- Some features may require further compiler flags or metadata to activate even if using the specialized build
- For example, you may need to supply ^:static metadata on a function to get access to this feature
- Ideally, use of the special feature should be simply ignored if not using the specialized build (or a warning should direct you)
- Produce multiple output jars each time we build a version (snapshot or release)
- The jars will contain different versions of the classes
- Specialized build flags or env variables will need to be set during the build for each variant
- Hudson CI server must be able to build and test custom builds as part of the dependency matrix
- Hudson CI server must be able to perform releases for all of the variants
- The compiler will need a standard way to determine which variant is active to make decisions during compilation.
Clojure dyn (AOT)
- Same as now, can redef calls to core Clojure fns both externally and internally
- Metadata available
Clojure prod (AOT)
- Clojure code is compiled with direct linking
- Internal calls from one Clojure core function to another cannot be redefined (due to AOT)
- External code can redef calls to core Clojure fns
- Some metadata is trimmed (docstrings) to save space and loading time
- Same as now (but use of Clojure core depends on which version you use as above)
- As source is loaded and compiled, internal references are statically linked
- External calls to application functions can be redefined
- Internal calls to application functions can be
Clojure prod (AOT) + Application dyn source
Clojure prod (AOT) + Library prod source + Application prod source
How do these scenarios affect what the end Application can do? How do intermediate libraries choose what they depend on and publicize what they're doing to users?
A particular ns could choose different choices for how to link other namespaces, its own namespace, or even some functions in its own namespace. Leaving a small number of "main loop" fns dynamic could allow for runtime system patching.
Maven coordinate options
- Different artifactId
- Releasing different artifacts would probably require running the build multiple times - there are probably some complexities in the CI and release process as a result.
- Different qualifier
- Qualifiers are really designed to give you more information about the purpose of a released version (alpha, beta, GA, build number), so this would be departing from standard Maven expectations.
- Different classifier
- Classifiers are designed for this use case - they allow you to create variants of the same group/artifact/version specialized for a particular environment (JDK variants is one example given)
- Lein and Maven can both specify a classifier - in lein with [org.clojure/clojure "1.6.0" :classifier "foo"]
- It seems good that the main build and variant share Maven co-ords - there is then a standard way to get from main artifact to (for example) the dev variant for tooling.
- How-to in Maven
- Create one Maven profile per output artifact, in each profile:
- Set a property that can be used as the <classifier> in the maven-jar-plugin
- Set <activation><active-by-default> to false
- Set system property compiler flags that need to be active during Clojure compilation for the foo variant
- Turn off generation of the -sources and -javadoc classifier artifacts in these variant builds
- Run the release once per profile - this can be automated in Maven with a matrix build
- How-to in Ant
- Define property that can be set to activate a particular variant
- Set that property when running the build
Maven custom assembly
- The Maven assembly plugin allows for arbitrary customization of the artifacts created by the build. However, creating multiple artifacts with different Maven coordinates will likely require multiple builds to produce each artifact for deployment.
Custom script (ant, make, etc)