Camel 3.0 Ideas
Camel is now almost 6 years old and its second revision camel-2.x is more than 4.5 years old already. Camel is extremely mature, used in production by a large number of organizations from small to large and even governments. We feel like we really hit the initial target of simplifying integration. Camel's middleware abstraction api and the eip based routing brought a lot of positive feedback from users.
There is however more that could be done to simplify the work of integration developers who need new components (not shipped with camel for licensing - copyleft of commercial - or other reasons) or new integration patterns or algorithms or even new tools. We learned a lot in the past years and benefited from a strong and continuously growing community. It's time to put what we learned to good use and re-engineer your favourite integration framework yet again.
The middleware abstractions look pretty solid, and aside from some possible reshuffling we don't expect major changes. As a consequence, most of the components will retain the same general feel. The core will however be rearchitected to become even more pluggable and modular. We will however spare no effort to make a new Camel 3 be as backward compatible as possible and when not possible at least provide a painless migration path.
This is a mindmap of ideas for improving Camel 3.0. Fell free to discuss this on the Camel Mailing Lists if you have other ideas or feedback.
Table of contents
(+1: claus, cmueller)
We should drop support for JDK6, and require JDK7 as minimim version. eg build and compile the release with JDK7.
We should aim to be compatible with JDK8.
(hz) Why? Isn't OpenJDK 6 still supported by RedHat and very much in use? Even Oracle's Java 6 is still in use. Maybe some bundles would require JDK7+, buy why not keep things like the api for instance still supported with Java 6? A strong argument may sway my opinion, but for now, I fail to see the reason.
JDK8 Java DSL
It would be good to have a camel-java8-dsl component that offers a JDK8 DSL which uses all the nice new stuff from JDK8 with higher order functions, closures, et all.
Though this may comes later in Camel 3.x when JDK8 is GA.
At least stuff like Predicate, Expression, AggregationStrategy etc. are "functional interfaces" (containing only one method) and Java 8 applications can implement them using lambdas. That's only a start, but it doesn't require a specific DSL.
Routing Core Re-engineering (raulk)
The routing core of Camel 2.x is heavily based on a recursive call pattern, where Processors are responsible for calling the next one along the chain. This results in lengthy and meaningless stacktraces (difficult to make sense out of and debug for newcomers) and higher memory usage due to retention of local variables for a longer time than strictly needed.
Moreover, Camel weaves a large number of "plumbing" processors along the way which should not really be processors because they form part of the very essence of the routing core, e.g. error handlers, stream caching interceptors, trace interceptors, async processor helpers, MDC, etc.
The proposal is to shift towards an iterative model, by redesigning the logic of Camel routing. The suggested model is defined by these pillars:
- A single class, or a limited set of them, contain the routing logic of Camel. Package name:
org.apache.camel.core.routing. Central (abstract) class:
RoutingCore. Concrete realisations could be: PipeliningRoutingCore, MulticastRoutingCore, depending on the fundamental routing pattern.
- The RoutingCore iteratively calls the routing steps, one after another. The routing steps return their result to the
RoutingCore, who is in charge of calling the next element subsequently. OUT and IN are bridged if necessary (PipeliningRoutingCore).
- The Processor interface is crumbled up into its many specialisations, each of which represents a distinct concept of the Camel framework: RoutingDecider (EIPs should only take decisions about the routing, but not perform the routing itself, e.g. choice, filter, loop, throttle, etc.; see examples in subsection below.), Actions, ErrorHandler (already exists), Interceptor, etc.
- The RoutingCore is responsible of all the "magic" now disseminated across a number of processors. Assisted by Helper classes.
The goal of this idea isn't to zap off recursion altogether, just to consolidate the routing logic into a handful of cornerstone classes.
Camel is no longer a baby and the framework concepts are well mature, thus they should be transferred to the API and avoid making everything a raw Processor.
Converting some EIPs from "performers" to mere "deciders"
- choice() => evaluates the predicates and returns the List of Processors or Endpoints to invoke.
- filter() => same as choice(), but returning null if the filter doesn't match, to continue to the next routing step.
- loop() => evaluates whether the looping control predicate still stands. If yes, it returns the processors to invoke, where the last is itself (to trigger the looping logic again); else, it returns null to continue to the next routing step.
- throttle() => pauses accordingly and then returns the endpoint/processors to invoke.
Clearer Architecture of Camel Core
- The camel components should know as little as possible about camel core
- The classes needed to setup camel should be separate from the things needed at run time
- Camel Core should be tiny as possible and only contain what really is core
So why should this be important? Currently components depend on camel-core as a whole and there are no further rules which classes the components should use and which classes should be private to core. Even classes from the impl package are needed. So this means that any refactoring we do in camel core could affect all components. As camel is growing steadily this can become quite problematic.
Split camel-core into multiple parts (hadrian)
(+1: cmueller, hadrian, claus)
Claus: Important to be 99+% backwards compatible with Camel 2.x.
There are multiple benefits and less constraints. A separate api jar would allow the definition of a 'route container' which is currently one of the roles of the CamelContext. This allows primarily alternative implementations of camel for constrained environments (such as real time systems, for instance). Processors/Routes/Components written against the api could be deployed on any camel implementation (as long as all necessary features are supported).
These should be structured in a way that these big building blocks do not have cyclic dependencies. Any other cycles can be ignored in this step.
Allowed depdencies ( "->" means may use, may depend on):
- * -> api
- end user config code -> builder
- builder -> impl
Avoid shading google concurrent linked map in camel-core
The shaded Google concurrent map should IMHO be pluggable, so people can run without this as default. And then people can install that google JAR on their classpath and Camel can pickup and use that. This JAR only helps in SMX/Karaf installations when having concurrent startup of many Camel apps. For regular users this does not bring any benefits to the table. This can help slim down the size of the camel-core JAR.
We can either auto detect the google class, as people did in the past with JDK1.3/1.4 and the apache commons collection. eg using commons collection on JDK1.3 and not in JDK1.4 as it had that out of the box.
Define scope and rules for camel-core packages (champion?)
In extension to the previous paragraph each camel package should have a clear scope that defines what to put in the package and what not. There should be rules that define what dependencies are allowed for classes in a package. The minimum goal is to guarantee that by following the rules dependency cycles can not happen. Additionally the rules should minimize dependencies between packages to achieve loose coupling between packages and high coherence inside a package.
More flexible routes at runtime (claus)
When routes is added in Camel 2.x architecture, global cross cutting concerns such as error handlers, interceptors, onCompletion etc. is applied when the route is added. We need to separate this and have those applied during routing. The
Channel needs to do this and therefore it must be more dynamic than its currently is. And we need to enlist the various global cross cutting concerns by their xxxDefintions in the CamelContext, so we can access them at any time. This allows end users also much more easily to add/remove interceptors, error handlers and whatnot at runtime. And it makes it much easier to add routes generated from JAXB or other sources, as we don't need to prepare or anyhow mold the
RouteDefinition given. See ticket CAMEL-3024 for some details.
Fix routes with multiple inputs (claus)
The current implementation of routes with multiple inputs is to clone the route, which means you essentially got 2+ routes if a route has multiple inputs. However routes with multiple inputs is seldom used. The actual solution will depend on the api refactoring.
Route initialization logic for Java DSL and XML DSLs (claus)
The Java DSL does its route initialization slightly a bit different than the XML DSLs, due the nature of it, and the fact the fluent builders can do additional logic, which the JAXB model of XML DSLs does not. We should align the initialization logic so Java DSL and XML DSLs does the same thing. They setup the pure model at first. So the configure method in the RouteBuilder should setup the model as the XML DSL would do. Then the prepare route logic which follows could be the same in all cases. This would also allow us to ensure when people use multiple RouteBuilder classes in Java DSL, then context scoped onException, interceptors is applied for all RouteBuilders.
Add OnException, Interceptor, etc. to JAXB model for a CamelContextDefinition (claus)
Configuring context scoped onException, interceptors etc. is woven into the RouteDefinition as part of the route initialization logic. When we have a dynamic routing engine (see above) that can at runtime support this without the need for woven into the routes. Then we should also ensure the context scoped onException, interceptors etc. is available in a CamelContextDefinition. This ensures the models is always 100% kept as it was provided, and we can fully export the model to XML and other languages (having a supported render).
Tighten up route definitions (claus)
Currently cross cutting concerns such as error handlers, interceptors, onCompletion etc. can be define anywhere in the route. We should tighten this up and only allow this to be configured in the start of the route. This also ensures when end users use code assistance in their route development, the IDE will not popup a big list which includes these cross cutting concerns. See also next note. (ProcessorDefinition will therefore be trimmed)
Message History EIP/Message Store (Christian Ohr)
This has been moved to its own Wiki page.
We should upgrade Jetty to 8.x as minimum. And if possible support Jetty 9.x which is in the works.
Currently we are stuck on 7.x due CXF / Karaf etc uses that old version, and thus we have been good citizen to align and use same version.
AMQ is also using older Jetty, but that is easier to upgrade as well.
We should avoid using the hostname in the JMX MBeans as its better to have a consistent naming that tooling and other parties can rely on. Having the hostname in there just add complexity to the mix. Also Camel may quote the MBean name for the CamelContextMBean and use " " in the mbean name, as the only mbean in there. (will need to double check exactly which mbean it was).
We should consider improve on this.
Remove not used components
We should consider removing
- org.apache.camel.view from came-core
- dot maven generator
- ... (there could be other stuff to remove)
The BAM has not changed in 5 years, and very seldom used by end users. And neither has the functionality you need. There is much better solutions outside ASF Camel for a BAM solution.
The DOT generator is not up to date and maintained. Also it requires binary generator to generate a route diagram; we never managed to find a good java library for that.
The MSV component is never/rarely used, and is causing some issues for cutting releases, due some weird maven issues / download of JARs etc. And the codebase has basically been left unchanged for 5+ years now.
Split camel-cxf into WS and REST
The camel-cxf component has grown too fat and has too many dependencies. People would like to use a light-weight RS. We have already talked on Camel @dev about splitting camel-cxf into a WS and RS modules. As well refactor the code-base as there is potential overlap with CXF itself and stuff to be removed/trimmed/optimized etc.
We can have a camel-cfx-core where we can have shared logic if that makes sense.
To be better defined and moved to the section above or removed
Support for asynchronous transactions
When using the asynchronous routing engine it would be desirable of transactional context could be propagated to the new threads.
This requires the TX manager supports suspend/resume on the TX. G.Nodet have worked a bit on this. See CAMEL-2902. Also see CAMEL-2729.
We could add support for using HawtDB as the persistent store for streams which overflow to disk store.
This might be implemented with the message store when it is used for stream caching.
The Resequencer EIP currently doesn't support persistence, we could introduce this and let it leverage HawtDB such as we did with the Aggregator2 EIP.
This might be implemented with the message store when it is used for temporarily saving exchanes until they are in order.
Schedule in DSL
We could consider adding DSL syntax sugar for scheduling routes. For example currently you have to use Quartz or a
ScheduledPollingConsumer which has the
delay option. We could add DSL which has something like:
The hard part is to come up with a good DSL syntax. We can look at BAM and see what we got there as well.
Advanced Scala support
A proposal for advanced Scala support is scalaz-camel which is an alternative to Camel's existing Scala DSL.
More EIPs as @annotations
Currently its only the Routing Slip, Recipient List and Dynamic Router which are avail as @annotation as well. We could add more EIPs as annotations such as Splitter.
And also maybe annotations for
AggregationStrategy to make this less Camel API dependent, so you can use a plain POJO for that.
-1: claus. We keep the JMX API and ppl can use jolokia for REST over JMX etc.
Currently the performance statistics is only avail when using JMX. We should allow those stats to be enabled regardless if JMX is enabled or not. Then we can use those stats from the web console. This also allows to expose those stats in the cloud where JMX is often not possible to be used.
camel-jpa component we could offer JPA annotated entities with the performance status (just as we do for the tracer). Then end user can more easily use that if they want the stats to be persisted in a database using JPA.
SEDA/VM components to leverage async routing engine
When testing your Camel apps with OSGi you may use PaxExam for that. We should create a test kit for osgi, like we have camel-test for regular junit testing. The test kit should make it easy for end users to have their apps tested with OSGi. We already have pieces in the
tests/camel-itest-osgi. We just need to clean and shape it up so its ready for end users as well. And of course add documentation as well.
And then we should use it in
camel-itest-osgi of course. See CAMEL-3777.
We already have REST support with CXFRS and Restlet but it can be better. We should make sure those components is dead easy to use and you can invoke REST services in one line of code etc. And we should make more examples and tidy up the CXFRS documentation.
More load tests (cmueller)
More load tests for frequently used Camel components (jetty, jms ...) and camel-core. If we have an defined runtime environment which is in general accessable/available for all users (like a public image on Amazon EC2) and a data store for the performance numbers, than we could easily collect and share the numbers.
- Ensure correct behavior under load
- Source for performance numbers (throughput etc).
- Detection of memory leaks
- Detection of performance decreases after refactorings
- create a single type converter registry available as OSGi service to all bundles serving as a single registry to lookup converters or to add/remove converters from custom bundles.
Easier commit/rollback for component developers
Maybe expose some interface having commit / rollback methods to make it easier for component developers to implement custom logic. Currently they may not know about OnCompletion and how to use
UnitOfWork on Exchange to do this today.
The implementation of DefaultUnitOfWork seems to have transformed itself into a to broad concern where unit of work is doing a bit more work than the transactional aspect that ties to its name.
Maybe this implementation should be named ExchangeContext and we can introduce a simpler UnitOfWork concept. This would also allow us to refactor the SubUnitOfWork into a general parent/child unit of work concept.
However this requires API changes and thus is best kept for Camel 3.0
This is also needed by the refactor of the Message History EIP.
Improvements to ThreadPoolProfile for thread management
We could move
org.apache.camel and have it in the root package.
Tighten up onException
We should consider tighten up the onException DSL a bit, for example to make it more clear that if you have processing steps in there, they only occur after the message has been exhausted. So maybe we need to have a onExhausted to make this clear
So in this example its only after the 3 failed redeliveries, then its exhausted, and then we handle the message and route it to a log where we ignore the message.
Also currently you can do this:
Which will handle the message and after this let the error handler deal with the message. See CAMEL-5059.
We should in this case use a NoopProcessor so the message is handled, but the regular error handler does not react. Then its consistent.
However we have not changed this in Camel 2.x to keep backwards compatibility.
Likewise there has been ideas to move onRedeliverRef to <redeliveryPolicy> as currently its to be configured outside the policy.
It may make more sense to move onRedeliverRef to the policy to keep it together.
Add composite EIP to compose a number of EIPs into a single EIP
It would be nice if we have a DSL to compose a number of child EIPs into a single EIP itself. eg a bit like <pipeline>. This would allow people to make it easier to group together a number of EIPs into a single "unit". This allows for example the error handler to redeliver to the composite EIP instead of at the point of failure within the group. There are use-cases where people want to do that. And today the solution is to split this into a new route, which you then disable error handler by setting it to no error handler. And then call the route using the direct endpoint. So instead if you could do <composite> ... stuff goes here </composite> then that would be neater, as you dont have to split into multiple routes. I think there is an old JIRA ticket created about this a long time ago.
Remove Spring transaction dependency from JMS component (Removed because SJMS will be non-spring version)
Refactor the JMS component to remove the dependency on Spring PlatformTransactionManager and replace with JTA TransactionManager.
Default to use Transacted mode by default
One thing I regret about the design of Camel was defaulting to non-transacted mode by default and forcing users to explicitly enable transactional model with a Transaction Manager.
It might be nice in Camel 3 to assume everything is transacted by default; and mimmick a form of transaction manager for non-transactional resource.
e.g. its pretty common to work with messaging like JMS and databases as a common set of endpoints; the easiest thing that could possibly work for end users is for those to just be transacted by default without the user having to grok how to tinker with component configurations & Transaction Manager registration.
Note I'm not advocating XA here; I think the default for camel should be for Idempotent Consumer and simple approach to middleware
Add trace information to the Camel exceptions
If and exception occurs, it would be helpful if the exception contains trace information like the passed endpoints/processors and the current endpoint/processor. We may could also extend the error handler to log (or not) this information.
That make it much easier to dig into the problem when we encounter an exception.
Avoid throws Exception on end user API
org.apache.camel.Processor should stay as is. This API has been like this since the first commit, and we should let it be backwards compatible.
But there is a number of end user APIs such as
Service, eg when you call
start on a CamelContext. It would be better to have that not throws Exception but use runtime exceptions if failed to start etc.
An API which uses unchecked exceptions is the
ProducerTemplate which end users uses. So its nice and easy to use.
Another example would be
MainSupport where the
run method throws exception etc.
There is possible other end user APIs which we can make a bit more friendly.