OSDir


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: A "Kamel" crazy idea


Le lun. 30 juil. 2018 à 17:49, Nicola Ferraro <ni.ferraro@xxxxxxxxx> a
écrit :

> Hi Cameleers,
> it seems from the comments that this "Kamel" subproject is something we
> want to start and I think that also the main camel core will benefit from
> the new features it will bring.
>
> I would like to donate the current "Kamel" code to Apache Camel, in order
> to have a initial brick to start from.
>
> From your reactions, the name "Kamel" seems cool and I'd like to keep it.
> However if you talk to people about the "Kamel" project, they won't
> understand if you're talking about "Kamel with the K" or "Camel with the
> C".
> For this reasons I propose to keep "Kamel", but also use "Camel K" [ˈkæməl
> keɪ] as friendly name when talking about it. This is in line with the
> repository we should create for the subproject, that needs to start with
> "camel-" (it will be "apache/camel-k" in that case).
>
> These days I've been experimenting different ideas with Luca Burgazzoli. In
> particular, there have been some concerns here on the power of a
> declarative DSL (a new one or the existing XML one) and we've found a
> strategy that will allow us to use also the Java DSL in Kamel integrations.
> Without adding too many details here, at integration build time we can run
> the user code in a build container and inspect the produced routes to get
> metadata for the optimizer. This is one of the first things we should do
> next.
>
> This possibility of running code that produces integrations opened a lot of
> unexpected paths, that I've written down in the project roadmap (
> https://github.com/nicolaferraro/integration-operator) and summarized
> below.
>
> One nice feature (kudos to Luca) is that we can simplify life for Kamel
> users up to the point that they'll just need to write their routes on a
> Java file and run them with e.g.:
> "kamel run Routes.java"
>
> Where "kamel" is a binary we release within the project. Kubernetes custom
> resources will be used under the hood, but the "kamel" binary is a utility
> that will provide a user experience comparable to that of current
> serverless platforms (or even better).
>
> But there's not just that.
>
> Luca wrote this weekend a prototype for having a polyglot Camel (
> https://github.com/lburgazzoli/camel-routes-loader).
>
> With that, you will be able to e.g. write integrations in groovy and use
> the same Kamel engine:
> "kamel run routes.groovy"
>
> But also JavaScript:
>
> https://github.com/lburgazzoli/camel-routes-loader/blob/ca986541f7c422ee02c21727cdfe4293d64a364e/src/test/resources/ext/camel/routes.js#L2
>
> This is a prototype right now, but a similar approach based on GraalVM has
> a lot of potential, because users can not only use their preferred language
> to write routes and processors, but also bind them to functionalities
> available in their preferred libraries.
>
> How much this will be feasible depends on the adoption of GraalVM, but I'm
> seeing many frameworks adding metadata to make GraalVM work with
> reflection. We've also done some work in Camel, there are some Jiras for it
> and first tests made by Guillaume signal that it's something feasible, at
> least for the Camel core and a subset of components. GraalVM would be
> important also to reduce memory footprint and improve startup time, as
> already said.
>

I'll continue experimenting and I'll report back.
I did some initial experiments leveraging some modifications I did  to
speed up the start up time [1] and on a single route [2].
There may be some limitations down the road of course, but at least it
shows that it's feasible. Fwiw, the experiments lead to a startup time of
14 ms compared to a few hundreds milliseconds (roughly 680 with [1]).


>
> If the GraalVM approach works (it is working for other frameworks), instead
> of rewriting a subset of Camel in Go (as the original proposal mentions),
> we can just sanitize and recompile our existing codebase: this way we fully
> leverage the strength of Apache Camel.
>

Right, if we can avoid rewriting Camel, that would be much better !

[1] https://issues.apache.org/jira/browse/CAMEL-12688
[2]
https://github.com/gnodet/openwhisk-runtime-camel/blob/master/camel-openwhisk-example/src/main/java/org/jboss/fuse/openwhisk/camel/example/SimpleCamelFunction.java


>
> Another important thing in the roadmap is actually how we'll integrate with
> current FaaS platforms. We will evaluate together if it's better to have a
> tighter integration with some of them or to leverage knative for some
> aspects of Kamel...
>
> But, let's get started!
>
> Il 19 lug 2018 10:45, "Nicola Ferraro" <ni.ferraro@xxxxxxxxx> ha scritto:
>
> It's clear to me that we need to add support for our existing XML DSL, that
> is powerful. But there are multiple reasons why I'd like to also "add" a
> limited yaml/json notation to "Kamel".
>
> The first one (and simplest) is that json/yaml is the primary encoding for
> all resources exchanged with the Kubernetes API server. I'm not saying that
> for this reason the Integration resource cannot contain a XML
> (/Java/Ballerina) section, but from a pure presentational point of view,
> having the possibility to write simple use cases (even a "hello world") in
> pure json/yaml is better than always requiring a mix of json and XML.
>
> The second one is simplicity. Writing a optimizer (the module that
> materializes the integration into running code, choosing a specific runtime
> and configuration) for a turing-complete language like Java is not easy:
> even if you manage to create a good parser, it may be able to do
> optimizations only if you write routes in a particular way, without complex
> constructs.
> XML is ok from this point of view. The thing is that both json/yaml and XML
> are just two different ways to serialize object trees, that can be then
> statically analyzed.
> The point is not XML vs json/yaml, it's more about tailoring a new
> minimalistic DSL to the emerging use cases vs proposing "only" our classic
> way of writing integrations. I think XML can be the "advanced" way. We can
> experiment optimizations easily with the new DSL, and enable them also on
> XML if it's worth.
>
> Scripting should be part of the spec, but I'd try to use programming
> languages only for processing/transformation, not for the route definition.
>
> Third one is performance. Apart from the fact that json parsers are in
> general said to be faster than XML parsers... Given the "fast startup"
> target that we want to reach, we may think e.g. to translate the new DSL
> into Java or Go code, then compiling it. This allows doing parsing at build
> time in order to avoid it on startup. This kind of improvements are much
> easier with a limited DSL but much more difficult with a existing
> fully-fledged DSL..
>
> Nicola
>
>
>
> On Wed, Jul 18, 2018 at 11:51 PM Pontus Ullgren <ullgren@xxxxxxxxx> wrote:
>
> > I also like the idea but with some comments.
> >
> > As Hiram Chirino I'm not sure YAML/JSON is the best language for this.
> > Perhaps a more fluent DSL like the Camel Java DSL or perhaps something
> like
> > Ballerina language would be better suited ?
> > Also in my experience even simple integrations, that is simple real world
> > integration and not just hello world demos, requires
> > you to add one or more Java class or scripting in addition to the core
> > components. So for it to be useful there must be some way to add custom
> > code in some way for aggregation/enrichment strategies.
> >
> > If you go with a GO implementation I would also like some fallback for
> > using Java since I find it unlikely that the existing base of components
> > will be ported any time soon.
> >
> > As for the name as other I like the "Kamel"  name however a being a
> newborn
> > kotlin fan I know of one pre-existing project with that name
> > that is close enough to allow for potential confusion.
> > https://github.com/topicusoverheid/kamel
> >
> > Looking forward to see the result.
> >
> > // Pontus
> >
> >
> >
> >
> >
> >
> > On Mon, 16 Jul 2018 at 15:15 Hiram Chirino <hiram@xxxxxxxxxxxxxxxx>
> wrote:
> >
> > > Love the idea.
> > >
> > > Personally, I'd keep using the existing Camel XML DSL if possible.  You
> > can
> > > still embed it in the CRD.  The operator that deploys it could inspect
> it
> > > and figure whats the most optimized runtime that can support the DSL.
> > > Perhaps if it's only using the restricted set of camel components
> > supported
> > > by camel-go (https://github.com/lburgazzoli/camel-go) then is uses
> that.
> > > Otherwise it falls back to using camel spring boot.
> > >
> > > For that to work I think we need the different runtime implementations
> to
> > > provide a way to ask them: 'hey do you support running this camel
> route?'
> > > Not a trivial thing to respond to, it might require a build step in
> there
> > > for traditional Camel runtimes.
> > >
> > >
> > > On Fri, Jul 13, 2018 at 4:32 AM Antonin Stefanutti <
> > antonin@xxxxxxxxxxxxx>
> > > wrote:
> > >
> > > > Hi Nicola,
> > > >
> > > > I love the idea.
> > > >
> > > > I just wonder whether YAML/JSON is an expressive enough format in the
> > > long
> > > > term. But as you’ve mentioned, starting simple would enable
> > experimenting
> > > > some very interesting / promising optimisations. So it seems worth
> > taking
> > > > that path, instead of trying to embed a complex DSL or the existing
> XML
> > > DSL
> > > > into the CRD.
> > > >
> > > > Definitely +1
> > > >
> > > > > On 13 Jul 2018, at 01:30, Nicola Ferraro <ni.ferraro@xxxxxxxxx>
> > wrote:
> > > > >
> > > > > Hi Cameleers,
> > > > > it's now passed some time since I started thinking about a new
> > project
> > > > that
> > > > > we can begin here at Apache Camel, and I'd like to have your
> opinion.
> > > > >
> > > > > We've already been targeting cloud-native applications with Camel,
> > > > > especially on top of Kubernetes, that is becoming "the standard"
> > cloud
> > > > > platform. But writing a Camel integration and running it on
> > Kubernetes
> > > > > requires some effort: choosing the base platform (spring-boot,
> karaf,
> > > > > simple main?), adding health checks (actuator?), packaging a docker
> > > image
> > > > > and creating the Kubernetes resources (fabric8-maven-plugin,
> helm?),
> > > > > publishing the image on a docker registry, then finally deploying
> the
> > > > > resources on a Kubernetes cluster.
> > > > >
> > > > > The resulting integration container is then far from being optimal
> > > from a
> > > > > resource consumption point of view: it is likely that a Camel
> > > Spring-Boot
> > > > > application will require at least 200MB of RAM and also some CPU
> > shares
> > > > > because of polling threads used by many components.
> > > > >
> > > > > In case people use a CI/CD pipeline, it will take also a long time
> to
> > > get
> > > > > from a code update to having a Kubernetes POD up and running.
> > > > > Apart from compilation and image push/pull time, also startup time
> is
> > > > often
> > > > > ~10 seconds for Camel + Spring-Boot in a container with standard
> > limits
> > > > on
> > > > > resources, making it difficult to propose this combination for
> > > > "serverless
> > > > > integration" (this term is becoming increasingly more popular).
> > > > >
> > > > > So, my proposal is to start to investigate a "more cloud-native"
> > > approach
> > > > > to integration: *making Camel integrations first-class citizens in
> > > > > Kubernetes, and making them super fast and lightweight.*
> > > > >
> > > > > We can base the project on Kubernetes Custom Resource Definitions
> > (CRD)
> > > > > <
> > > >
> > >
> >
> https://kubernetes.io/docs/concepts/extend-kubernetes/api-extension/custom-resources/
> > > > >,
> > > > > for example a Integration CRD and have a Kubernetes "operator"
> > > > > <https://coreos.com/operators/> taking care of:
> > > > > - Optimizing the integration that we want to run
> > > > > - Packaging in a container
> > > > > - Running it on Kubernetes
> > > > > - Managing its entire lifecycle
> > > > >
> > > > > A Kubernetes-native integration may look like:
> > > > >
> > > > > -------------------
> > > > > kind: "Integration"
> > > > > apiVersion: "camel.apache.org/v1alpha1"
> > > > > metadata:
> > > > > name: "example"
> > > > > spec:
> > > > > replicas: 1
> > > > > routes:
> > > > >  - id: timer
> > > > >    route:
> > > > >     - type: endpoint
> > > > >       uri: timer:tick
> > > > >     - type: endpoint
> > > > >       uri: log:info
> > > > > -------------------
> > > > >
> > > > > For those who are not familiar with Kubernetes resources, this kind
> > of
> > > > > YAML/JSON resource definitions are really common.
> > > > > The example route is embedded in the Kubernetes resource
> declaration
> > > and
> > > > > follows a basic "flow DSL". We may start from a basic one and
> evolve
> > it
> > > > as
> > > > > new requirements arrive from the community.
> > > > >
> > > > > I've made a very simple (but working) POC here:
> > > > > https://github.com/nicolaferraro/integration-operator.
> > > > >
> > > > > This idea of a "Cloud-Native Camel" on Kubernetes (project codename
> > can
> > > > be "
> > > > > *Kamel*", if you like it :D), will be an enabler for a lot of nice
> > > > features.
> > > > >
> > > > > For example, we can propose "Kamel" as "ideal" platform for
> > "serverless
> > > > > integration" (I see many people reinventing the wheel out there):
> the
> > > > > operator can reduce resource consumption of a single integration by
> > > > > optimizing the runtime and also pause/resume integrations when they
> > are
> > > > not
> > > > > used, that is the basic idea behind "serverless" (e.g. think to
> > > > > HTTP-triggered integrations, but not only).
> > > > > Focusing on serverless will bring more emphasis on push-based
> > > > notifications
> > > > > (webhooks, cloud events <https://cloudevents.io/>), that are
> rarely
> > > > used in
> > > > > Camel components, that prefer a poll based approach being it
> simpler
> > to
> > > > use
> > > > > in classic deployments, but not so good in the cloud, where more
> > > > resources
> > > > > become higher direct costs for the users.
> > > > >
> > > > > The presence of the simplified DSL enables also experimenting on
> > > > "*reduced*
> > > > > subsets of Camel" implemented in languages other than Java, for
> > example
> > > > one
> > > > > language that has a reactive approach on thread scheduling and a
> > really
> > > > low
> > > > > memory footprint, like Go.
> > > > >
> > > > > But apart from this kind of experiments (that are valid IMO), the
> > > "Kamel"
> > > > > optimizer will have free room to choose the right platform for the
> > > > > integration that the user wants to run, including, in the future,
> > doing
> > > > AOT
> > > > > compilation using Graal/VM (less memory, faster startup) if the
> > > features
> > > > > (components) used in the integration are supporting it (maybe we
> can
> > > add
> > > > > AOT compilation in the roadmap for Camel 3).
> > > > > A silly optimization: integrations starting from "timer:..." may be
> > > > > scheduled directly with Kubernetes CronJobs, so they will consume
> > > > resources
> > > > > only when actually running.
> > > > >
> > > > > Being the final integrations lightweight and being the DSL
> > > > > language-independent, we may see a increased adoption of Camel also
> > as
> > > > > agile integration layer for not-only-java applications (both
> "cloud"
> > > and
> > > > > "serverless" applications).
> > > > >
> > > > > I'm the first one that would like to work on a project ilke this.
> > I've
> > > > > worked on many Kubernetes/Openshift based applications and
> frameworks
> > > in
> > > > > the past years, also on operators and CRDs, and I think this way of
> > > > > redesigning integrations has a lot of potential.
> > > > >
> > > > > Integrations will not be necessarily limited to the simplified DSL,
> > but
> > > > we
> > > > > can add extension points for scripting and even custom libraries
> > > > (although
> > > > > limiting the freedom of the optimizer).
> > > > >
> > > > > The most important thing: it may become a great project, since it's
> > > > driven
> > > > > by a great community.
> > > > >
> > > > > So, what do you think? Is it crazy enough?
> > > > >
> > > > > Nicola
> > > >
> > > >
> > >
> > > --
> > > Hiram Chirino
> > > Engineering | Red Hat, Inc.
> > > hchirino@xxxxxxxxxx | fusesource.com | redhat.com
> > > skype: hiramchirino | twitter: @hiramchirino
> > >
> >
>


-- 
------------------------
Guillaume Nodet