fineract-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Keith Woodlock <keithwoodl...@gmail.com>
Subject Re: Microservices
Date Mon, 23 May 2016 09:16:11 GMT
Mryle,

>> Now please excuse me while I get back to just doing it. ; o)

Ok sounds like decision made that yourself and Markus will be going ahead
with this :)

Can you let us know where you will be checking in the code so we in the
community can follow along.

Given your are going to go ahead with developing a proof of concept
starting from a blank canvas it might be good to layout a scope and some
goals to get to in a time-boxed period of time. I would suggest something
like:

1) A single microservice implemented from API to Database including
deployment related capabilities

The emphasis on this would be to show off the tech approach/pattern
involved in implementing a microservice over the domain of the microservice
so feel free to choose a simpler area

This hopefully could be provided in a much shorter time-line (a month or so)

2) A microservice that encapsulates the full Loan Product / Loan area of
Fineract

Makes sense to takcle the meat of the platform to show off qualities of
microservices approach

By the end of this you will have two microservice examples and that will
bring up the questions around deployment and service discovery and how that
will be implemented.

One last question would be to ask if Markus or yourself have actually
developed/being part of a project previously that has used the microservice
approach specifically you are talking about?

If yes thats great, if no, is it worth seeking someone with some expertise
in the area of microservices specifically in java area?

regards,
Keith.


On Mon, May 23, 2016 at 8:26 AM, Myrle Krantz <mkrantz@mifos.org> wrote:

> Hi,
>
> I won't be able to write a complete answer.  Nobody would read an e-mail
> that long.  So I’m going to “nibble” at it bit by bit.
>
> Of the three approaches Keith lists, I favor the third approach "Re-imagine
> the fineract platform from ground up again (hope it solves the problems you
> see and doesnt introduce new problems)".  I've been in the industry long
> enough (16 years) to understand that there are significant risks in
> "starting over from scratch", and I hope that we can learn enough from the
> existing code base to avoid some of those road bumps.
>
>
> I would imagine multiple phases here:
>
> 1.) Markus and I start work on a framework, and basic services for this
> refactoring.  Everyone else continues work on MifosX
>
> 2.) We do a proof of concept with a bare bones implementation.
>
> 3.) Report to the community based on what we learn by end of this year.
>
> 4.) Based on what we have learned, together with the community consider
> migration of:
>
>    a.) New microservice-based code into the Fineract community.
>
>    b.) Community man-power from the existing codebase to the new code base.
>
>    c.) Existing customer data into the new microservice-based code via
> migration scripts.
>
>    d.) Existing projects onto the new microservice-based product.
>
>    e.) The community app onto the new microservice-based backend.
>
> We can do a subset of the above migration steps.  There are several
> possible approaches.  But I really want to do this together with the
> Fineract community.  There’s much know-how here, and I think we can copy
> code over wholesale in many areas, and in other areas plagiarize good
> architectural decisions and problem-solving approaches which the existing
> Fineract codebase illustrates.
>
> Because I want to do this with the Fineract community, I want to start
> getting people in the community up to speed on the concepts early.  Even if
> people decide they don't like this approach, nobody will be poorer for the
> extra information.  I also very much want people’s feedback on where they
> see problems with our approach. I haven’t given much in the way of
> specifics yet; I want people to have some background first. Which was the
> reason for my original mail.
>
> Now please excuse me while I get back to just doing it. ; o)
>
> (My next email will come tomorrow, and will focus on Roman’s very important
> deployment question.)
>
> Greets from the Voreifel, Germany,
> Myrle
>
>
> *Myrle Krantz*
> Solutions Architect
> RɅĐɅЯ, The Mifos Initiative
> mkrantz@mifos.org | Skype: mkrantz.mifos.org | http://mifos.org
> <http://facebook.com/mifos>  <http://www.twitter.com/mifos>
>
>
> On Thu, May 19, 2016 at 11:49 PM, Keith Woodlock <keithwoodlock@gmail.com>
> wrote:
>
> > Hi Mryle,
> >
> > Thanks for your detailed reply, I added my comments (please remember I am
> > not involved in code base for sometime so might be missing alot) below
> but
> > its a pretty long email so I summarise what I think here:
> >
> > The decision to move from one approach to developing a system to another
> is
> > a big one and not without its costs. I am not really for or against it
> and
> > I'll be interested in seeing how things turn out whichever way the
> > community decides to go. If everyone really feels it will be the solution
> > to any problems that exist then they should go for it. The risk of course
> > is that its a very tech solution to things.
> >
> > I think Chris Richardson does a good job on http://microservices.io of
> > outlining the differences between the monolith (
> > http://microservices.io/patterns/monolithic.html) and the microservices
> (
> > http://microservices.io/patterns/microservices.html) approaches.
> >
> > You probably have the following 3 directions you can go:
> >
> > 1) Stay with the monolith and attempt to tackle the problems you see with
> > the system (domain, performance etc)
> >
> > 2) Gradually move services from the monolith into microservices (hope it
> > solves the problems you see and doesnt introduce new problems)
> >
> > 3) Re-imagine the fineract platform from ground up again (hope it solves
> > the problems you see and doesnt introduce new problems)
> >
> > - Approach 2 sounds like it will be a drawn out affair and take a long
> time
> > to reach the end with the drawback of having to support both approaches
> >
> > - Approach 3 it too would be alot of effort but you could use something
> > like http://eventuate.io/ to get to a poof of concept faster
> >
> >
> >
> >
> > Before devling into some of the points you raised I wanted to highlight
> two
> > aspects of the fineract platform:
> >
> > *- Everything is exposed a HTTP API / Service*
> >
> > All the functionality that exists in the platform is or should be exposed
> > by a HTTP API. This HTTP API is a techonology agonistic contract that
> > developers can build their apps/clients on or integrate with easily.
> >
> > Developers can choose to build their apps/clients in any technology they
> > wish and to use as much or as little of the platform API in their app.
> >
> > API documented here https://demo.openmf.org/api-docs/apiLive.htm
> >
> > *- All HTTP API Endpoints deployed together*
> >
> > The code is compiled and bundled as a .war file that is deployed to an
> > application server like tomcat.
> >
> > If one application server isnt able to handle the request load you can
> > easily scale horizontally by adding more application servers (as the
> > platform is stateless) and load balance the requests between the
> serrvers.
> > This would typically only be needed in a larger scale deployment.
> >
> >
> > *The Problems You Listed*
> >
> >
> >
> >
> > *1.) We are starting to have performance problems.  Conflux in particular
> > is running into this.  Mussoni is starting to have problems too.  All of
> us
> > would like to move to bigger customers and bigger markets, so this will
> > only get worse.*
> >
> > You should open up a new email thread within the community on these
> issues.
> >
> > It would be useful to provide some more details on what API requests in
> > praticular Conflux/Musoni and others are seeing performance problems
> with.
> >
> >
> > *2.) As we move to bigger customers, down-time becomes less and less
> > acceptable.*
> >
> > There shouldnt need to be a need for downtime with the existing
> > deploy-all-services-as-one approach.
> >
> > On tomcat you can deploy your new application in parallel with your old
> > one.
> >
> >
> >
> >
> >
> >
> > *3.) Loans products are hard-coded.  A lot of dev list mails read like
> this
> > “For my case, I want my interest/fees calculated at a different time/ on
> a
> > different basis than the ‘standard’ use case, how do I do this?”  If I
> were
> > a customer asking a question like this, some of the hacks being suggested
> > would feel brittle and unsatisfying.  Savings products look like they
> were
> > implemented as an afterthought and are also hard-coded.*
> >
> > Sounds likes a domain modelling / requirements issue or an implementation
> > issue.
> >
> > Different financial institutions in different countries have many
> different
> > flavours of the way they want their loan or saving products to work and
> its
> > an area that will continue to change into the future.
> >
> > Mifos started out with supporting one or two of the common models in use
> in
> > the field and it looks like more have been added in since based on
> content
> > in http://fineract.incubator.apache.org/
> >
> > You might be suggesting though that its time to revisit the Loan Product
> /
> > Loan area and re-model it based on what you know today?
> >
> >
> >
> >
> >
> >
> > *4.) Our adherence to the CQRS pattern is less than perfect.  Command
> > responses were hard coded.  Some work has been done to make proper
> command
> > handlers out of it, but there are still semi-commands handled in a big if
> > else block.  Handling commands is synchronous even though one of the
> > advantages of this pattern should be the ability to handle operations
> > asynchronously.  This contributes to the problems mentioned in 1 above.*
> >
> > Not sure what the hard-coded command response area is (been along time
> > since looked at code base)
> >
> > Its true that code that is responsible for identifying what command
> handler
> > to invoke based on the command is not great and suffered from the 'god
> > class' problems - the inital core of the platform was written in a
> > time-boxed three month period as proof of concept and as more
> apis/commands
> > got added it got bigger.
> >
> > That said, if you have a preferred solution its surely an area that could
> > of been improved in the last few years if it was that important?
> >
> > On the synchronous nature of them, yes we decided rather than going for
> an
> > eventually-consistent system that the api calls would be like a
> datatabase
> > transaction call, i.e when you get a response you know its either
> > successful or not
> >
> > Calling an api 1000's of times to do something in bulk of course would
> not
> > make much sense and you would expect to add a bulk api for the that type
> of
> > activity.
> >
> > Also if there are very long running tasks then of course they should not
> be
> > done through a synchronous api call, you should set up an api that
> supports
> > 'long running tasks' and pass back a id/key that the user can use to
> check
> > for the result later or something similiar.
> >
> >
> >
> >
> > *5.) There’s currently no standard way to connect to what I like to call
> > “money buses” (Swift, Stellar, UPI).  I did implement a bridge for
> Stellar
> > which listened for commands, but it wasn’t clear which commands to listen
> > for.  I ended up deciding on journal entries.*
> >
> > Not fully across what you are doing here but if its a case of integrating
> > some kind of third party payments api and invoking it based on some api
> > call like makePayment or disburse then that should be catered for by the
> > platform. If you werent clear on how to do its probably cause the
> platform
> > didnt try to support that type of integration out of the box first of
> all.
> >
> > Listening to journal entries might not be the best idea cause what If i
> > dont want to use the platforms accounting solution and turn it off?
> >
> > This use case should be designed into the praticular area in question.
> >
> >
> >
> >
> > *6.) In first world core banking systems, the back end and the front end
> > are carefully separated with clearly defined and limited communication.
> > This is to provide added protection against hacking of the type we’ve
> seen
> > in the news recently (Bangladesh’s central bank for example).*
> >
> > The front end (apps/clients) are totaly seperated from the back-end api
> at
> > present. In terms of limiting what apps can speak to what APIs, you
> > probably need to add that capability on top of the current
> > authentication/authorization capablity.
> >
> > At present there is fine-grained support for restricting what apis a user
> > is entitled to call, its not to much more to also limit what apis can be
> > called based on what app/client is being used.
> >
> > *How these things related to microservices pionts*
> >
> >
> >
> >
> >
> >
> > *1.) By chopping up database accesses into logical units, we can increase
> > capacity for only those units which need it.  This reduces costs.  In
> > particular, for the saving and reading of commands (for which we expect
> the
> > largest volume of write instructions), we can switch to a faster
> > persistence mechanism which works better in a distributed environment.*
> >
> > Your probably talking about switching to storing the commands/events in a
> > different persistence mechanism than a relational database. You could
> > choose to do that now if you wish.
> >
> > You might also be referring to breaking the current database schema into
> a
> > database per service approach (
> http://microservices.io/patterns/index.html
> > ).
> >
> >
> > I am not sure how this would reduce costs. Like any solution it has its
> > pros and cons (
> > http://microservices.io/patterns/data/database-per-service.html)
> >
> >
> >
> > *2.) We can run multiple instances of a microservice, and, with careful
> > design, perform rolling updates, moving us towards the ability to have
> > 99.99% uptime.*
> >
> > Much like point 2 earlier,  you should be able to do this with the
> current
> > setup.
> >
> >
> >
> >
> >
> > *3.) A more modular approach with well-defined APIs would make it
> possible
> > among other things to separate out the loans and savings products and let
> > anyone implement their own loans and savings products who wants to.
> Things
> > like interest rates, fees, and repayments rates are very cultural, and as
> > we move into more countries we’ll need more flexibility here.*
> >
> > There should be well defined APIs at present. If you have a solution for
> > allowing people to create their own loans and savings products with
> > complete flexibiliy you should be able to roll it out now in the current
> > setup.
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > *4.) CQRS is a truly awesome pattern for this domain, partly because it
> in
> > theory, makes it possible to handle writes asynchronously.  If we can
> > decouple the commands for different domains from each other and actually
> > handle them asynchronously that would be a major step forward.  Of course
> > this doesn’t necessitate a microservices deployment model, at least not
> > until the question of what we do with the data we create while handling
> the
> > commands.  If, for example, a command for creating a user causes
> different
> > data to be adjusted then a command for starting a payment, why should
> > creating a user cause a slowdown in working through payments?  Those two
> > kinds of processing don’t even have to execute on the same hardware...*
> >
> > As you say, if you want to implement a more fully blown implementation of
> > the CQRS+Event Sourcing pattern you could implement that now for the
> areas
> > your believe need it and still respect the existing API
> >
> > The example you given isnt a good one as you wouldnt expect there to be a
> > high throughput of users created but I understand that your saying that
> if
> > there was 'intensive' operations going on on the platform, it might be
> > usefult to be able to split them into a seperate unit that runs on
> seperate
> > hardware so it doesnt impact more important functionality that might have
> > an SLA on it like payment processing etc
> >
> > The question is, what are these intensive operations that might be
> > affecting other areas of the platform so badly?
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > *5.) Writing ‘money bus’ plugins as microservices is a great way to
> > separate functionality and be sure that even if one bus is down, the
> others
> > can still be up and running.  My stellar bridge component essentially
> > applied this pattern.  If we can standardize the interface that bridge
> > components implement, we could make these easier to write and improve
> Mifos
> > connectivity to other systems.  These don’t need to be a standard part of
> > deployment, in different countries, different networks are popular, and
> > these things change over time too.  Connectivity to other systems is, in
> my
> > humble opinion, one of the most important factors in the success of a
> > product.*
> >
> > Yes its very important to be able to integrate with other systems and I
> > dont see why the kind of functionality you want to implement and
> integrate
> > into mifos/fineract cant be done today in a clean way. I dont see how
> > having all the current services split out into deployable units helps
> this
> > use case.
> >
> >
> >
> >
> >
> >
> >
> >
> > *6.) Separation of core systems from customer-visible systems will
> require
> > a duplication of some functionality.  This functionality would need to
> run
> > in separate processes and only have access to data duplicated out of the
> > back end system.  Most data would only flow from back to front (probably
> > via some kind of ETL process).  If we divvy up our data based on the
> > functionality groups which need it, it will be much easier to determine
> > which data should be copied “forward” and which should “stay behind”.  A
> > microservice is such a functionality group.*
> >
> > Its probably not a good idea to mix customer-visble functionality with
> core
> > business-functionality. It terms of large financial institutions you
> would
> > be talking about 100,000s to millions of potentional users for customer
> > visible functionality whilst the core business-functionality would be
> > significantly less.
> >
> > So they really are two different systems at some level.
> >
> > regards,
> > Keith.
> >
> >
> > On Thu, May 19, 2016 at 2:07 PM, Myrle Krantz <mkrantz@mifos.org> wrote:
> >
> > > This is a big topic, and I’m trying to decide if I should chop it up
> into
> > > multiple threads.  I’m going to leave it as one for now, since I
> detect a
> > > healthy scepticism about the question of whether we need to do this at
> > all,
> > > and I don’t want to preclude that discussion.  Here are some of the
> > > problems I’m seeing:
> > >
> > > 1.) We are starting to have performance problems.  Conflux in
> particular
> > is
> > > running into this.  Mussoni is starting to have problems too.  All of
> us
> > > would like to move to bigger customers and bigger markets, so this will
> > > only get worse.
> > > 2.) As we move to bigger customers, down-time becomes less and less
> > > acceptable.
> > > 3.) Loans products are hard-coded.  A lot of dev list mails read like
> > this
> > > “For my case, I want my interest/fees calculated at a different time/
> on
> > a
> > > different basis than the ‘standard’ use case, how do I do this?”  If
I
> > were
> > > a customer asking a question like this, some of the hacks being
> suggested
> > > would feel brittle and unsatisfying.  Savings products look like they
> > were
> > > implemented as an afterthought and are also hard-coded.
> > > 4.) Our adherence to the CQRS pattern is less than perfect.  Command
> > > responses were hard coded.  Some work has been done to make proper
> > command
> > > handlers out of it, but there are still semi-commands handled in a big
> if
> > > else block.  Handling commands is synchronous even though one of the
> > > advantages of this pattern should be the ability to handle operations
> > > asynchronously.  This contributes to the problems mentioned in 1 above.
> > > 5.) There’s currently no standard way to connect to what I like to call
> > > “money buses” (Swift, Stellar, UPI).  I did implement a bridge for
> > Stellar
> > > which listened for commands, but it wasn’t clear which commands to
> listen
> > > for.  I ended up deciding on journal entries.
> > > 6.) In first world core banking systems, the back end and the front end
> > are
> > > carefully separated with clearly defined and limited communication.
> This
> > > is to provide added protection against hacking of the type we’ve seen
> in
> > > the news recently (Bangladesh’s central bank for example).
> > >
> > > So how does this relate to Microservices?  From here on out, I’m going
> to
> > > assume you all know what a Microservice is, so I hope you read those
> > links
> > > in my original e-mail : o).
> > >
> > > 1.) By chopping up database accesses into logical units, we can
> increase
> > > capacity for only those units which need it.  This reduces costs.  In
> > > particular, for the saving and reading of commands (for which we expect
> > the
> > > largest volume of write instructions), we can switch to a faster
> > > persistence mechanism which works better in a distributed environment.
> > > 2.) We can run multiple instances of a microservice, and, with careful
> > > design, perform rolling updates, moving us towards the ability to have
> > > 99.99% uptime.
> > > 3.) A more modular approach with well-defined APIs would make it
> possible
> > > among other things to separate out the loans and savings products and
> let
> > > anyone implement their own loans and savings products who wants to.
> > Things
> > > like interest rates, fees, and repayments rates are very cultural, and
> as
> > > we move into more countries we’ll need more flexibility here.
> > > 4.) CQRS is a truly awesome pattern for this domain, partly because it
> in
> > > theory, makes it possible to handle writes asynchronously.  If we can
> > > decouple the commands for different domains from each other and
> actually
> > > handle them asynchronously that would be a major step forward.  Of
> course
> > > this doesn’t necessitate a microservices deployment model, at least not
> > > until the question of what we do with the data we create while handling
> > the
> > > commands.  If, for example, a command for creating a user causes
> > different
> > > data to be adjusted then a command for starting a payment, why should
> > > creating a user cause a slowdown in working through payments?  Those
> two
> > > kinds of processing don’t even have to execute on the same hardware...
> > > 5.) Writing ‘money bus’ plugins as microservices is a great way to
> > separate
> > > functionality and be sure that even if one bus is down, the others can
> > > still be up and running.  My stellar bridge component essentially
> applied
> > > this pattern.  If we can standardize the interface that bridge
> components
> > > implement, we could make these easier to write and improve Mifos
> > > connectivity to other systems.  These don’t need to be a standard part
> of
> > > deployment, in different countries, different networks are popular, and
> > > these things change over time too.  Connectivity to other systems is,
> in
> > my
> > > humble opinion, one of the most important factors in the success of a
> > > product.
> > > 6.) Separation of core systems from customer-visible systems will
> > require a
> > > duplication of some functionality.  This functionality would need to
> run
> > in
> > > separate processes and only have access to data duplicated out of the
> > back
> > > end system.  Most data would only flow from back to front (probably via
> > > some kind of ETL process).  If we divvy up our data based on the
> > > functionality groups which need it, it will be much easier to determine
> > > which data should be copied “forward” and which should “stay behind”.
> A
> > > microservice is such a functionality group.
> > >
> > > There are other advantages to microservices that are more general in
> > nature
> > > and not specific to our use case; read the links I sent in my first
> > e-mail
> > > and you should start to get a feel for it.
> > >
> > > (Roman, I’m going to watch and see if your questions are answered as we
> > > discuss.  If not I’ll circle back to them, and see what was missed.)
> > >
> > > Greets,
> > > Myrle
> > >
> > >
> > > *Myrle Krantz*
> > > Solutions Architect
> > > RɅĐɅЯ, The Mifos Initiative
> > > mkrantz@mifos.org | Skype: mkrantz.mifos.org | http://mifos.org
> > > <http://facebook.com/mifos>  <http://www.twitter.com/mifos>
> > >
> > >
> > > On Thu, May 19, 2016 at 1:01 AM, Keith Woodlock <
> keithwoodlock@gmail.com
> > >
> > > wrote:
> > >
> > > > Hi Myrle,
> > > >
> > > > First off, thanks for starting a conversation around this within the
> > > > fineract dev community. I was wondering when this was going to come
> up.
> > > >
> > > > Rather than delving into the technicalities of what a microservice is
> > and
> > > > isnt I thought it would be good if the people proposing this could go
> > > into
> > > > a bit of detail here of why they are proposing it, the pros and cons
> > of a
> > > > microservice approach over the existing approach and the existing
> > > problems
> > > > within fineract/mifos that it will help solve/remove.
> > > >
> > > > I think that could promote some good discussion and following on from
> > > that
> > > > later we could delve into how it could be achieved if going that
> > > direction.
> > > >
> > > > regards,
> > > > Keith.
> > > >
> > > >
> > > > On Wed, May 18, 2016 at 11:14 PM, Roman Shaposhnik <
> > roman@shaposhnik.org
> > > >
> > > > wrote:
> > > >
> > > > > Hi Myrle,
> > > > >
> > > > > microservices are indeed a nice architectural pattern with an
> > > > exponentially
> > > > > growing adoption in the enterprise. It typically leverages some
> kind
> > > of a
> > > > > PaaS
> > > > > solution to be available for deployment. My current understanding
> is
> > > that
> > > > > your
> > > > > deployment model today is a traditional appserver. Are you going
to
> > > > change
> > > > > that?
> > > > > If so, what will be the deployment requirements for the project?
> > > > >
> > > > > Thanks,
> > > > > Roman.
> > > > >
> > > > > On Wed, May 18, 2016 at 2:35 AM, Myrle Krantz <mkrantz@mifos.org>
> > > wrote:
> > > > > > Hi all,
> > > > > >
> > > > > > As many of you know Markus and I are thinking through a
> > > rearchitecting
> > > > of
> > > > > > Fineract into microservices. Markus gave some of you a preview
> into
> > > our
> > > > > > work at the Mifos tech conference in Amsterdam in March (those
of
> > you
> > > > who
> > > > > > are interested and weren’t able to attend can check it out
on
> > Youtube
> > > > > here:
> > > > > > https://www.youtube.com/watch?v=iIrnZpoNZ9A -- my apologies
for
> > the
> > > > poor
> > > > > > video quality)
> > > > > >
> > > > > > If you want to get an introductory understanding of what a
> > > microservice
> > > > > is,
> > > > > > this is a good place to go:
> > > > > >
> > > > > > http://martinfowler.com/articles/microservices.html
> > > > > >
> > > > > > If there is interest, (and I have time), I’ll try to send
you
> more
> > > > > > information about my ideas as they take shape.
> > > > > >
> > > > > > Greets,
> > > > > >
> > > > > > Myrle
> > > > > >
> > > > > > P.S.  Here’s Spring’s take on this:
> > > > > > https://spring.io/blog/2015/07/14/microservices-with-spring
> > > > > >
> > > > > >
> > > > > > *Myrle Krantz*
> > > > > > Solutions Architect
> > > > > > RɅĐɅЯ, The Mifos Initiative
> > > > > > mkrantz@mifos.org | Skype: mkrantz.mifos.org | http://mifos.org
> > > > > > <http://facebook.com/mifos>  <http://www.twitter.com/mifos>
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message