YT YouTube2Text

YouTube2Text → Transcripts → Anyone can build a bank: creating a new banking backend by Vlad Yatsenko

Anyone can build a bank: creating a new banking backend by Vlad Yatsenko — Transcript

by Devoxx · 7,047 words · 1,019 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00[Music]
0:05- thanks for coming over if it's a small
0:09room then you know the number of people
0:11want to attend so thanks thanks a lot
0:13for coming over even taking the space on
0:16on the floor it's so my name is Vlad
0:21Sankar like typically shot usually Lots
0:25it's hard to pronounce my full name for
0:27mature people
0:28I'm nan confounding city or evolute just
0:34the quick audience check how many people
0:36have heard of reveal it like crazy okay
0:41this is a good so I don't have to do
0:42long introduction of what River is today
0:47I'm gonna talk about building a bank in
0:51general what kind of challenges is there
0:52I'm gonna go even on the code level as
0:56well not too much in order to develop a
0:59conference so it's kind of a balance
1:03between domain and take some technical
1:05problems that's a typical for this
1:07domain as well so let's go so what is a
1:13bank in general in the past we used to
1:18like people used to imagine a bank like
1:21this building with insured columns and
1:25because banks needed to portray that
1:28they are very safe that can be trusted
1:29people bring money very physically and
1:33essentially there was the main purpose
1:35and then they started doing lending
1:37borrowing etc but this thing is changing
1:42right because it was very in efficient
1:44you had to cross distances you had to do
1:47a lot of physical content
1:49these days bank banks realizing that
1:52this is not working because of the wave
1:54of FinTech and we see a lot more of this
1:58happening with the old banks and it's
2:04because of rise of FinTech that happen
2:06due to different factors due to the
2:10financial crisis in 2008
2:13that regulate this allows me change the
2:16rules in such a way so that allows new
2:19engines to come into this market and we
2:23use that opportunity so but you know
2:26it's important to think what was the
2:29purpose of the bank in general right
2:31obviously like if it's of any almost any
2:34business is to make money as mentioned
2:39in the past people used to use banks as
2:42just trusted vault where they put money
2:45instead of keeping that money under
2:46their mattress and then the main
2:49business of the banks and it is still is
2:52boring at one extreme interest rates
2:56from some people and lending that money
3:00at high interest rates to other people
3:02and the business basically making this
3:04difference on on the interest rate these
3:10days with the thing it's etc and in its
3:14financial institutions yeah we do a lot
3:17moment not just borrowing and lending
3:20how we think about banking today it's
3:23card payments currency exchanges
3:25investments insurance all the controls
3:29to manage your money better like
3:31analytics budgeting saving accounts the
3:35security of your accounts expense
3:37management's all these things at least
3:39can go on the automation the API the
3:42open banking ones in Europe etc so
3:46there's a lot of things that are
3:47happening in banking in general and this
3:51is how we see today and you can even do
3:55some gifs with your with your money
3:59these days so but how we see it what
4:04what the the main purpose of the bank is
4:06to help you get most value are the money
4:10this is how we see our purpose in
4:13general despite we often used for mainly
4:18currency exchange travel this is our how
4:23we started with starting with that
4:25problem
4:26and I'm kind of naturally started
4:27growing I mean revolute as a business
4:30into other areas thinking about it's
4:33like how to help you to make more get
4:38more value out of your money
4:42hence yeah banks today look usually like
4:46vases and up and the cut and this is
4:49what what you see most of time this is
4:51the the new Association of the for the
4:53bank is not that big building as in the
4:56past just a little bit about revolu how
5:01we value our main cool thing is the
5:07multi currency our mission is to offer
5:10global multi-currency accounts with all
5:14the functionality around the accounts
5:17that's linked to your financial life we
5:22did at the moment by allowing you to
5:24hold up to 29 different currencies move
5:27them frequency without friction and fees
5:31and also we how to get exposure to five
5:37cryptocurrencies and we got to this
5:41point from when we launched in 2015
5:43those we support only three currencies
5:46through many currencies pounds euros and
5:48dollars and they have escalated and not
5:51working outside of Europe launching some
5:54in other markets in Asia in an America
5:57so we are heading towards this mission
6:01to making money frequency no truly
6:03frictionless truly all of us just to
6:09give you some perspective as well on our
6:11current skills so at the moment we are
6:13around five million users I put five
6:16million users because it's gonna hit
6:17probably the mark five million people in
6:20the next two days this is the volume of
6:25transactions with processing on the
6:26daily basis so it's it's quite sizable
6:31if you look at the industry in general
6:32just to give you again a perspective
6:34perspective the first payments in the UK
6:36the the biggest banks process
6:40that's many transactions a month so here
6:43we're talking about a lot of other
6:44transactions of course a lot of them are
6:46cut transactions but just in general get
6:48to give you perspective and how it looks
6:54now it's around ninety of bagging
6:58applications supporting all of this for
7:02mobile applications because we do i OS
7:05Android and we have products for retail
7:08and business and a bunch of web
7:12applications manually mainly for
7:13internal use and this all of this is
7:19built and supported by 220 people and
7:22this number is changing every week so
7:30how all this simple what you used to see
7:34a very simple interface right this is
7:37like how you imagine the bank at the
7:39back of it this is just a very
7:42high-level approximation of what's
7:45happening at the back end depending on
7:49size of institution like a financial
7:53solution would be doing just some of it
7:55delegating outsourcing to to partners
7:58a lot of it and within our case we do
8:02too much all of it
8:04we started initially with partnerships
8:07to our source things like compliance
8:10things like payment payment processing
8:16etc and people started bringing all of
8:19it in-house kind of getting closer to
8:22come to the last mile - now we have
8:26connectivity to many payment networks of
8:31cut cut payment networks like Visa and
8:34MasterCard its direct it's now in beta
8:37we're gonna be switching all of traffic
8:41to to our own connectivity soon then
8:45Faster Payments swift SEPA there's a lot
8:48of proprietary integrations with banks
8:50to make this whole system
8:53work efficiently and for you to just
8:56click a button and make it as soon as
9:00possible at no cost so but it was not
9:07that the big and complex from the
9:09beginning from the beginning was in the
9:11MVP right we called monolithic viable
9:14project product so it was in a way
9:19monolith which looked like this we are a
9:24Java house so we use the replicate we
9:30use relational databases for storage so
9:33many sounds too fancy like there's no
9:36scale etc but this is kind of the very
9:40foundational Salvage we do and we
9:42started just with this for applications
9:45one was the API that was powering
9:48powering the mobile apps used by users
9:51cut processing that would process the
9:55the cut transactions when you pay online
9:58or didn't point-of-sale and our
10:01back-office application and all of that
10:03was built kind of out of one code-based
10:06was mana rifle very simple the you're
10:09making changes through fracturing very
10:11simple when you have just one code base
10:13and just out of that it was just for
10:16applications built of deployments were
10:19like non don't I'm very fast 30 seconds
10:22to deploy any application so it was the
10:25life was very easy we would think we in
10:31from the beginning we went with
10:32PostgreSQL as a storage and we learnt
10:36learning early on how to scale progress
10:42as well have replication backups with it
10:45we purchased feels probably one of the
10:48best databases that exists today to
10:51which these things don't are not I would
10:55say not as difficult they may sound to a
10:57lot of application developers so yeah so
11:02we I mentioned for
11:04doing all of it with one codebase
11:08applications were stateless which
11:09allowed to do non time downtime
11:11deployments everything was easy to
11:14manage very easy to manage logs just
11:16from logging cessation doing your grab
11:20some logs the monitoring was very easy
11:24as well we would try to set up our own
11:27infrastructure with al haystack but
11:30eventually decided where when we started
11:32with only four people as the whole
11:34company so with only two technical
11:38people it was very very hard initially
11:40to manage all of that we decided to
11:43forego it and use nearly for monitoring
11:46which works pretty well usually to
11:50enable monitoring zero effort for Java
11:52applications with New Relic in yeah as I
11:59said we started this team with only two
12:02engineers including myself and then
12:03after almost a year of being in
12:05production we grew to four people and it
12:11was quite a change for us at the time
12:12and but we at the same time when we were
12:17building the product those known had
12:22experience building a bank before
12:24understanding different concepts what
12:28was important in general rights like
12:30where engineers usually are generalists
12:31we can know how to code and we can take
12:35any problem right give us enough time
12:37and we will solve it but there are
12:40certain areas so from financing industry
12:43what's important is correctness first of
12:45all like we always tend to think about
12:47complexity performance etc but the most
12:49important thing in in finance its
12:52correctness all right everything should
12:53be consistent that's why transactional
12:56behavior is like number-one priority and
12:58the typical way to do the transaction
13:01management if you know is to have this
13:04annotation based approach in Java right
13:06you put your annotations then do some
13:09IEP with spring or some other framework
13:11and very very easy right to define your
13:13transaction boundaries but as in this
13:17example you see
13:20this method like transfer money as for
13:23instance starts growing it this transfer
13:25money it's not just as easy just
13:28changing to balances with a lot of
13:29things you need to do do some
13:32validations fraud checks etc and that
13:35can become quite complex and then you
13:37say okay you know this is too much like
13:39why do I need to start transaction so
13:42early I need I want to optimize we
13:45release that connection sooner let's
13:49make the transfer logic block only
13:52transactional right and then we just do
13:54this we just unite your ID you did you
13:58say extract method for this block of
14:02code and this is what you get but what's
14:06wrong with it does anyone know what's
14:08wrong with this okay I'm glad there are
14:11people who we see the problem before the
14:16transactional notation was on the public
14:18method now it's in private and it
14:20doesn't apply anymore and if you don't
14:23have tests for every piece of logic we
14:26need to explicitly say this must be
14:28transactional you're gonna hit a big
14:30problem here as you like if we're
14:32talking about money then you start
14:33having consistencies and it seems
14:37initially oh why do we do this or
14:39because because it's it's very easy it's
14:43declarative way of doing it right that's
14:46what we're gonna generate and try to do
14:48not to be very empirical to use the
14:50declarative ways to do such things
14:52there's a more stable solution right is
14:57it less declarative but I don't think so
15:00but it always works there's no magic
15:02here at all right just explicitly say I
15:04want to call this block of code in
15:07transaction and since Java 8 it's very
15:09easy with lambdas the code doesn't look
15:12ugly anymore
15:13etc so this you know just using quote
15:20from Warren Buffett's rule number one
15:22never lose money but to never forget
15:25rule number one so in fairness it's it's
15:28super important we should never lose
15:31mine
15:32our money or customers money it's like
15:34it's like the most important thing
15:38especially consider these days all the
15:41money is in reality is virtual eyes it's
15:44like it's just numbers in some databases
15:47regardless on what the the financial
15:52institution is like unless you have that
15:54in cash this even that cash is still
15:56recorded some way to solve this problem
16:00on the other level as well you know we
16:04said we need to define a transaction
16:06button boundaries very clearly and in
16:09such a way so it doesn't so it's not
16:12fragile this something called free
16:16cancellation which is another important
16:17process because still there are too many
16:20different systems involved in processing
16:22financial transactions and we end up
16:25doing a lot of reconciliations
16:27especially when we cross borders when
16:29money goes out so of revolute to other
16:33banks and the other way around
16:35there is no transaction where right
16:38we cannot say okay let's do distribute
16:40transaction with a financial institution
16:42so we have another process called for
16:44cancellations which basically is about
16:46taking two data sets and making sure
16:48they they agree and of course due to
16:54this you know the story of SQL for us is
16:57like very obvious there's no
16:59considerations here for no SQL
17:02processing transactions no skill it is
17:05not really something you should consider
17:11just example of code that I wanted to
17:16show as well rational databases can do a
17:23lot of things that many java application
17:27developers are not aware of I've noticed
17:29one of these things is streaming on
17:34database level it's a called cursor in
17:36gdb sea level is called scrollable
17:39statements which means you don't have to
17:41create any complex batch processing if
17:44you want
17:45process efficiently a lot of data he
17:49just opened scorable statement behind
17:52the scenes I'm not showing the details
17:55of this because it's kind of in our case
17:56is hidden behind framework but here in
17:59this piece of code in this line here
18:03final processed standing always on a
18:06specific date it can be a lot of them
18:09and these things this is not a process
18:12that requires your immediate processing
18:16of the whole population on standing
18:18orders created by the users we need to
18:20post them on a specific date and that's
18:21important we can do it even sequential
18:24if want is but generally to process them
18:27efficiently so we don't have you know
18:29put all of that in memory you know or
18:32have very complex process bashing that
18:34just streaming from database and process
18:37as we go and whether we use will rely a
18:45lot on a library of you can call it
18:48framework as well
18:49it's called Juke if you've never tried
18:53it and you do SQL like I recommend Lucas
18:56is awesome is once you start using joke
18:58you never have mistakes in your SQL
19:02statements it's probably like I'm a big
19:05fan of of Juke and Lucas adder who is
19:09creator of it's in Montana right and
19:26another interesting problem in finish
19:28the domain is currencies of seems to be
19:34kind of simple it's you have one number
19:36to convert to another currency just
19:37multiply it by rate and you have amount
19:40in a different currency the rarity is
19:43how translation works there are always
19:47two rates bid and offer also called buy
19:50and sell and what I know if any of you
19:53still see or use the you know the brick
19:56and motar exchange bureaus
19:58apples they show you two rates right two
20:01columns and well see a lot of people
20:03even people would experience in infants
20:06an urgency they sometimes when they
20:07start thinking about this let's take a
20:09pause which one do I need to use like it
20:13can get tricky some but generally people
20:17make mistake all the time about this so
20:22we may be of course super simple for you
20:26is yes you put a number one one side it
20:29multiplied by the rate shown it will be
20:33the other number all right in the we
20:35kind of tried try to make it bit more
20:38clear to to our users and you know you
20:41don't have to have your own calculations
20:43in your head for us finally in the
20:48beginning and when we were building this
20:49like just currency exchange it's
20:54probably just the exchange rate class in
20:58our code base it has probably around 300
21:02lines just just that class and the whole
21:05area of currency exchange ended up
21:08because of kind of developers in the
21:11beginnings learning this space making
21:14mistakes in the back end and in the
21:17front end code so we ended up having
21:21probably more than hundred tests just
21:24around currency exchange so just to
21:31cover all the possible cases runnings
21:33etc and around crisis as well another
21:38problem that we have in this space is as
21:41we have this is another difference
21:44between the US and a lot of other
21:46fintechs Bank challenges etc where we
21:49kind of we try to solve this problem of
21:52currency coupla complexity and you know
21:56operating in different markets with
21:58different payment channels currencies
22:01they seem to be like a stable thing like
22:03they exist like in total are less than
22:06200 currencies in the world the royalty
22:08when we started in 2015
22:11we had few events where currencies were
22:14changing to would be a denomination
22:16event when the government decided you
22:18know this currency you know it's to blow
22:20that now we need to issue new currency
22:23that will replace it and we period of
22:25some there will be transaction coming in
22:27in both currencies those kinds this
22:30would get new isocodes etc and you can
22:33have to follow this adopt so you this
22:37transaction don't come and you like oh
22:39we cannot get exchange rates Buddhist
22:40currency because we don't know about it
22:42yet all right or some Connie cases like
22:45have you heard of that that I love men
22:49of Jersey against the Channel Islands
22:53they have their own pound currency it's
22:58generally we think it's it's it's a
23:00British pound but actually it's a
23:01different currency it's pegged to pounds
23:04it's always one-to-one it doesn't have
23:05its own exchange rate because of this
23:08market data providers Bloomberg etc they
23:10don't provide rates for those cars it's
23:12just they just tell them can you give me
23:14like I'm P is a love man pounds in give
23:17me right for it so even know I don't
23:19know what I'm P is but there is ISO code
23:22for this currency and transactions can
23:23come in in this currency so we end up
23:26having a lot of logic kind of this on
23:29this edges trying to you know put adopts
23:32all these two patrols as saying hey
23:35these pounds actual very same as British
23:37Pound right another thing that we've
23:46made mistake here things many startups
23:49they try to go fast and stuff we ended
23:52up has said it was monolith and we ended
23:55up in the beginning with this spaghetti
23:59architecture was called service
24:02spaghetti a lecture where you know this
24:06example here shows that it was a point
24:09when I said you know I don't want this
24:12automatic injection I need to run some
24:14something simple I I'm gonna just
24:17initialize it manually rights just
24:20explicitly and I started unrolling and
24:22rolling and rolling depending
24:24and I didn't get to the end of it I I it
24:26didn't work so because I hit a cyclic
24:30dependency because we use jook jook
24:34results depends this for you you don't
24:36even know you know it doesn't tell you
24:38that you have some better design code so
24:42this kind of thing made matters realize
24:47we shouldn't do this anymore
24:49we serve a spaghetti key texture like
24:53this is an example how you you like more
24:56explicitly you can get in this situation
24:58right you have dependence or something
25:00in change it's injected but then another
25:03thing also has dependency on it and it's
25:05like and you can get very complicated
25:06you don't even know so basically my
25:10advice will be especially a new world of
25:12micro services when we say we're gonna
25:15fill this very focused on some piece of
25:18functionality small applications
25:20services just injection of the injection
25:23doesn't give you any benefits really
25:25like there's not much blood that you
25:28actually create by doing it manually so
25:30by not doing this you making sure that
25:32your code is actually was simply and you
25:34are able to then to CC when it becomes
25:36too complex so we then switch because of
25:44this spaghetti code we switched into
25:49even driven mana let's face okay all
25:53these services like to process the
25:55transaction it's not just explicitly
25:57moving money from one account to another
25:59it's sort of other things maybe we need
26:01to do some checks injects one service we
26:04need to do fraud detection you inject
26:06another service then at the end we want
26:07to check out maybe we need to trigger
26:09some promotion here or when this send
26:11the push message it's a lot of stuff
26:12going on when one transaction happens
26:15and kind of this becomes as on that
26:18picture very very complicated so for
26:20then we resolved it by billing only in
26:25house even during framework here's just
26:29a fragment of you know we started
26:32defining logic in just smaller pieces
26:36it's it's a secret like framework where
26:39this that commands and there are
26:41repositories to to fetch dates are kind
26:43of a certain separation and those more
26:47compact pieces of logic commands we call
26:50them actions they the whole frame of
26:53bills aren't again as I said DDG
26:56concepts everything ends up being every
27:00changes up issuing an event and then you
27:03can do on top of those events create
27:06things like this and whatever you want
27:09any side effect you want to do you don't
27:11have to inject anything anymore you just
27:13create it it works independently anyway
27:15and just because of principle is
27:17everything issues an event just not to
27:21confuse it's not
27:23sakura-san the sense of the see us model
27:26where there is even sourcing where
27:27everything starts with an event and of
27:30that we create our we'd not don't we
27:33didn't do that this because we didn't
27:34have it first of all from the very
27:36beginning we and we didn't want to
27:39switch because even thousand has a lot
27:42of own problems with latencies etc
27:45there's all of complexity we decided to
27:47have dual model in a way we consistently
27:50always write in even log and at the same
27:53time your new state so which allows us
27:58quickly the next decision on the state
27:59could be done without calculating it and
28:02events are used for have to have audit
28:06log of everything and have even driven
28:09architecture so initially was I said
28:12it's even driven monolith because all of
28:14this was still within the code base but
28:16the code was already segregated into
28:18these kind of smaller modules which
28:22allows us then to move into next phase
28:25of check or even druvan services I don't
28:30I don't like using micro service system
28:32because it's kind of it's the name
28:36itself is maybe a bit exaggerated
28:39because when we say micro before we had
28:42service oriented architecture and those
28:43were services now we have micro services
28:46was mean there
28:48million times smaller or in what senses
28:51when is this problem because of this not
28:55everyone agrees on what how big that
28:57micro service should be etc so if we
29:00just say our services should perform you
29:05know if they separate by domain
29:09boundaries and you know they linked
29:13through three events and so in that
29:17previous thing by the way on the
29:21previous slide was on the front that we
29:23used angular at the time react was not
29:26seen fashion yet and then in 2016
29:29everything was was written into react
29:33which is a good thing angular was first
29:37of all it was not done by people
29:39specializing that was done by Java
29:41developers it was a nice so now
29:45everything is on racks is a lot cleaner
29:47on the front end as well so and the
29:50whole thing this ops this done being
29:54more applications in fact I didn't show
29:57here the whole thing at that stage
29:59already was there was 2016 there was a
30:04lot more services but they too was small
30:07only see like bit of monolithic code one
30:10one code base and we introduced this
30:13even store where we started assisting
30:15all these events and which allowed us
30:16not to have this even driven model just
30:19within this monolith application but
30:21start moving this modules into just
30:24separate services separate code means
30:27because we started sending these events
30:29to even stove assisting them there and
30:31stream from there and then it would
30:33allow us to do things like this which is
30:36out of trying to explain what was going
30:39here so what one of the problems will
30:43have because we allow you to exchange
30:46currency instantly but reality is we
30:53become like a broker for you really is
30:55nothing happens instantly in real world
30:58like somewhere in the bank this there
31:00are accounts and then
31:02the money should actually move from on
31:03layer below it search and lie below the
31:05set model processes around this etc so
31:08what it means when we as a user you have
31:13a contract with us I mean we tell you we
31:15give you the service we exchange money
31:17quickly and start using it we cannot
31:21pull back like okay we it means we're
31:24accumulating risk the risk that we when
31:27you safe exchange hundred pounds to two
31:30euros we need to get this euros from
31:32some who you will gave it to you now
31:34but we don't have it at that moment
31:36potentially right because it's not like
31:39people have all in equal amount of
31:42pounds and euros if they are constantly
31:44exchanging between themselves it's not
31:46balanced in this way but there is this
31:49offsetting effects anyway which helps
31:51but so we this kind of even driven model
31:55allowed us to so initially how we did it
31:57all our risk calculations happening was
32:00just one big query on transactions table
32:04on our ledger and it would produce like
32:07this and the numbers these our how
32:08sending risks and our finance department
32:12would go and act innate initially it was
32:14not big volumes that's why it was okay
32:17and but I'm falling started growing that
32:19calculation that crews are being slower
32:21slow and hit a point we're doing some
32:24optimizations on a career level and
32:26doing caching as I said was all
32:28pointless
32:29it was timing out in we wanted to have
32:32instant reaction to this so even driven
32:36allowed us to have this thing the risk
32:40calculation started happening in a
32:42separate application which on every
32:44event every literally transaction every
32:47financial movement it would update it
32:49risk online pretty much would be with
32:53delay probably of up to 100 milliseconds
32:58and that allowed us them to build this
33:01whole hedging of our risk in automated
33:05manner as well so which you know of
33:10volumes I was bigger and our finance
33:13you don't have to go on and go and trade
33:17it on the market that's accumulated risk
33:23another big problem in finance when you
33:26starts moving money around especially
33:28with us when we move with around the
33:31world
33:31and we try to make it as simple as
33:33possible is the financial crime which is
33:39one of those things is money laundering
33:42so as any financial institution using
33:48responsible food taking care of this
33:50problem and the biggest issue that is
33:53it's not deterministic it's you cannot
33:57just say one day we've fully solved it's
34:00like we clearly can say this is money
34:03laundering
34:04this is geniune transaction like there's
34:07no ever a hundred percent guarantee on
34:10anything this problem ends up being
34:12probabilistic and this is where a lot of
34:15our data scientists they build much
34:18machine learning models Concilium etc
34:23because we gonna have very complex
34:26business model all of different users
34:28different currencies country says it's
34:30very complex and guys it's amazing jobs
34:35they're our levels so on this it's hard
34:40to measure what say on fraud size which
34:43is a similar problem where the different
34:47types of for when people process still
34:49people's card details and then try to
34:51make transactions on behalf etcetera so
34:54for us to say with introduction of
34:58machine learning models filiz we
35:01decreased the level of fraud from
35:08recalls so it went down probably about
35:1140 times like the level like we probably
35:16about eight times so below the standards
35:20industry standards so it's it's quite
35:26quite impressive I know if you--if
35:27and you ever had a notification from
35:29system called shellac so it's like our
35:32in system based on this ml Allegan that
35:37detects the card card fraud and as I
35:41said it's very public problem and it's
35:43two-sided problem so on one side we want
35:46to catch these bad guys on the other
35:47side we don't want to affect the gene in
35:50users right and it's it's very hard
35:52problem is kind of the teams working on
35:54this they constantly try to optimize to
35:58decrease this so-called false positive
36:01right so that's and it's not possible to
36:05decrease it to zero unfortunately but in
36:09addition to this there are just more
36:11items in process when someone gets into
36:13that situation a genie the user gets
36:15caught into this trap process how to get
36:18them out who it's you know quickly
36:23through certain things like you may be
36:26asked or you may be asked on the up to
36:28upload documents proving maybe your
36:31transactional session so that it avoids
36:34this kind of way you deal with
36:36traditional bank I happened to be in the
36:40situation quite a few times when you do
36:42an international transaction it just
36:45binds just by default blocks it and you
36:47sits in anyways and then you have to
36:49call them and then it takes days
36:51sometimes week which is resolved you
36:53without kind of even being able to do
36:57anything at the end you just go to the
36:58branch so it's very very complex you can
37:01get very complex so how we dealt this
37:05with this promise in the beginning for
37:07that monolithic actually there was just
37:10within transaction processing there were
37:12all these checks and they were not the
37:15work because of this they were limited
37:17and we couldn't do too much while we
37:20process transaction at that point to do
37:23much of smart calculations to figure out
37:25like okay based on the transaction
37:27history do we think this is authorized
37:30by you or not so this was not obviously
37:35scaling and we went into this machine
37:41learning this algorithms and that allows
37:45us to do something this this kind of
37:46architecture where our transaction
37:50authorization engine it would ask this
37:53code brains this most of that stuff is
37:57written in Python because that's what
37:59data science is usually the most yeah
38:03that's the best fit for that kind of
38:05problem
38:07gets an answer but an answer here will
38:09be really prepared already this for
38:12detection system it will already know
38:16before the next transaction happens and
38:19in addition with this new information
38:21will be able to kind of to give an
38:23answer very quickly learn based on that
38:24there will be an event generated based
38:28on you know you were successful
38:29transaction or was a decline due to
38:31suspicion and fraud would go through our
38:34even stall which would stream back into
38:36the system it will updates everything
38:38and then this how it knows them for the
38:40next transaction already to give you a
38:42quick answer because in some cases is
38:46very critical
38:46where on card transactions we don't have
38:50that much time to do a lot of smart
38:52things and that's why we have come to
38:53have this model pre calculating all of
38:55things because we would be limited with
38:57like 500 milliseconds to do all of it
39:08slowly I'm gonna touch quickly on the
39:14process itself on the culture how we do
39:16these rights it's we known as a company
39:21of growing fast building order features
39:24fast etc how does it how does it work
39:27how do we do it without breaking stuff
39:30so the basic principles people who do
39:36this stuff they they're responsible
39:39there is no pipeline how it would be
39:41typical in big organizations and begin
39:44the price the business only is preparing
39:48requirements and developers coding that
39:51and
39:51the testers testing and then the release
39:55manager etc some organizations can be
39:57very complex and they may call it agile
40:00as well
40:00what father Jaya was something something
40:03else growing like from just few people
40:06developers become these kind of multi
40:10faceted all rights you use code just as
40:13a tool you your programming skills to
40:15solve problems so engineers because
40:19we're on the cloud nose while dealing
40:21with the fractious actually is more
40:23about coding you then you know actually
40:27as in the past dealing with physical
40:28Hardware etcetera the all the complexity
40:31of the hardware is hidden from you just
40:33work with these abstract notions you
40:35know VPC sub networks etc which are in
40:39anyone who studied computer science
40:43should understand at least on the high
40:45level this wife in in our case we have
40:48specialized DevOps guys who do very
40:52specialized projects and they would do a
40:54lot of heavyweight stuff maybe doing
40:59some working on scalability of this
41:01infrastructure on on skill build of
41:05databases on on network layer physical
41:10connectivity will have some physical
41:12connectivity to to payment networks in
41:15physical data centers around security
41:19etc and huge thing which which is like
41:26in 2017 when we started growing fast and
41:30a team of ten people quickly turned into
41:33a team of 50 developers people coming
41:37from different places from different
41:39understanding of HL processes was
41:44interesting as well to hear about one of
41:46the common answers to would is what it a
41:50gel is for you what do you mean by agile
41:51process from a genius to here oh it's
41:54scrum which is you know we don't see
41:58scrum is a gel and if it is because it's
42:02kind of rigid process in a way and
42:05from my experience evolved in many
42:07organizations in the past teams struggle
42:10to adopt scrum which shouldn't be
42:13initialized so how agile is about people
42:15and processes right process should be
42:17helping not saying in a way so we're
42:20kind of not focused on scram or scrum
42:22just do things that work and one thing
42:27that definitely works and that face when
42:30we started growing fast and developers
42:34DG was set as like a engineering
42:39strategy like TDD is for everyone who
42:43does coding everyone has to do TDD like
42:47not everyone does it on the same level
42:48of I'll be honest new people as well and
42:52realized in some organizations TDD
42:56understanding of TDD is just having
42:58tests which is surprising to me at the
43:01time what was made it very clear DD
43:06means you have to test first if you talk
43:08about quality you have to put quality
43:09first and so you have to write your
43:11tests first and then focus on the
43:13problem describe it and then you go into
43:14implementing it like very important
43:17concept this is why because of this vs.
43:21we didn't have historically test this we
43:22had mobile justice would test UI in the
43:26beginning and that didn't work well
43:29because the absolute time as well the
43:34weight of box was a lot higher because
43:37engineers didn't have culture of
43:39test-first development always kind of
43:43lay off loaded this responsibility of
43:46quality to to testers right
43:48which wasn't scaling well so we decided
43:52to change this and this is like we
43:56stopped hiring here.i testers it's
43:58completely and those who left those
44:01people they focused on creating
44:04frameworks and infrastructure so no one
44:08would often is called a software
44:10engineer in test so they would do
44:12automation help developers to focus on
44:15solving the determine problems
44:19them tools to write tests efficiently
44:22and run them and since since then
44:26so the like we started tracking the red
44:30of box crashes in the ABS in both areas
44:33Android it's that it's give motivation
44:36to develop it's actually to protect
44:38themselves from these problems because
44:41they come back to them at the end of the
44:43day so they started focusing on testing
44:46a lot and builds all the friends for
44:48this and as kinda mentioned a few times
44:55through the talk we one of the how
45:01shallow things is we're very problem
45:03driven not solution driven what it means
45:06is solution becomes secondary for us
45:10when we like the arguments what we do
45:13this way this way to use this technology
45:16or that when you start putting this
45:20around the problem definition when you
45:22define problem well but everyone who
45:25works on it understands it the same way
45:26then this becomes a criteria does it
45:29something work there or not and are the
45:32differences that outside of this process
45:34kind of become we're less important than
45:37and like the this Hollywood's generally
45:39don't don't happen in company between
45:41engineers and of course another very
45:48important aspect is peer reviews it's a
45:52way to challenge each other and learn
45:53from each other so it's General General
45:56Lee looks like everyone is a way of code
46:02review process with full requests etc
46:06we kind of would do this as part of this
46:10peer review but focus again on the
46:12problem not on the code what it means is
46:15in this process engineers are actually
46:18they have a slide where we coded is a
46:20specification just again to focus you
46:24have this focus on ddd describing first
46:27the problem with test where it starts
46:30from understanding the problem
46:32thinking about it designing some
46:34solution discussing it with your
46:35reviewers so that's where they Sam you
46:38prepare some code a pull request they
46:41wouldn't know what it is about and they
46:44spend a lot less time you know they ask
46:46you question they expect to see
46:47something in the code by the time and
46:49they focus on this thing if you solve
46:51the problem first not on you know second
46:54syntactical things that are resolved by
46:57introducing conventions and it's like
47:00not the main focus and I'm gonna speak
47:09because not much time this and yeah and
47:14really obviously all of those things
47:17they're part of in provinces but without
47:19people involved in this not having drive
47:22to create the great experience for fully
47:24customers to be involved and you know
47:28understanding the impact engineers crazy
47:31this all of that wouldn't matter really
47:33so this is probably the main thing in
47:37the whole yeah and as a pin is a
47:41conclusion to conclude so you a problem
47:45driven you learn and apply new learnings
47:48fast you work as a team you can build
47:50anything not not just buying as
47:53engineers as a nation we we have these
47:55benefits of knowing the tools
47:58understanding abstraction we can apply
48:01these skills pretty much to any domain
48:03which is possessing in awesome position
48:06really compared to to some other people
48:10so what yeah don't don't just try to do
48:13all your ideas at once just for be
48:15focused that's what I was saying any
48:23questions
48:33I'm sorry so we're talking about a
48:48giving example of cap transaction
48:50authorization what means this signal
48:53when you say you touch the controllers
48:58payment attach terminal movie cards from
49:01their signal go through multiple parties
49:04we will go through some payment gateways
49:06acquiring banks it will go through
49:08central network or senator service of
49:11MasterCard or Visa then we'll go to
49:13process it to us and this is what I mean
49:16I bet point when we receive this message
49:18we have this 500 milliseconds to say
49:22okay this we can authorize it
49:24yes so the inside the direct end-to-end
49:29yes it takes longer the before cut
49:31transactions there is certain limit most
49:33economies they have some differences
49:35depending on the market as well
49:37depending how all the system is in
49:39different countries but as far as
49:42remember it's about five five seconds or
49:465 seconds and like this and end to end
49:48yeah sure yeah so the question is how we
49:56handled technical debt over time so
50:00technical debt is something we don't
50:03define it as a separate thing so focus
50:08on things like refactoring and since we
50:11moved to this architecture of you know
50:16even architecture with secure s with
50:19small components this becomes more
50:22problem driven like you make the
50:25technical debt becomes at certain points
50:27is just a new requirement we say okay
50:31now we which in certain level of
50:34customers we need to think how to scale
50:36for the next year like well can we scale
50:39with the current solution can we process
50:41the transactions what do we do just
50:43scale out by you
50:46you know allocating more resources
50:48processing resources always should we
50:50change something in the code and we make
50:51this decision we don't look at it as a
50:53technical debt technical debt was at the
50:56point when we transition to that those
50:58literally arising all the code and that
51:01happened and we have technical that
51:04probably define it so yeah it's just not
51:07give me absolute answer there is
51:09technical debt in mobile apps so in on
51:14Android app that has been closed
51:16everything was written from Java 6 to
51:19Kotlin which made all Android developers
51:23lives a lot easier
51:24and then on is it still these mix of
51:28Objective C and Swift so there's a bit
51:30of technical debt and rewriting of
51:34internal frameworks thank you
51:42[Applause]
51:46[Music]

About this transcript

This page contains the full transcript of Anyone can build a bank: creating a new banking backend by Vlad Yatsenko by Devoxx, generated from the public captions YouTube serves with the video. The transcript has 7,047 words across 1,019 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.