Engineers do not get to make startup mistakes when they build ledgers | Hacker News Reader

Engineers do not get to make startup mistakes when they build ledgers

news.alvaroduran.com

590 pointsfagnerbrack551 comments

Tell that to Synapse customers. Many millions of dollars are missing.

Banks have to follow strict rules to account for where all the money goes. But the way fintechs work, they usually just have one or a couple underlying "FBO" accounts where all the pooled money is held, but then the fintech builds a ledger on top of this (and, as the article points out, to varying levels of engineering competence) to track each individual customer's balance within this big pool of money. In Synapse's case, their ledger said the total amount of all of their individual customer balances ended up being much more than the actual funds held in the underlying FBO accounts. Lots of folks are assuming fraud but I'm willing to put money that it was just a shitty, buggy ledger.

FWIW, after seeing "how the sausage is made", I would never put money into a fintech depository account. Use a real bank. Fintechs also often put out the fake promise that deposits are FDIC insured, but this only protects you if the underlying bank goes belly up, not if the fintech loses track of your money.

See https://www.forbes.com/sites/zennonkapron/2024/11/08/what-th...

90 million dollars missing and 250 million dollars frozen. That 250m probably needed by some people to pay rent.

Backed by Andreesen Horowitz who are conducting a scorched earth jihad against all government regulation.

https://finance.yahoo.com/personal-finance/synapse-bankruptc...

The sad thing is that most people don't learn lessons from history. It took me far too long to start learning lessons from history.

After an asset bubble and collapse people will understand why we have a lot of the regulations from the 1930s.

That's not really a lot compared to what Wallstreet is steeling daily.

You do understand what you described is basically Bolshevik/French revolution only different times.

Some men with some power using young starry eyed young people to grab much more power from incumbents.

At a big co I worked at, the lack of consistency between trading systems caused money to (dis)appear (into)out of thin air.

Prior to one of these hiccups, I hypothesized, given how shitty the codebase was, that they must be tracking this stuff poorly.

This led to an argument with my boss, who assumed things magically worked.

Days later, we received an email announcing an audit one one of these accounting discrepancies.

JPMC proposed using crypto, internally, to consistently manage cash flow.

Not sure if it went anywhere.

At all of the exchanges and trading firms I’ve worked with (granted none in crypto) one of the “must haves” has been a reconciliation system out of band of the trading platforms. In practice one of these almost always belongs to the risk group (this is usually dependent on drop copy), but the other is entirely based on pcaps at the point of contact with every counterparty and positions/trades reconstructed from there.

If any discrepancies are found that persist over some time horizon it can be cause to stop all activity.

Fun fact, centralized crypto exchanges don't use crypto internally, it's simply too slow.

As a contractor, I helped do some auditing on one crypto exchange. At least they used a proper double-entry ledger for tracking internal transactions (built on top of an SQL database), so it stayed consistent with itself (though accounts would sometimes go negative, which was a problem).

The main problem is that the internal ledger simply wasn't reconciled with with the dozens of external blockchains, and problems crept in all the time.

> JPMC proposed using crypto, internally, to consistently manage cash flow.

Yikes, how hard is it to just capture an immutable event log. Way cheaper than running crypto, even if only internally.

It's all merkle trees under the hood. I feel like the crypto coin stuff has overshadowed the useful bits.

If its for internal why not just use a normal append only log. x amount transferred from account y to account z. A three column csv oughta do it.

> I hypothesized, given how shitty the codebase was, that they must be tracking this stuff poorly.

That is like half of the plot of Office Space

This sounds like a situation that I know about at the placed identified by name in your comment. It took months to track down the issue.

Synapse says that it was actually the Bank (Evolve) that made the accounting mistakes, including missing transactions, debits that weren't reported, sending in flight transaction to Mercury while debiting Synapse incorrectly etc.

https://lex.substack.com/p/podcast-what-really-happened-at-s...

Thanks for posting this, I will definitely listen to it.

While I haven't listened yet, one thing I don't really buy when it comes to blaming Evolve is that it should fundamentally be Synapse's responsibility to do reconciliation. This is what completely baffled me when I first worked with another BaaS company - they weren't doing any reconciliation of their ledgered accounts with the underlying FBO balances at the partner bank! This was insane to me, and it sound likes Synapse didn't do it either.

So even if Evolve did make accounting mistakes and have missing transactions, Synapse should have caught this much earlier by having regular reconciliations and audits.

Rambling interview. As best as I can tell Synapse said there were technical issues with Evolve the bank.

Meanwhile this article said Synapse withdrew from Evolve the end user funds. Mr. Holmes of Evolve said the bank “transferred all end user funds” to other banks at the request of Synapse, but declined to identify them

https://www.nytimes.com/2024/07/09/business/synapse-bankrupt...

Wise also recently switched their US bank provider from Evolve to Community Federal Savings Bank. Maybe they had similar issues?

From a cursory look at how it describes itself (BaaS, etc), Evolve is hardly a "bank" in the traditional sense of the word.

Sankaet is full of shit.

I had money disappear from my HSBC account. As in, the double entries didn't match by a small amount (it was a payment between two of my accounts that didn't match up, which I couldn't trivially reconcile in the books). I pursued this for a while but they never even properly acknowledged it let alone fix it.

I had my unfounded suspicion it was some internal subtle theft going on, but incompetence is probably a better explanation.

If you live in a developed country it should be sufficient to ask them to account for it with a note that a formal complaint will be sent to relevant authorities if nor dealt with in timely manner.

That stuff like this is in order is the foundation of kapital societies and is taken quite seriously.

I had a similar situation with Santander many years ago: it was a small amount and happened when I was closing my account, bank manager couldn't explain it and escalating the problem was a pain - especially because I was about to move to another country and had more urgent things to do.

I wonder how common issues like these are...

You're likely correct it was theft. I was told by a CTO there that topping up accounts with bank money where it has been hacked away was daily routine and cheaper than fixing the systems. Incompetence would not manifest on routine operations like this.

> I had my unfounded suspicion it was some internal subtle theft going on

Had you watched Office Space recently?

I’m skittish about real banking institutions as well. Vanguard for example outsourced a bunch of their dev work to India a few years ago. Had a friend that worked as a sysadmin for BoA. They were required to keep certain logging for 7 years but he would just delete it anyway when disks were starting to get full.

But the fundamental difference is that the regulatory structures are in place to recover your money if a bank loses it. That's not the case with fintech middlemen. Take the Synapse case:

* End customers are really a customer of Yotta, a (silly IMO) fintech where interest was essentially pooled into a sweepstakes prize.

* Yotta was a customer of Synapse - they used Synapse BaaS APIs to open customer accounts (again, these accounts were really just entries in Synapse's ledger, and they underlying funds were supposed to be stored in an FBO account on Evolve).

* Synapse partnered with Evolve, who is the FDIC insured bank.

Synapse went bankrupt, and Yotta customers are finding out they can't access their money. But the bankruptcy court is at a loss as to really what to do. FDIC isn't getting involved, because as far as they can tell, Evolve hasn't failed. Synapse is basically out of the picture at this point as they are bankrupt and there isn't even enough money left to do a real audit, and Yotta is suing Evolve alleging they lost customer funds. But, in the meantime, Yotta customers are SOL.

If you had a direct relationship with an FDIC-insured bank, and they lost your money, there would be a much clearer path for the FDIC to get involved and make you whole (up to $250k).

> They were required to keep certain logging for 7 years but he would just delete it anyway when disks were starting to get full.

I’m highly skeptical of this claim. Every bank I’ve worked with adheres to their records requirements like it’s life or death (because it kind of is for the bank).

Tell your friend he’s exposing himself to hard prison time if he’s not just making up a story. If his boss tells him that they don’t have budget to retain the logs he should be whistle blowing, not violating federal banking laws to save what is a rounding error in their IT budget.

Coincidentally written something about this yesterday [1], but the gist of my take summed up is that the nature of accounting oriented data models doesn’t help when dealing with multiple FBO accounts.

The main problem is that accounting defaults to full fungibility of monetary amounts represented in a ledger, which has the effect of losing track of the precise mapping between assets and liabilities, so you end up in a position where you simply cannot tell precisely to a bank _who_ are the actual customers they owe money to.

[1] https://www.formance.com/blog/engineering/warehousing-promis...

I like Cory Doctorow's saying: "When you hear the term 'fintech,' think 'unlicensed bank.'

"Fintechs also often put out the fake promise that deposits are FDIC insured, but this only protects you if the underlying bank goes belly up, not if the fintech loses track of your money".

Would you count Wealthfront as a fintech? I was finding their marketing compelling, but this thread makes me think twice.

There is a pretty fundamental difference, and it’s that Wealthfront (and M1, Robinhood, Fidelity, etc) are registered broker-dealers. Broker-dealers are regulated just as stringently as banks, but by the SEC and FINRA as opposed to the Fed. Broker-dealers have been running passthrough FDIC programs for decades, and in a lot of ways have more stringent regulations than banks. The most notable is that they are forced to segregate assets (can’t put client assets on the balance sheet, they have to custody them separately), which is the ultimate way banks fail and need the FDIC to bail them out. Source: used to work in broker-dealer auditing

Yes, it's the same basic principle going on at Wealthfront/etc.

It's possible (probable?) that they have better accounting controls. But I personally wouldn't keep anything above SIPC limits at Wealthfront (or any near competitor like Robinhood, M1, etc). And I'd be keeping records on my own.

And I'd make peace with the fact that SIPC resolution is a completely different ballgame from FDIC turnaround for assets held directly at an insured bank (which is like single business day don't-even-notice fast). I.e. not use it as the sole source of emergency funds, have months of expenses at a different institution, etc.

In addition to a ledger, fintechs need a reconciliation system to ensure the ledger is correct. Does the card processor audit files match your ledger? Does your ACH and check processing systems match the ledger? What about external money movements at the sponsor bank. Are they recording in the ledger?

> In Synapse's case, their ledger said the total amount of all of their individual customer balances ended up being much more than the actual funds held in the underlying FBO accounts. Lots of folks are assuming fraud but I'm willing to put money that it was just a shitty, buggy ledger.

Bugs are as likely to show more and less money than there really are. But bugs in production will almost always show more :)

I had been debating the merits of using Flourish, but I'm sticking with SNOXX on Schwab. Same rate and I think SNOXX has to be safer, right? Even with the Flourish FDIC guarantee, as others have pointed out, it's only for the underlying back not Flourish itself.

In Synapse's case, their ledger said the total amount of all of their individual customer balances ended up being much more than the actual funds held in the underlying FBO accounts. Lots of folks are assuming fraud but I'm willing to put money that it was just a shitty, buggy ledger.

If there was no malfeasance then no money would be gone. The totals would add up, they just wouldn’t no know who was owed what. Since the totals don’t add up, someone got the extra.

    > Fintechs also often put out the fake promise that deposits are FDIC insured

Does this still happen?

Many fintechs are not licensed to hold funds and work with bank partners who hold your actual funds. That allows them to say they're insured because they're not co-mingled with the corporate funds in the event of insolvency. This doesn't stop them from making accounting errors.

The FDIC said you can't do this anymore starting Jan 1, 2025. So I expect it to stop in about 30 days. The FDIC will probably find a few laggards and throw some fines at them, and the process will then probably completely stop.

the problem is the discrepancy between what the Fintech means when they say fdic insured, and what the customer hears when they're told fdic insured. the customer (erroneously) assumes it means that if the Fintech or anyone else has problems, the customer is covered up to the 250k fdic limit. what the Fintech means, is that there's someone they're partnered with that is a bank and is fdic covered. How the money is deposited into the bank is up for interperation. if there Fintech is being dishonest, they have one bank account at a bank, and all of the customers money goes into that one shared account, they're not technically lying - the money is fdic insured. unfortunately for the customers, that's not the same as each of them being fdic insured is the Fintech goes under. fdic doesn't seem to want to clarify this issue either, which is a problem.

The fact that practically all funding most of the world runs on these days is just a bunch of variables in some shitty program never stops being weird to think about. All it takes to create or destroy trillions is one (or maybe a few) CPU ops.

It really stretches the belief into fiat money to the absolute limit.

Dade: This is every financial transaction Ellingson conducts, yeah? From million dollar deals to the ten bucks some guy pays for gas.

Kate: The worm eats a few cents from each transaction.

Dade: And no one's caught it because the money isn't really gone. It's just data being shifted around.

> Lots of folks are assuming fraud but I'm willing to put money that it was just a shitty, buggy ledger.

I'm not sure there's much difference. Intent only matters so much.

I mean... Fraud is defined by intent.

You can argue negligence over mistake. But fraud definitely requires intent.

> In Synapse's case, their ledger said the total amount of all of their individual customer balances ended up being much more than the actual funds held in the underlying FBO accounts.

When the banks do this it's called "fractional reserve banking", and they sell it as a good thing. :)

I’m constantly amazed by how much the crypto community thinks they understand fractional reserve banking while getting it so completely wrong.

In fractional reserve banking, money that is loaned out is accounted for as liabilities. These liabilities subtract from the overall balance stored (reserved) at the bank. The bank is not printing money new money, no matter how many times this idea gets repeated by people who are, ironically, pumping crypto coins that were printed out of thin air.

I think it’s incredible that cryptocurrencies were literally manifested out of bits, but the same people try to criticize banks for doing this same thing (which they don’t).

There's more to it than that; balances are exceeded by the sum of "assets held by the bank" and "assets owed to the bank".

@serbuvlad: “When the banks do this it's called "fractional reserve banking", and they sell it as a good thing. :)”

How dare you criticize our holy banking system /s

One of the things it took a little time to wrap my head around when I started working at Google was trading off reliability or correctness for scaling.

I had previously built things like billing systems and small-scale OLTP web applications, where the concept of even asking the question of whether there could be any acceptable data loss or a nonzero error rate hadn't occurred to me. It was eye-opening to see, not so much that if you're doing millions of qps, some are going to fail, but more the difference in engineering attitude. Right this instant, there are probably thousands of people who opened their Gmail and it didn't load right or gave them a 500 error. Nobody is chasing down why that happened, because those users will just hit reload and go on with their day. Or from another perspective, if your storage has an impressive 99.99999% durability over a year, when you have two billion customers, 200 people had a really miserable day.

It was a jarring transition from investigating every error in the logs to getting used to everything being a little bit broken all the time and always considering the cost before trying to do something about it.

Durability levels that poor aren’t state of the art any more.

The rule of thumb I’ve seen at most places (running at similar scale) is to target one data loss, fleet wide, per century.

That usually increases costs by << 10%, but you have to have someone that understands combinatorics design your data placement algorithms.

The copyset paper is a good place to start if you need to understand that stuff.

That sounds a lot better than some number of nines standing by itself.

99.(9)x percent durability is almost meaningless without a description of what the unit of data is, what a loss looks like. There's too many orders of magnitude between a chunky file having an error, a transaction having an error, a block having an error, a bit having an error...

This all makes a lot of sense, but I have seen this a lot in the opposite direction. Specifically, people from social media companies with very squishy ideas around failures coming into financial applications. Generally does not go well.

While I was at Google I was loaned out to the Android team to work on contact synchronizing. There was a problem that I was running into, but it was an extremely rare situation. I pulled aggregated data from production, and it looked like it would affect 0.01% of people. When I presented the solution and mentioned this error rate I was asked how those 200,000 Android users could remediate the situation. They couldn't, their contact sync would just be broken, so I was told to go back to the drawing board.

But yeah, it was the number that was humbling.

There are definitely areas where 4 nines are good enough, but there are just as many areas where the aren't.

This is the type of thing where it helps to hire the right people from the start. You hire a bunch of leetcode experts, but never stop to ask if, besides conjuring data structures and algos in their minds they actually can build the thing you want to build. If your people know how to build the thing, you don't need to sacrifice growth, it gets built right from the start.

Sometimes you need engineers that have some other type of education, maybe accounting, maybe finance, maybe biology. I always thought that the most important part of my career was understanding every industry I built for, deeply, and knowing experts in those areas so you could ask the really important questions. That is problem solving and engineering. The rest is programming/coding.

I have never come across a post on HN that was so scarily describing my current day to day and with a comment I agree with so wholeheartedly.

I’ve spent the majority of my career in tech with a finance angle to it, mostly sales and use tax compliance.

What I never fully appreciated was how much those accountants, controllers, and lawyers were rubbing off on me.

I was recently advising a pretty old startup on their ledgering system and was beyond appalled at what it looks like when a bunch of engineers with no finance or accounting background build an accounting system.

We don’t have to find magical accountant engineers either, it will wholly suffice if we sit actual accountants in with our engineering team during the design process.

After my design of their complete overhaul I had a friend who is a CPA completely review my work, we found a few holes in certain scenarios, but by and large we were good.

Money is a difficult engineering problem because with money comes all the human goofery that surrounds it.

This. Ledgers are a domain knowledge. Anything else should be built on top, including high load optimizations.

In some circles, there is the irritating tendency to believe that technology can solve every problem. Experts are eschewed because innovation is valued above all else.

This reinforces for me the importance of domain knowledge in engineering leadership. i.e. if you work for a finance company you need to have a decent understanding of finance in order to make the correct technical decisions and tradeoffs, same goes for journalism, same goes for commerce etc. etc. Successful organisations I've worked for have always included domain specific, non-technical, questions to tech team interviews whilst some of the most technically acomplished teams I've worked with have floundered through a lack of domain insight.

I’m a software engineer and a CPA, so I’ve got the domain knowledge. However, I’m not sure how to find a gig where I can apply that. It seems everywhere I look vastly prefers someone with twice my experience as a software engineer and zero domain knowledge over my years of experience in accounting and fewer years of experience as a software engineer. Any insight on how to utilize it effectively?

Unfortunately, the likely solution to this is to keep banging out work at wherever you can gain experience as a software engineer.

Modern hiring practices are too rigid and filter-friendly for you to likely appear as a strong candidate based on the fact you have good accounting experience on top of your growing software skills.

What will really help you though, is having friends who work at a bank in the software departments. It's almost always who you know in this world. You need to network, network, network!

You know, it's kind of funny, but it seems like most businesses are not interested in doing things "right". As doing things "right" is time consuming, hard, and cuts off avenues to "innovate". "That's not how accounting works" is like telling management to clean their room. No one wants to hear it.

As someone who worked on a bookkeeping system without being an accountant (or bookkeeper in any sense), I'd say your challenge is that it's very possible to learn enough to build a decent system, assuming your engineering knowledge is strong.

I don't say this to blow my own trumpet, only to say that the non-engineering leadership at the company in question were very invested in the product details and making sure I understood basic accounting principles.

That said, I went from that role to working in a billing system that was originally built by a team where leadership weren't invested in the details. The result was a lot of issues and frustration from Finance leadership. Storing monetary values in float was not the biggest issue, either.

That being said, maybe branch out of just looking at accounting/bookkeeping and market yourself as "I know how money works in software systems." Your skills are extremely transferrable and knowing the Finance expectations will make it easier to make good design choices further upstream.

Your ideal job for utilizing both would either be to work for one of the ERP providers or accounting software providers (Oracle, NetSuite, Workday, Xero, etc.) or to launch your own targeting a specific need.

I would network to get over the ATS barrier and focus on smaller companies.

I work at a company that wants to start fixing their accounting system...what stacks are you familiar with?

I’m a general backend dev. I’ve used python, rust, and TypeScript to varying degrees along with SQL and noSQL databases. I’m pretty comfortable learning new stacks I haven’t used before.

Claude is familiar with all of them, and can describe each in terms of others without risking company code :-)

I may be able to help, I just sent you an email.

I agree with your sentiment here, but I think it helps to think of those domain specific questions as technical questions, just from a different technical domain. Finance is technical, mechanical engineering is technical, even sports management and sociology have big technical components too. I think having an expansive view of what technical competence is breeds the humility necessary to work across domains.

Yes, sure, if it helps you to think in those terms I guess that makes sense. Being a bit reductive perhaps but I think what it comes down to is having an interest in the "why" as much as the "what" and the "how".

ever since i started working for an insurance company, i realized that understanding the insurance industry is far more difficult than understanding the codebases. if the codebase is acting weird, at least i can step through it with a debugger!

Reinsurance is a beast. I spent a good time cracking layers and trying to implement the logic with a team.

I don't know about that -- I've always understood that to be the role of the product manager, to have all the domain knowledge.

It's the PM's job to work with engineering to ensure that the requirements are correct, and that the built product meets those requirements. And in an agile setting, those conversations and verifications are happening every single sprint, so something can't go for too long without being caught.

If you don't have a PM, then sure I guess your engineering team had better have deep domain knowledge. But otherwise, no -- it's not engineering's responsibility. It's products's responsibility.

There should be a system analyst. Every ERP or core banking big name works this way. Requirements have to be processed and then passed for implementation. In an agile setup rinse and repeat. As a side effect the whole team gains the domain knowledge. Or part of it.

> There should be a system analyst.

Exactly, this what we try to advance our tech people to, someone with the combination of domain+systems knowledge.

It results in more independent+correct action which is much more powerful than having to specify significant details to less knowledgeable people.

Pardon me an old story... I never built a double entry accounting system but decades ago I did build a billing system for a internet/telcom startup that grew to a modest 8 figures revenue.

By accident and not knowing any better as a young dev, I ended up building the billing logic from day one, and for better and worse building it in two places in the system (on a consumer-facing billing webpage, and on a separate backed process that generated invoices and charged credit cards.)

It turned out to be remarkably hard to keep them in sync. We were constantly iterating trying to get traction as we burned down our capital, releasing new products and services, new ways of discounting and pricing (per use, per month, first X free, etc), features like masterpayer/subaccounts for corporate accounts, user-assignable cost centers, tax allocation to those cost centers with penny allocation, etc such that new wrinkles and corner cases would keep popping up causing the numbers on my two screens/methods not to match.

Being personally responsible for the billing, I would go over all the invoices by hand for a couple days each month to insure they matched before we charged the credit cards and mailed out printed invoices as a final check to prevent mistakes. There was always/often some new problem I'd find affecting one or a small handful of customers which I would then fix the code before we billed. I never felt good letting go and not doublechecking everything by hand.

I thought about refactoring the billing logic to occur in one place to eliminate these mismatches and my manual crosschecking, but after a lot of thought I realized I wasn't comfortable with a single codebase and liked having two codebases as it helped me catch my own errors. I then just made it easier and easier to run automate and crosschecks between the two. The billing code was a little too gnarly to be proud of, but I was very proud of the outcome in how accurate our billing was, the lack of complaints and many near misses we avoided for many years. I do feel twinges of guilt for the complexity I left my successors but I still don't really regret it.

After that experience, the motivation for double entry bookkeeping has always made a lot of sense to me. I had sort of reinvented it in my own hacky way with double logic billing code to prevent my mistakes from causing problems for my customers...

Billing, or anything else involving money, is so easy to get wrong.

The data team I ended up leading at a previous company, had an unfortunate habit of "losing" money - it wasn't real money being lost in transit elsewhere, but records of something we should charge a customer.

Or if the team wasn't losing revenue, it was double charging, etc. etc.

Took us 3 years of hard work to regain the trust of the business leaders.

No offense but this sounds like a nightmare. It also sounds like you did a fantastic job achieving accuracy despite the complexity of the system. That’s something to be proud of.

This is The Nightmare. Devs building systems they barely understand, complexity leaking all over the place, and someone inheriting that awful job of keeping it running without having made any bad decisions themselves.

Software is full of these systems.

The more I learn about computers/software, the more I think it is a miracle that anything works at all.

This is also very similar to N-version programming :)

Good comment, thanks for sharing

No tests either? If you lose track of enough money every transaction that you can make an example of 'Every $5 purchased resulted in $4.98 in the transaction log' I think your problem is far, far bigger than not having double entry bookkeeping.

Who builds a financial system like that an considers it normal? The compensation is one thing, but you'd flee a service like that with all possible haste.

These guys. They said it themselves. “We could have built it right, but we didn’t.” They chose not to. It was not an accident. They made jokes about “dancing cents”. They did these things because there would never be meaningful consequences for doing them, and they knew it. They move fast, they broke things - money things - and they laughed about it. And now they’re lecturing people as if having willfully made these decisions gives them both moral and technical authority. This is magnificently pompous, startup VC culture nonsense.

Sounds like they refunded anything that went wrong so it's not really as bad as you make it sound.

critically, only when customers reached out. which means tons of people that weren't eagle eyed got defrauded.

No, they refunded when a customer complained. Any customer that didn't notice or didn't complain didn't get what was rightfully theirs. That should be criminal, if it isn't.

Refunds limit damages in a lawsuit, but don’t prevent legal issues.

Especially important when explicitly saying you’ve done these things.

I still didn't understand how they lost the money. Yes, double entry bookkeeping might have helped diagnose it, but how were they losing it?

Yeah that was my thought as well. Ledgers have tons of benefits but they’re not going to fix your dancing cents problem. You’re going to have bad number on the ledger.

Sure, maybe that points you to the bugs, but so would writing basic tests.

A ledger where you insist that every entry touches exactly two accounts, in a business where transactions regularly involve various types of fees, could easily misplace or misattribute a few cents here and there.

This type of business can also have fun dealing with accruals. One can easily do a bunch of transactions, have them settle, and then get an invoice for the associated fees at variable time in the future.

> where you insist that every entry touches exactly two accounts

A ledger is where every transaction balances to 0. It can involve multiple accounts, but the sum of all transfers between all accounts in a single transaction must sum to 0. This is the property of double entry that actually matters.

A good design principle is worth a 1000 tests.

How do you figure the story is true?

I don't understand why the author chooses to bring up the mantra “make it work, make it right, make it fast“ in a negative light, perhaps he misunderstands where the "make it fast" comes in?

to clarify: "make it right" is the second step, and until you make things work correctly ("right"), that's where the work stops, you have to make the system work soundly. the "make it fast", as in, optimize, comes in only after you have got it right, that all correctness and soundness issues are resolved perfectly. then you can start optimizing it (making it run fast).

it has nothing to do with delivery speed. it has nothing to do with working quickly. it's about optimizing only as a last step.

perhaps the author is lamenting the fact that it is possible for something to sort of "work", but to be so far from being "right" that you can't go back and make it right retroactively, that it has to be "right" from the inception, even before it starts barely working?

I think your last sentence is correct. I think author is saying that you can't make it work and THEN make it right with payment systems.

This is my opinion as well and I've been involved in the audit of a fintech system where auditors had to download EVERYTHING into excel spreadsheets and make the numbers balance before they would sign off on the books. That took a lot of time and money I'm guessing made a difference of at least .1 unicorns in the liquidity even that took place 3 years later.

I think author chose the wrong mantra too!

What happens in fast paced startups is that you ship what essentially is a MVP as soon as possible (i.e. you stop at the "make it work" step) because you need to build your customer base, finances, etc.

A better mantra would've been Facebook's "move fast and break things". But, that only works if you can fix the things later. You wouldn't do it if you're building an aircraft for example.

They said the engineering team followed this mantra, which includes the phrase "make it right". But they didn't. They didn't even bother to find out what they didn't know about fintech before they started building a product.

Given the context in which he used it, I think the misunderstanding you suggest in the first sentence is most likely. Immediately afterward he talks about the time pressure startups face.

Most of the comments here echo what the article is criticising. I can see countless threads of back-and-forth defending single-entry bookkeeping.

Sure, single-entry bookkeeping might be easier and more normalized, but sometimes it is a good idea to just stick with the systems and abstractions that have been developed over centuries.

Just use double-entry bookkeeping unless you definitely need something else. Sure, it might be icky for the programmer inside you, but I think you'll be thankful if you ever need to get actual accountants involved to sort out a mismatch.

On a related note: does anybody know of any good resources for programmers in payments and adjacent fields? Something like an "Accounting for Programmers"?

These articles might be a good starting point:

https://martin.kleppmann.com/2011/03/07/accounting-for-compu...

https://www.moderntreasury.com/journal/accounting-for-develo...

I read an article from Modern Treasury that advocated for mutable pending transactions to vary entry amounts by replacing entries¹, which is just about the worst idea I ever heard in the design of a DE system, and their reasoning boiled down to: if you're running a settlement system but are too lazy to implement a clearinghouse layer separately, no worries, just violate the entire DE covenant instead. So I'd take anything they write with a pinch of salt.

[1] https://www.moderntreasury.com/journal/how-to-scale-a-ledger...

Out of curiosity, how would a clearinghouse layer plug into this in practice? Thinking aloud, would you have an event stream of, say, EntryCreatedEvent, and the clearinghouse would provide streams of of EntryClearedEvent and EntryRejectedEvent - would you join those streams together to derive EffectiveEntry, EffectiveTransaction, PendingEntry, PendingTransaction based on whether all clearing is done on both sides?

I consider that first link, Accounting for Computer Scientists, as the canonical guide for computer scientists as to wtf double entry accounting is and why it's the right way to do it.

Kleppmann's Designing Data Intensive Applications is also pure gold.

Worth reading once a year imo

A Basic Introduction to Accounting[0].

[0]: https://www.winstoncooke.com/blog/a-basic-introduction-to-ac...

> On a related note: does anybody know of any good resources for programmers in payments and adjacent fields? Something like an "Accounting for Programmers"?

Get a grip on the accounting basics first. I built my own bookkeeping system with double entries and realized the design and the programming was the easy part.

If told people that I was using a database system where one and 100th of the data was missing after every 10 transactions would you seriously take my advice as an engineer blog post like these that seem to be calmly focused toward an advertisement of a person’s or a group’s promotion and that it’s just introducing concepts.

Wat

The user name "zitterbewegung" sounds like a German. They build sentences like that.

You have to think about a sentence like a stack and read it slowly and carefully and take note of the push and pop operations.

Germans are like RPN where everyone else is a regular calculator.

I'm afraid it's not just sentence structure that is the problem here:

> one and 100th of the data was missing

No idea what that means

> an advertisement of a person’s or a group’s promotion

What in god's name is an "advertisement of a group's promotion"?

I don't know German I was using the voice recognition on an iPad.

For a glimpse at the _real_, human consequences of this kind of slipshod mentality toward money shuffling software see the Post Office scandal[1].

https://en.wikipedia.org/wiki/British_Post_Office_scandal

Anything that moves money should be treated with upmost seriousness and be aware of as many historical mistakes as possible.

> https://en.wikipedia.org/wiki/British_Post_Office_scandal

The four-part miniseries with Toby Jones (mentioned above in §Media:Dramatisation) was really good and goes over things pretty well:

* https://en.wikipedia.org/wiki/Mr_Bates_vs_The_Post_Office

Does anyone have a good explanation for why a ledger database should care about credits and debits and normal balances? This has always seemed more natural to me as a front-end / presentation layer concept. Your ledger entries always sum to zero for each transaction, your income account has a negative balance, and you display it in a sensible manner.

I’m also surprised that this whole article starts by discussing stock trading but has no mention of how to represent stock trades. I assume they are “Sagas” consisting of money moving from the customer to the clearinghouse (or prime broker or PFOF provider or whatever) and shares moving from that provider to the account at which the shares are held. And maybe other associated entries representing fees? It seems to me that this is multi-entry accounting, which is quite common, and that entries don’t actually come in pairs as the article would like us to think.

That part of the article felt quite wrong to me as well. I've built accounting systems that worked well for a decade, where internally the values were a single amount column in the journal table. If it was a debit, it'd be positive, if a credit, it'd be negative.

In fact, we could call these values yin and yang, for all it mattered.

Also, I'm not able to really follow what he means by "money = assets in the future".

Money is money, but if you wanted to track the intermediate state until the customer gets receipt, you would use an In Transit account (Good In Transit / Service In Transit etc.)

Yet, it doesn't change the fundamental definition of the value in the accounting system. I think the author confuses an engineering concept (sagas, or thunks, or delayed but introspectable/cancellable actions in general) with accounting.

> Also, I'm not able to really follow what he means by "money = assets in the future".

I’m guessing it’s one of two things:

1. A transaction might fail. If you enter a transaction into your bank’s website or your credit card company’s website, you should probably record it in your ledger right away. But the transaction might get canceled for any number of reasons. And the money will not actually move instantly, at least in the US with some of the slower money moving mechanisms.

2. In stocks and other markets, settlement is not immediate. A trade is actually a promise by the parties to deliver the assets being traded at a specific time or range of times in the future. One probably could model this with “in transit” accounts, but that sounds quite unpleasant.

FWIW, I’ve never really been happy with any way that I’ve seen accounting systems model accruals and things in transit. I’ve seen actual professional accountants thoroughly lose track of balance sheet assets that are worth an exactly known amount of cash but are a little bit intangible in the sense that they’re not in a bank account with a nice monthly statement.

IMO it's entirely wrong, and it also makes it a lot more difficult to programmatically create transactions with 3+ legs (For example: A payment with a line item + sales tax).

I think the author is just wrong on that point, but the rest is sound. (Source: I've built bookkeeping software)

It is best explained by common scenarios an Italian merchant in the Middle Ages experienced. The basic concept is Assets==Liability (plus Equity). Where positive Assets are entered on the left hand side (debit). And positive Liabilities are entered on the right hand side (credit). In accounting, debit and credit just means left and right.

1. Merchant takes out a loan for $5,000 and receives $5,000 in cash. • Assets (Cash) increase by $5,000 (Debit). • Liabilities (Loan Payable) increase by $5,000 (Credit). • Equity remains unchanged.

2. Merchant buys inventory for $1,000 cash. • Assets (Cash) decrease by $1,000 (Credit). • Assets (Inventory) increase by $1,000 (Debit). • Total assets remain unchanged, and liabilities and equity are unaffected.

3. Merchant sells all inventory for $1,500 cash. • Assets (Cash) increase by $1,500 (Debit). • Assets (Inventory) decrease by $1,000 (Credit) (recording cost of goods sold). • Equity (Retained Earnings) increases by $500 (Credit), representing the profit ($1,500 sales - $1,000 cost).

4. Customer1 deposits $500 in cash for future delivery of goods. • Assets (Cash) increase by $500 (Debit). • Liabilities (Unearned Revenue) increase by $500 (Credit). • Equity remains unchanged.

5. Customer1 transfers half of the future delivery of goods to Customer2. • No changes to assets, liabilities, or equity occur at this point. The merchant’s obligation to deliver goods (reflected as Unearned Revenue) is still $500 but now split between two customers (Customer1 and Customer2). Internal tracking of this obligation may be updated, but the total financial liability remains the same.

Actually it is clear as long as you remember that main point you made: debit and credit just means left and right.

We are all spoiled by thinking of debit/credit as equal to decrease/increase respectively because that how we interpret our bank accounts. That understanding totally collides with formal accounting where debit/credit DON'T mean decrease/increase respectively. I think this is the root cause of all confusion about double-entry accounting. I may be wrong about this, happy to be corrected but that is the bit my brain grinds against when trying to make sense of things.

E.g. I replaced all instance of debit with "Left" and credit with "Right" in your example:

    1. Merchant takes out a loan for $5,000 and receives $5,000 in cash. • Assets (Cash) increase by $5,000 (Left). • Liabilities (Loan Payable) increase by $5,000 (Right). • Equity remains unchanged.

    2. Merchant buys inventory for $1,000 cash. • Assets (Cash) decrease by $1,000 (Right). • Assets (Inventory) increase by $1,000 (Left). • Total assets remain unchanged, and liabilities and equity are unaffected.

    3. Merchant sells all inventory for $1,500 cash. • Assets (Cash) increase by $1,500 (Left). • Assets (Inventory) decrease by $1,000 (Right) (recording cost of goods sold). • Equity (Retained Earnings) increases by $500 (Right), representing the profit ($1,500 sales - $1,000 cost).

    4. Customer1 deposits $500 in cash for future delivery of goods. • Assets (Cash) increase by $500 (Left). • Liabilities (Unearned Revenue) increase by $500 (Right). • Equity remains unchanged.

    5. Customer1 transfers half of the future delivery of goods to Customer2. • No changes to assets, liabilities, or equity occur at this point. The merchant’s obligation to deliver goods (reflected as Unearned Revenue) is still $500 but now split between two customers (Customer1 and Customer2). Internal tracking of this obligation may be updated, but the total financial liability remains the same.

I find this much more easier to reason with.

I understand this. But we’re talking about computers, not Italian merchants. Italian merchants had actual pieces of paper. Computers have tables and views and frontends that are separate from the tables.

Any self-respecting accounting system should be able to produce a balance sheet that matches the conventions you’re describing. I don’t think it follows that the actual numbers in the database that get summed to produce the total liabilities should be positive.

Thank you for an example. But I don't see how it explains why debit/credit should be used instead of simple signed amount. Like how Transaction(from, to) where `from` and `to` are Entry(account, credit|debit, unsigned amount) make things easier than Entry(account, signed amount).

You basically used different labels for positive or negative amount in the example.

yes, agree, I think a 'source' 'destination' model is significantly more straight-forward. Just record the 'source' account and the destination account and you essentially end up with a ledger as a directed graph (Martin Kleppmann wrote a great post on it)

I also wrote a super short post on how to model such a system on postgres https://blog.nxos.io/A-simple-double-entry-ledger-with-sourc...

Blockchain actually kinda nails it, that's in essence a source/destination ledger, no 'postings' or similar needed, and from a balance calculation POV has been working pretty well

One reason this model isn't applied in accounting, in my personal view :), is simply historical and the fact that the number 0 didn't exist when accounting principles were created.

Wrote another post on how to model debit/credits on a source/destination ledger here: https://blog.nxos.io/Debit-and-Credits-on-a-Source-Destinati...

It's very straight-forward, you just have to accept that asset accounts have negative balances and present the absolute amount instead of a negative amount in a view.

It isn't always clear which is "source" and which is "destination" and now you need a bunch of new conventions about these things. Accounting already has these (admittedly arbitrary) conventions so we might as well use those.

> you essentially end up with a ledger as a directed graph

The page contains a comment from Matheus Portela who pointed to a blogpost of his about "Double-Entry Bookkeeping as a Directed Graph" [0].

"I've also had the same problems you described here and double-entry bookkeeping is the way to go for financial accuracy. As a programmer, things clicked when I realized this system is an extended directed graph.". It turned out that: "Hi Matheus! Would you believe me if I told you that I read your post in preparation for this article?"

[0] https://matheusportela.com/double-entry-bookkeeping-as-a-dir...

Source/destination seems fail on actual DB implementation. How to sum all entries for a particular account? It could be on either side. It complicates queries and could trigger unoptimal query plans.

okay but how do you model a three-party transaction? say you want to collect fees or taxes

In double-entry, a transaction is the tuple (amount, credit_account, debit_account).

In singly-entry, it is the tuple (amount, account).

> In double-entry, a transaction is the tuple (amount, credit_account, debit_account).

Every “double entry” accounting package I’ve ever used can easily handle transactions that are awkward in this schema and transactions that don’t fit at all.

Moving $1 from one current account to another? I guess you declare that the $1 needs to be a positive amount, but your two accounts have the same normal balance, and calling one a “debit account” is a bit awkward.

Adding an accounts payable entry that is split between two expense accounts? Not so easy.

I've never understood that either. As a computer guy it always struck me as a redundancy -- the kind of redundancy where one thing will always eventually be wrong.

I assumed it has to do with the fact that it was invented for bookkeeping by hand. Accountants assure me that it's even more important with computers, but I've never managed to figure out how that works

Double entry accounting is still error prone, but single entry accounting is fraud prone.

A single debit can result in many credits.

A single record can (and will) be lost. Network issue, db issue, etc., some transactions will not make it to the db.

With double entry you at least have a chance of reconciling the lost transaction.

> As a computer guy it always struck me as a redundancy -- the kind of redundancy where one thing will always eventually be wrong.

That's the purpose. If you have a system with no redundancy, it's equally true that something will always eventually be wrong. But in that case, you'll have no way of knowing what's wrong.

With the redundancy, you can detect problems and often determine what happened.

It is not a redundancy. Actually the meaning of a transaction encoded in a pair of accounts. You will need to store it somewhere anyway in a real world scenario, but predefined account pairs is the universal language. The accounting system.

I've been doing this so long that I've finally realized that if someone can't explain why something is the way it is, that means it's wrong, or at least some arbitrary choice among several equally good alternatives.

Neither have I. It always seems like a massive cargo culting. Human accountants are liable to make very different kinds of mistakes than computers.

Double entry book keeping implements an error correction and detection algorithm.

> This has always seemed more natural to me as a front-end / presentation layer concept.

Consistence is a property of the backend, if that is wrong there is not hope for later

> Your ledger entries always sum to zero for each transaction, your income account has a negative balance, and you display it in a sensible manner.

'sensible manner' is the problem here. The data/money will be diverge with time, and without proper storage of the data it will by impossible to figure out.

The problem here is NOT store 'a transaction'. That with a RDBMs works. Is to store the FLOW of MANY transactions and the divergent ways things works.

Like, your bank is telling you has $100 and your system $120. And your system sum right, but the bank rules.

Or when you do a return and cents are lost in the interchanges and chargebacks.

Or, just wrong data entry, sync, import/export, etc.

---

The way to see this is that `double entry` is a variation of `inmutable data that don't mutate and always track the flow of it' that is golden for business apps.

Great article. I have an observation to the "engineers should know this and do good engineering" though: I work for a payments company and there is a fundamental career problem with becoming the ledger ninja: It's not enough work, and it's eventually done!

I've seen the following major phases of this work: 1) Build the ledger (correctly), and it will work well for a while. 2) Add convenience code for the callers, assist finance in doing reports/journaling, fix some minor bugs, take care of the operational bits (keep the database up). 3) Reach the scaling limits of your initial approach, but there are some obvious (not trivial) things to do: re-implement the transaction creation directly in the database (10x perf gain), maybe sharding, maybe putting old tx into colder storage, etc.

This is spread out over a while, so I haven't seen it be a full-time job, even at real startup-level (+10% MoM) growth. Even if it was, that's one person, not a whole team. I understand engineers that instead are pulled towards projects where they are in higher demand.

In another comment somebody said ledger systems are trivial when done right and super hard when done wrong - so if you did a good job it kinda looks like you just created 3 tables and some code. That seems thankless, and job searching as this type of specialist is harder than just being a generalist.

I suspect this is a "small company" problem which don't sell too many different things. A larger enterprise might have a platform for connecting businesses with businesses and businesses with customers. They might sell services (subscriptions) but also one-time purchases which require different forms of revenue recognition. Those might even be split up across different revenue streams. You end up building sub-ledgers for each one of them because the ERP can't handle the scale. Oh, and you're a public company so make sure everything is SOX compliant and easy to audit. Ah, and you operate on a global scale so do all those things in different currencies for different legal entities.

There's a reason Stripe is as successful as it is. And then there's a world where a company outgrows Stripe.

There are worse career choices ("prompt engineer" LOL) than financial engineering.

Stripe isn't a bookkeeper/accountant. Something like http://Pilot.com is though.

stripe on bookkeeping: https://stripe.com/guides/atlas/bookkeeping-and-accounting

I’ve built my career on cleaning up should-have-used-a-ledger messes. it’s hard but there’s always another company that needs it, and I get bored staying in one place.

recently I discovered that in a medical billing context the system is way, way weirder than I had seen before. I shipped the three tables, but getting everything into it seems like it might be an endless battle

If you are trying to reinvent parts of banking, you could maybe start with what already works as a thought experiment.

Here's an example of a core banking deposit transaction schema that has been extensively battle tested in many small & mid-size US institutions:

https://jackhenry.dev/open-enterprise-api-docs/operational-d...

You may note fields like "Effective Date" & "Affects Balance/Interest", which imply doing this correctly may involve exploring some interesting edge cases. Wouldn't it be cool if you could just cheat and start with an approach that already considers them?

I don't know what the fuck I just read here. I've worked on ledgers for years with massive codebases with millions of loc and I've never seen an error of "a few cents". Feels like I just read a alibi. Don't develop financial software from scratch.

That was my reaction too.

It was like reading an engineering team saying their attempt to design a new lightweight mountain bike failed. It turned out saving weight by omitting brakes wasn't a good idea, and their attempt to fix it by riding beside each bike and manually slowing it down wasn't too popular either. Then they have the hubris to follow that up with, "you can avoid the same mistakes by reading what we have to say on bicycle design".

The lessons they can take away have very little to do with engineering or FinTech. I'd file it under doing a bit of research before committing to any major task. Basic accounting principles are centuries old now. They could have learnt about them by buying a high school text book, reading it cover to cover and use doing the exercises. It would have taken them less than a week.

Admittedly that only gives you the "how", not the "why". You realise much, much later that the engineering equivalent of double entry accounting is two systems design by different teams that are constantly comparing their outputs. By the time you've figured that out, it's caught so many errors you realise "holy shit, these systems are really easy to screw up, and are under constant attack by malicious actors".

There is a hidden trap at every step - floating point being imprecise, currency rounding not being reversible, tax calculations being one way, ACID being a hard requirement. I'm being this mob screwed up tax calculations. Floating point throwing out 1 in 100 million transactions was one of the joys they never got to experience.

The lesson here is not to never develop from scratch. In fact it may be faster and safer to do so than to vet all of the open source junk that modern ByteByteGo-system-designers think they need. TigerBeetle and Jepsen-proof-engineering are paths to dig down here, along with basic knowledge of computer number storage.

The lesson for me here is that there is no substitute for knowing or caring about what you're doing. Dancing cents don't happen when eng cared about correctness in the first place. I feel like I just read a college freshman explaining how there's no possible way they could have passed their first exam. Better yet he's here to sell us on his thought leadership without ever explaining why he failed the first exam.

I read the intro and decided I wouldn't learn much from someone willing and eager to admit he eats paint chips.

I've seen it, sighed, fixed it, and ensured the necessary accounting adjustments assigned the loss to the business instead of the customer. Seems that last step, which is the most important one, wasn't followed here, which is the real shame.

To be fair, I have seen it before because of a modification I did to the code base. It was when I first started working with it so I didn't really understand how to correctly implement extensions. But still, i was able to see the error in the ledger and connected to my change.

I had the same thought. I mean, were they actually using ordinary floating-point numbers to represent amounts in their ledger? This sets off so many alarm bells.

It's a suicide to build a finance system or similar without double entry ledger.

Worse, I've worked at where the transaction reference id from vendor is not recorded, it's lost so the past data cannot be reconciled!

is there no individual accountability regime in the US?

in the UK, as an engineer, if I'd built this I would expect the regulator to come after me personally for not ensuring the system had adequate controls to protect clients money/investments

with a potentially unlimited fine + prison time

Has that ever happened? It's incredibly hard to prosecute directors in the UK for obvious malfeasance. I have never heard of a software engineer being sanctioned for crap code.

for engineers[1] it is relatively new, having only been introduced in the last couple of years

[1]: technically those performing a Certified Function

It would be nearly impossible to prosecute for just bad code. It would require more and is limited very small scope of people.

>is there no individual accountability regime in the US?

Here in the US, programmers like to call themselves Engineers, forcing everyone else to use the term "Professional Engineer" or "Licensed Engineer" or some other modifier before their title. I hate it, I wish they would stop, but it's not going to happen.

Software here is a wild, wild, West. The motto most live by is "move fast and break things"... even when those things are people's lives.

There’s a few more professionals that are called engineers.

Railroad locomotive operators, ship engine operators.

The name precedes the creation of licensed tertiary education level engineers.

A lot of people seems to ignore the fact that licensed professions that require an accredited diploma in a tertiary level education program is a relatively recent feature of our societies.

The PE thing is more than 100 years old in the US. By 1947 every state had a PE licensure program. It has nothing to do with programmers.

I leave you with the 2008 financial crisis as exhibit A on exactly how nobody gets in trouble for lack of financial accountability

Not entirely accurate. Exactly one bank (a small Asian American community bank, the 2651st largest in the U.S.) was prosecuted after the subprime mortgage crisis.

https://en.wikipedia.org/wiki/Abacus:_Small_Enough_to_Jail

The secret is to have everyone in on it. Everyone is guilty but nobody is quite culpable enough to punish.

The low level guys were just doing their jobs, and each individual transaction was mostly legal. A few weren't but it's hard to sort out which ones. Maybe the management should be responsible for ensuring that nobody ever did anything illegal, but they can't really supervise everything all the time, can they?

Poof. Guilt is just a rounding error that all rounds down to zero. The government passes some new regulations to prevent that particular scenario from happening again, and the same people set about finding a new scam.

Not for software that I'm aware of.

For certain regulated professions there is, if a building falls down due to a bad design the professional engineer (PE) that signed and sealed the plans can be held personally liable.

I don't see how a rank and file programmer would ever be personally responsible for their code. You can blame management for forcing untested or known flawed logic, but not some shmoe that pushes an "off by 1" bug while working weekends and late nights with no testing and hard deadlines.

Programmers should ensure they understand the requirements of what they are being asked to build, and not just blindly build things. If I was asked to build an accounting system, I would insist on speaking with an accountant to understand the requirements. If I was asked to work on a medical imagining system, I would want to be working with a qualified radiologist and probably a PhD holder in a relevant field too.

Overconfidence can quite literally be fatal.

Pushing a bug, yeah that happens.

Deliberately implementing a financial system that ignores established (and probably legally required) accounting practices? That's kind of like a structural engineer willfully disregarding the building code because that's what management asked for.

Can you please provide a reference to the regulators coming after an engineer in the UK.

Also you say regulations in the UK have been changed recently.

I'm not aware of regulations that apply to software engineers.

Depends if the particular activity is regulated or not.

a stock trading platform, as described in the article?

Are you a professionally qualified engineer?

It really irks me that the author assumes I know how double-entry accounting works and doesn't mention a single sentence about it. I read half way through the article and couldn't follow it, except that single-entry is bad and double-entry is good.

A double-entry system is one where you can't change the balance of an account, you can only record transfers between accounts. This means it's impossible to forget to update the other side of a transaction, it's a single step. A consequence of that is you can check that the sum of all accounts is always 0.

In practice you have virtual accounts like "cloud expenses" and "customer subscription" that only go up/down over time, to be the counter-party for transactions in/out of your company. So it's not impossible to mess up, but it eliminates a class of mistakes.

Hm, but what's the second entry? Do you not just add a single entry from X to Y?

Entry1: Account A: receives 1000 from Account B

Entry2: Account B: sends 1000 to Account A

And from GP:

> A consequence of that is you can check that the sum of all accounts is always 0.

Entry1 + Entry2 = 1000 + -1000 = 0

Amusingly, when I made my own personal money tracking program over a decade ago for my bank accounts, this was how I implemented transfers between them just because it was simpler to do. Years later when I heard this name, I also had trouble understanding it because I assumed I had done it the bad way and banks did somehow did something more in-depth.

The change to X and the change to Y are the two entries.

In practice most systems allow for more than two entries in a single transaction (ex: take from bank account, give to salary, give to taxes) but always at least two, and always adding up to 0.

Could have been clearer but it is there. Here is the relevant section. In short single entry is basically the Account view and Double entry is the full ledger. (Called double because of the hard requirement that all Entries come in pairs.)

> Ledgers are conceptually a data model, represented by three entities: Accounts, Entries and Transactions.

> Most people think of money in terms of what’s theirs, and Accounts are the representation of that point of view. They are the reason why engineers naturally gravitate towards the “simplicity” of single-entry systems. Accounts are both buckets of value, and a particular point of view of how its value changes over time.

> Entries represent the flow of funds between Accounts. Crucially, they are always an exchange of value. Therefore, they always come in pairs: an Entry represents one leg of the exchange.

> The way we ensure that Entries are paired correctly is with Transactions. Ledgers shouldn’t interact with Entries directly, but through the Transaction entity.

That’s more than halfway through the post. The GP’s point that railing for pages about how X is good and Y is bad before defining them stands.

I've been trying to figure out double-entry accounting for years now and still don't get it. Most explanations are along the lines of "Here is a simple explanation that you are guaranteed to understand: <proceeds to describe the procedure>" which lacks intuition on why this is useful. I suspect I would need to do a gig as an accountant and run into some error conditions that double-entry solves to really grok it.

Edit: no offense but sibling comment is an example :P

Double-entry: each row of your DB stores {debit account, credit account, amount}.

Single-entry: each row of your DB stores {account, delta}.

With double-entry you are guaranteed via your schema that sum(debit delta) = sum(credit delta), that's it. Money is "conserved".

It's easy to denormalize the double-entry into a single-entry view.

That kleppman article talking about movements as edges of a DAG is the only place this is ever talked about clearly.

From what I've learned from a few guys, double ledger accounting is a technique which optimizes for consistency, error detection and fraud detection. Each movement of money, or material should always be written down in two or possibly more places. Ideally by independent people or systems.

Once you pair this with another entity or company doing the same, it becomes very hard for money or goods to vanish without the possibility to track it down. Either your books are consistent (sum of "stock" + sum of "ingress" - sum of "egress" - sum of "waste" makes sense), or something is weird. Either your incoming or outgoing goods match the other entities incoming or outgoing records, or something is amiss in between.

This is more about anomaly detection, because paying a clerk can be cheaper than losing millions of dollars of material by people unloading it off of a truck en-route.

> Most explanations are along the lines of "Here is a simple explanation that you are guaranteed to understand: <proceeds to describe the procedure>" which lacks intuition on why this is useful

Double entry accounting is useful because it enables local reasoning. Let me explain why! If you've remembered those other explanations, you hopefully remember the fundamental equation of accounting:

    assets - liabilities = shareholder equity

Well, you can also define this a bit more broadly as

    assets + Δassets - liabilities - Δliabilities = equity

In other words, changes in assets or liabilities must also balance. I sometimes think of these as "income" and "expense," probably because GNUcash has accounts of that type. If you rearrange that expanded equation you get

    (assets - liabilities) + (Δassets - Δliabilities) = equity

If the grouping on the left represents the state before a transaction, and the grouping on the right represents a single transaction, then we get a super useful auditing tool: as long as the book is balanced before any given transaction, the book will remain balanced after the transaction as long as the transaction itself is balanced. You can now reason about a narrow set of things instead of the whole book!

In practice what this means is if your system is losing cents on every transaction as OP article states, each transaction should be flagged as imbalanced. From there it should be super easy to see why, since you can examine a single transaction.

To achieve this you need all entries / actions to have a unique transaction ID to group by, and a notion of which accounts are liabilities and which are assets, etc. As OP article mentions, there's a ton of implementation nuance.

Agreed, exactly the same here. I was hoping the article would explain, but sadly, no.

Imagine you’re running a physics simulation. With conservation of energy: if energy shows up somewhere, it went down somewhere else. We shouldn’t assume a balance changed on its own without any other effects in the system. If you start “losing net energy” you can see what’s going on.

> A double-entry system is an accounting method that tracks money at both its source and destination.

There's an entire section on double-entry accounting in the article. The tl;dr is that if you take money out of an account, you need to place money in another account and vice versa. So, you have a table called "accounts receivable" which track money the company is owed. If you actually get paid, you remove money from "accounts receivable" and add money to the "cash" account, instead of just increasing the amount of cash you have.

It makes it much more difficult to lose track of money or have it stolen. In a single-entry system, I could receive money from a customer for services owed and just keep it for myself.

If you're looking for a pre-built double entry accounting Postgresql schema that provides database-level integrity checks:

https://github.com/adamcharnock/django-hordak

I created and maintain this, along with a couple of others. It is built for Django (so great if you're using Django), but extracting the schema wouldn't be too hard. It also has MySQL support, but the integrity checks are more limited.

(Side note: I'm a freelancer and available!)

There is a bigger lesson here. When you are creating something new it pays to understand what came before. Do you need to disgard everything?

I gave a boorish lecture to a junior accountant recently...

When you make things up yourself it is just you and Excel. When you use double entry you have 100s of years of history to fall back on. https://en.m.wikipedia.org/wiki/Double-entry_bookkeeping

Though there's a weirder life lesson here : a lot (most / all ?) of successes come from founders not having enough experience to know just how difficult their enterprise is.

More experienced founders would take a look at it and immediately nope out of it, and so not achieve success (after a grueling amount of work).

Accounting systems are super hard when you do them wrong and kind of trivial when you do it right.

There is no in-between.

Martin fowler wrote quite a bit on the subject and it's a good match for event-sourcing.

literally every startup I’ve worked at has had busted accounting system that had to be painfully and expensively rewritten as a proper ledger. We should be teaching “Accounting for Programmers” as a required course in CS programs, or at least spreading memes about how money data must be append-only and immutable or you are being negligent in ways far more serious than most programming errors

The Software Engineering department next door from my CS department had a mandatory course on accounting. I don't think CS needs accounting, but I did learn two skills that would have helped here: why one should never use floating point for money (which may not apply here, who knows), and how to stick to a formal specification. "Money is neither lost nor created" is the most trivial example one can think for an invariant.

Relatedly, some months ago I asked how to correctly record food in a double-entry bookkeeping system [1] and it triggered a 1300-words, 8-levels-deep discussion. We should remember more often than accounting is a degree by itself.

[1] https://news.ycombinator.com/item?id=39992107

Haha, same. https://news.ycombinator.com/item?id=39994886

I kept digging and digging on a "sell some lemonade for $5" example, and ended up at:

  - $5 debit to cash (asset => debit means +5)
  - $5 credit to revenue (equity => credit means + 5)
  - $X debit to cost of goods sold (liability => debit means - X)
  - $X credit to inventory (asset => credits mean - X)

A double-entry for the money, and a double-entry for the inventory, for a total of 4 entries.

It's too complicated for me. I'd model it as a Sale{lemonade:1,price:$5} and be done with it. Nothing sums to zero and there's no "Equity + Income + Liabilities = Assets + Expenses" in my version.

But this is HN, and I think a lot of people would call my way of doing things "double" because it has both the lemonade and the money in it. So when I say I'm not sold on doing actual double-entry [https://news.ycombinator.com/item?id=42270721] I get sweet down-votes.

if you’re just doing first-party sales, the single-entry model you described is probably fine!

but every startup these days wants to become a marketplace where you are facilitating multiple third-party lemonade vendors and taking a cut for letting them use your platform. In that case, the flow of money quickly gets too hard to understand unless you have double-entry

I think because programmer find it confusing about translation of model in computer to real world

Accounting predates computer by hundreds of years, there are better way to do certain thing of course but we must follow convention here because that's the norm and everyone understood

computers evolved out of tabulating machines and tabulating machines were designed to make pen&paper accounting easier. These are our roots and we should appreciate them

> Instead of using negative numbers, Accounts have normal balance: normal credit balance literally means that they are normal when its associated entries with type credit have a total amount that outweighs its associated entries with type debit. The reverse is true for normal debit balance.

But that is an interpretation made by the viewer. A customer typically is an asset account, whose balances are in the debit column. But if we somehow owe them money because let's say they paid us an advance, then their balance should be in the credit column. The accounting system need not bother with what the "right" place for each account is.

It is quite practical to have only a simple amount column rather than separate debit/credit columns in a database for journal entries. As long as we follow a consistent pattern in mapping user input (debit = positive, credit = negative) into the underlying tables, and the same when rendering accounting statements back, it would remain consistent and correct.

> It is quite practical to have only a simple amount column rather than separate debit/credit columns in a database for journal entries. As long as we follow a consistent pattern in mapping user input (debit = positive, credit = negative) into the underlying tables, and the same when rendering accounting statements back, it would remain consistent and correct.

Another benefit of Credit / Debit side on double-entry bookkeeping is you need to balance both side in a single transaction. Say if the user account 2003201 is in Credit and it got an addition of 1000 value, a same value need to be added on Debit side. If it's (1) a cash topup, then 1000 value need to be added to Cash account (let's say 1001001) on Debit side. Otherwise if it's a transfer (2) from another user account 203235, then the account need to be Debited 1000 value as well.

It's Asset = Liabilities + Equity, while the left equation is Debit (which increase value when a Debit transaction happen, and the right equation is Credit, which increase when a Credit transaction happen. In (1) case, the cash account increase since it's on Debit account, while in (2) case, the user account decrease because it's a debit transaction on Credit account.

With negative/positive, the invariant would be sum(amount) = 0; with your approach, it would be sum(debit-credit)=0. Both are valid, it is just two ways of expressing the same thing.

I think it is useful to think about double-entry book-keeping in two layers. One is the base primitive of the journal - where each transaction has a set of debits and credits to different accounts, which all total to 0.

Then above that there is the chart of accounts, and how real-world transactions are modelled. For an engineer, to build the base primitive, we only need a simple schema for accounts and transactions. You can use either amount (+/-ve), or debit/credit for each line item.

Then if you're building the application layer which creates entries, like your top-up example, then you also need to know how to _structure_ those entries. If you have a transfer between two customer accounts, then you debit the one who's receiving the money (because assets are marked on the debit side) and credit the other (because liabilities are on the credit side). If you receive payment, then cash is debited (due to assets), and the income account is credited (because income balances are on the credit side).

However, all of this has nothing to do with how we structure the fundamental primitive of the journalling system. It is just a list of accounts, and then a list of transactions, where each transaction has a set of accounts that get either debited/credit, with the sum of the entire transaction coming to 0. That's it -- that constraint is all there is to double-entry book-keeping from a schema point.

> When I decided to write a post on ledgers, I already knew there were a few good resources out there to help me.

One that's not referenced in this article and compile all of them is: https://github.com/kdeldycke/awesome-billing#readme

I saw this so many times. It is rather depressing. Even experienced devs hardly know enough about floats, rounding etc to implement these things, but they do. We were asked to troubleshoot a project for an online insurer ; the devs did two conversions; js and php. So the amount from the frontend ended on the server and got converted and got into the db as an entry. Definitely depressing how frontend numbers are converted and actually this means something to the end result. But the php conversion was also wrong, so it was often cents off. I believe people should never be allowed to have a 'senior developer' job again, but this is really quite commonplace. Most companies don't have enough revenue to notice or care, but given how often we see it, I would say it's actually normal.

While I've never worked on financial systems myself, I was told many times that monetary amounts should always be stored and manipulated as integer amounts of smallest units of currency (cents). One reason being that integer math is always exact, while floating-point isn't. And when money is involved, you want your math to be exact so everything always adds up every step of the way. So I'm always calling it out when someone is using floats to store money.

well, you could do this, but it would make world much more complex when working with the numbers in the code - todays highlevel languages like Java or C# have a "decimal" datatype which should be used for money/accounting, these types also offer several digits of accuracy, if required.

Everyone keeps focusing on double entry book keeping, but that's a ledger that's more suited to manual book keeping. We're in the computer age, people should be using the richer accounting model of REA:

https://en.wikipedia.org/wiki/Resources%2C_Events%2C_Agents

You can see that this model has all of the features discussed in the article, and then some, and REA events map naturally to something like event sourcing. You can project a REA dataset into a double entry ledger, but you often can't go the other way around.

I read the source you sited but it seems like a foundational part of the model is basically double book accounting:

“At the heart of each REA model there is usually a pair of events, linked by an exchange relationship, typically referred to as the "duality" relation. One of these events usually represents a resource being given away or lost, while the other represents a resource being received or gained.”

This is what I see as the main difference between single book accounting and double book accounting, with REA having some OO things added to the model to more accurately represent business objects when a computer is used. What am I missing about REA that makes it better than double book as implemented in the way this post was talking about implementing it?

Double entry bookkeeping (DEB) records entries in ledgers. There are no ledgers in REA, therefore REA is not double-entry bookkeeping.

You could argue that when representing both DEB and REA in an entity-relational model, they might have some similar looking tuples and relations, but that does not entail that they have the same data model. As I said in my initial post, REA is a richer data model that captures more information. You can reproduce the ledgers of DEB from a REA system, but you cannot go the other way in all cases.

I feel there is no excuse. Startups cut all sorts of corners but why? Get rich at other people's expense?

The tech founder should know their shit prior to building it so that time and runway and deadlines are no excuse. Really if you are doing fintech you should have employment experience and understand all the operations well.

Otherwise they are no better than say home builders who mismanage and go bankrupt. Or less geneously, they are conmen.

I didn't understand this part, can someone give examples of good and bad approaches for both credit and debit accounts?

Could someone explain something fundamental to me about the need for double-entry accounting? Why can't your source of truth instead be proper SQL tables following these schemas (using Python tuple/list notation for convenience)?

  transactions: [
    (txID, timestamp, [
      (accountID, delta, otherInfo),
      ...
    ], reason),
    ...
  ]

  accounts: [
    (accountID, routingNumber, accountNumber, ownerID),
    ...
  ]

Crucially, notice that "accounts" don't track balances here, and so I'm asking: why would I need TWO entries per transaction, and why would I need to do my own tracking of the balance of every account, when I can just keep a single entry per transaction in a proper ACID database and then build any view I need on top (such as running balances) with a battle-tested projection mechanism (like a SQL View) so that I still track a single source of truth?

what you’ve sketched out already looks like a double-entry bookkeeping system, if you have the constraint that all of the deltas in your array on your transaction sum to zero (so money always _moves_ and is not created or destroyed)

calling it “double” might be misleading - in the ledger systems I’ve worked on, a transaction that has only two entries - a single source and a single destination - is actually rare. In practice, a payment tends to have a source, the intended destination, a secondary destination for platform fees, possibly another destination for taxes.

And much more complicated situations are possible - I currently work in medical billing where there are multiple virtual internal accounts involved in each transaction for different kinds of revenue (including, for example, money we billed for but don’t expect will actually get paid)

so a transaction becomes a bundle of a set of several debits or credits that happen simultaneously and sum to zero. if you have that, you have double-entry bookkeeping, even if you schema puts it all in one table

Ahh thanks so much, that clears it up for me! Every description of double entry I'd seen so far seemed to have the notion of keeping a running balance for each account (and then having a separate transaction for each one where they all sum to zero) as a central feature, so I thought avoiding that would make it no longer double-entry, and it always puzzled me why that was so crucial. Your comment made me see that wasn't actually the point!

You transaction includes multiple entries. Your schema is multi-entry by definition :P

It's funny how many commentators here confuse debit/credit with double-entry.

SQL or not doesn't really matter. Double entry provides redundancy, and redundancy allows for auditing. It's like checksumming to detect errors.

I guess I'm asking: what would go wrong with the design in my comment, which makes transactions first-class and has a single entry for each one? What errors would double entry prevent that this doesn't?

Startups could improve in 2 wasy:

Perhaps popular culture should review the use of nosql databases, and then spending tremendous effort to try and make it into a relational database, while nosql databases can be setup to be eventually consistent.

Money, and accurate accounting don't work too well when the inputs to a calculation of a balance are eventually consistent.

So much work is avoided at times it seems to not learn SQL that it can rival or end up being more work once NOSQL leads you down the path of inevitable relational needs.

In addition to this, maybe it's time to build honeypot ledgers with a prize and reward in it for anyone who can hack or undermine it, similar to vulnerability bounties. Run for long enough, it would determine at least that one side of security mistakes in startups being reduced. Run a sprint, with prizes and let them run like daily deal fantasy and watch things harden.

Losing cents is because somebody didn't use Decimal for currency. And that's just flat out malfeasance--possibly even regulatory malfeasance.

Double entry is irrelevant, here.

I love the "use crypto" to fix the problem suggestion. LOL!

It's kind of poignant since crypto has given me a fantastic club to beat executives over the head with whenever they want to do stupid handling of money. Since crypto has so many decimal places, you break dumbass floating point handling of money immediately and irretrievably for way more than just "a couple cents". Your CFO starts jumping up and down really excitedly about handling money properly when he has to worry about a rounding error forcing him to compensate a Bitcoin price jump/crash.

It could be rounding issues as well (roundings do not all cancel out), or having to divide an even amount unevenly (how do you pay $10 in three "equal" installments?) Even with infinite decimal precision there will be special handling involved.

The payments industry is rife with nonsense like this.

At a fintech startup I was working with, we built a "shadow ledger" because we couldn't trust a 3rd party PP to give us the correct information, which would otherwise allow double spending on accounts.

We tried 3 different (major!) PPs - they ALL had similar flaws.

It's just like it was with pen and paper... Everyone trusts what is written on the financial statement but when they come to withdraw the cash, all at once, suddenly everyone finds out that the gold isn't there.

It's the same with computer systems. The charts show something, but until enough people decide to all withdraw their money or sell their stock at the same time, nobody has any idea that the money or asset simply isn't there or nobody knows just how frothy the valuation is.

Social media and search algorithms are highly optimized to ensure that people don't sell or withdraw stuff at the same time. Modern media directs massive attention towards certain topics as a way to draw attention away from other topics which could collapse the economy.

Also, imagine a bank has a serious bug which causes millions or billions of dollars to disappear every year or creates extra illegitimate dollars. Imagine they only discover this bug after a few years of operation... How likely is it that they will report it to an authority? They didn't notice it for years, why not pretend they didn't notice it for a few MORE years... The incentive to delay the reckoning is an extremely powerful one.

Precisely. Designing a transactional system can be solved. Designed a transactional system that properly entangles the bits with the assets they represent is the hard part.

Hey, great works

I literally build the same system and ask for HN Ask but no one answer it and already using double entry and ledger based on some research and AI advice lol

I implement all the check its mention in your page and quite satisfied using postgress constraint and trigger function to detect abnormality before bad row is inserted

but the problem now is populate the database using fake data because how rigirous my constraint and rule are, I need to simulate real world use case and duplicate it on development machine but this are hard

if you ever read this, can you give some advice please, because simply looping that function is not gonna work because constraint of double entry and ledger

A related pet peeve of mine is UIs that round off your investment holdings or cash.

Sometimes to 2 digits. Sometimes 3. Or 4. How many significant digits are there even in fractional SPY holdings? You just resort to looking in all the places (statements, transaction history, overview dashboard, …) and going with the one that shows the most digits, and assume that's all of them.

Big and small companies do this. And when I've reported it, they don't care. They don't see the problem.

Building fintech platforms are not easy, specially for large systems which not only require robustness, scale, but also traceability, daily reconciliations. We even had features to undo and redo (correct and replay) transactions from a time in history.

I did not like finance much, but building a fintech system really teaches a lot, not only from technology perspective but from managing stakeholders, processes, compliance, dealing with all kinds of finance-specific issues.

One should always work in fintech, at some point of time in career.

Once upon a time I worked for a company that made accounting software. While there I had to explain to the other devs:

1) That being able to escape characters in CSV files is kind of important. (Some customers noticed that a single quote in a transaction description would make the importer silently fail to import the rest of the transactions... :S ), and

2) Why it wasn't good to store all dollar values as doubles. I'm reasonably sure I quoted Superman 3 at some point during this discussion.

One of my $previousjob's managed to invent their own kind-of sort-of double entry accounting system, but not quite. So whenever anything went sideways (and we interacted with Stripe, so something always went sideways), it required a lot of manual intervention with changing of amounts in various tables.

I pushed for offsetting entries instead, but was overruled. It was a _nightmare_.

Thankfully, it is not $previousjob.

If the author is reading this: Please stop hiding the scroll bar. It's user hostile.

I want to see if I have time to finish this article before I have to leave. And now I have to waste time copy-pasting the contents into a text file, just to see where I am.

Edit: Actually, it's invisible only in Firefox. Chrome shows the scrollbar still. So this may be a bug in the author's CSS or something.

I think “make it work, make it right, make it fast” is a solid approach.

It appears the startup in question just never even did step one. A financial system that is imprecise does not work.

Additionally that mantra is typically applied at the story/issue level in my experience as in:

1. Get the tests passing 2. Get the code clean 3. Identify scaling bottlenecks and remedy if necessary.

At $currentJob, we have decided to stick with Fineract for ledger maintenance and that helped us scale without having to deal with the nitty gritty details of accounting. If you don’t have the necessary expertise around you, offload that responsibility to a third party system that is maintained by those who do.

I don't understand what the difference in modeling between:

Entry(account, direction, non-negative amount), direction is debit or credit.

Entry(account, signed amount), + is debit, - is credit (for example).

It's a two way mapping and should be equivalent. Unless debit or credit amount could be negative. But as I understand it's a big NO-NO in accounting.

I guess the lesson here is to not bank with an a16z company. I have been using Wise for years. Very very happy.

> It’s just that…it goes without saying that fintech companies should know better

haha, if only.

move fast, break everything if it gives profit

The number of engineers I've met that realized only too late that, by definition, floats and doubles are absolutely *not good* to deal with money is unbelievable. At this point it should be universal knowledge, but it isn't.

nit: double entry accounting only goes back to the 13th century, not thousands of years.

Fibonacci, if I remember correctly.

You do not remember correctly: https://en.wikipedia.org/wiki/Double-entry_bookkeeping#Histo...

It took me a while to realise what double-entry bookkeeping actually was, then one day it hit me: it's literally just the flow of money through a system. Understand that, and you understand the cornerstone of accountancy.

if only accountancy 'learned' from OOP and just had one mutable balance for each account :)

Blockchains are examples worth learning from about how to build ledgers. Append-only transactions which start as pending and then eventually confirmed. Balances are cumsums of confirmed transactions.

> Not losing track of money is the bare minimum for fintech companies.

The last fintech I worked for had a joke about how you weren't really an employee until your code had lost some money.

Feeling vindicated for the double entry transaction system we built at clearvoice.com for our two-sided marketplace, leveraging the fantastic DoubleEntry Ruby Gem from Envato.

As an SWE, I've seen how the sausage is made and is the reason why I'm very very careful when using fintechs for anything.

This was many years ago. I has an injurie that prevented me from driving, so I asked if there was any project on the backburner I could do on my own in 3 weeks without too much interaction (this was long before easy remote work infra).

Sure there was. A financial services company wanted to replace their repayment plan generator that ran on an aging AS/400 with someting running in .Net on Windows.

I dug in an learned all about time value of money, numerical formats and precision, rounding strategies, day counting strategies for incomplete periods (did you know there are many dozens of those) etc.

I made everyting as configurable as I could so we had the option to offer to other financial service clients by just setting up a different profile.

Since I had no interaction with the client before the first presentation meeting, I put in what to me felt like the most plausible config (I had managed projects in the domain before so I was not completely clueless to the mindset).

We had the meeting. I showed a few calculated plans which they compared on the spot to the AS/400 output. I deployed on a test VM for them so they could do more extensive testing. Code was accepted with 0 change requests and put into production shortly thereafter. Don't think they ever changed from the default settings.

An incredibly depressing thing in our profession is how we collectively lack memory.

We routinely re-discover stuff that probably was already solved by a quiet lady writing a CICS transaction in a s-360 system in 1969.

Yeah, dealing with money and especially others people money without double entry bookkeeping is a bad practice.

But this particular problem is the consequence of the choice of using floating point binary math to deal with monetary quantities.

Given the fact that most modern languages have very awkward support for arbitrary precision decimal numbers the most sensible way to deal with money usually boils down to store it as an integer.

I don't see how these are related (except of course for the errors in the second that could be caught by the first), and hopefully nobody in fintech is so ignorant/dumb to use floating point for money ?

Or are we talking about random scammers as 'fintech' now rather than banks (the fintech of which might indeed be old enough to be still running on COBOL-like systems) ?

Hot take: I blame the culture of obsession with CS topics, leetcode, tail recursion, dependent types etc and not enough focus on solid engineering and domain knowledge in areas encountered by 90% of real life jobs.

Starts in the education and perpetrates via hiring, blogosphere and programmer celebrities.

Wait, why can’t I just use a Postgres database with the Decimal type and call it a day?

You need to keep state, so that when something goes wrong you can figure out that something went wrong.

PG's rounding can be summed up as: we round to the nearest digit you specify, but how we get there is mostly undefined.

Every bank and organization you do business with(especially if you cross-jurisdictions) will likely round pennies differently. PG and the decimal data type(which is actually the numeric data type) are only part of the solution, as they won't handle rounding pennies correctly for you, but they will store it, if you round properly yourself.

PG also has a money data type, but it also doesn't let you specify the rounding rules, so you have to be careful. Also, money is tied to lc_monetary, which is handled outside of PG's control, so you have to be very careful with it, when moving data across PG instances, or you might be surprised. Also, it doesn't let you specify which currency this money is, which means you have to store that yourself, if you ever have more than one currency to care about.

If you don't care about balancing to the penny with your external organizations(banks, CC processors, etc), you are 100% guaranteed to have a bad time eventually.

Interesting article, thanks for sharing! I believe the things it explains are very important, and not only in fintech. A few years ago, even in the case of trying to keep track of a small organisation spendings and incomes in a spreadsheet I had some of the problems that are mentioned in the article and had to completely rethink and remake the spreadsheet in order to make it actually work.

But apart from the technical points, another interesting thing in the article is its introduction and what it says there about money as a concept. Yes, money is debt, both philosophically and technically it is the right way to think about it. And I believe that's something the crypto-assets industry and enthusiasts as a whole fundamentally gets wrong (mostly because of the libertarian political point of view blockchain tech has been designed with).

I'm not sold on double entry here.

If a new transaction enters the system now, I could follow the advice and record it as two sources of truth. Or I could just record it once.

If I could turn a single transaction into two ledger entries today, I could do the same later if I needed to.

the systems I’ve worked on that use your design end up in a state I call “the zoo” - because you never actually have simple transactions like you’re imagining. You end up with one object for credit card payments with stripe fees, another one for purchases with taxes, another one for ACH payouts, all meaning different things

double entry is actually the simplest projection of any possible billing system because anything that moves money can be described as a bundle of credits and debits that happen simultaneously and sum to zero. so you get accounts for taxes collected and fees charged automatically and don’t have to write custom logic for each thing

You are thinking it's duplicating work, but it isn't. You are simply recording the whole transaction in a double entry system. i.e. you know where the money came from and where it went. That's all a double entry system does, records both sides of the transaction.

This means when something goes wrong, you have a chance to figure out what in the world happened.

That's it. If you want to think of it as a state machine, you are storing all the state that happened with the resource you are tracking(usually money). This way when some part of the resource is missing, you can go figure out what happened.

Think about it another way: always record enough of the transaction, that if you were required to sit in a courtroom in front of a jury and explain what happened, you could do so. If you can't do that, you didn't store enough information.

There's no redundancy. Imagine you have 2 buckets, you reach into bucket A and grab some cash and put it in bucket B. $x dollars leaving bucket A is entry part 1, $x dollars entering bucket B is entry part 2. That's all it is. I honestly don't understand what single entry is by comparison - it sounds like losing an essential attribute of the transaction.

"Two sources of truth" is an anti pattern in computing and isn't redundancy. Use replication or backups for that.

You can't imagine simpler than what you described because it's single entry.

Double-entry is twice as complicated, makes sense only to accountants and not to computing people. Your example of 1 transaction would be doubly-kept as some nonsense like

  BucketA revenue:$x CR cash:$x CR
  BucketB revenue:$x DB cash:$x DB

Single would be recording the two accounts, one being the source account and the other the destination, and a quantity.

Keeping entries separate makes it easier to deal with data for specific accounts, the transaction will contain both of them.

Learning and using the ledger-cli accounting tool taught me a lot about this. It's incredible how messy seemingly simple things can be and how much trouble a bunch of cents can cause. Seems to be the accounting version of off-by-one errors. It is very tempting to just write them off as losses and forget about them forever.

Rounding in particular is a truly endless source of trouble and has caused me to chase after a lot of cents. Dividing up a large payment into multiple installments is the major cause of rounding in my use case. Life starts to suck the second things fail to be evenly divisible. I created an account to track gains and losses due to rounding and over time and it's adding up to quite the chunk of change.

Hilariously, the payment systems would charge me incorrect rounded up amounts and then they would refund the difference to my credit card at some undefined time in the future. Tracking and correlating all these seemingly random one or two cent transactions has got to be one of the most annoying activities I've ever learned to put up with. Not only do I have to figure out why things aren't quite adding up, I have to patch things up in the future when they fix the mistake.

Why can't these things just charge the correct amounts? For example, imagine splitting up $55.53 into four installments. That's 4x$13.8825. They could charge me 3x$13.88 + 1x$14.89. Instead they round it up to $55.56 and charge me 4x$13.89, then maybe they refund me $0.03 some unknown day in the future. It's like the systems go out of their way to be as annoying as possible. Some systems do this silly single cent refund dance in my credit card statement even though they print the exact same 3xN + 1xN+0.01 solution in the receipt. Makes absolutely no sense to me.

It's getting to the point I'm trying to avoid this nonsense by structuring purchases so the final price is evenly divisible by some common factors. I never liked the .99 cents manipulation trick but I seriously hate it now.

Im sure there is a point in the article but ive never seen a dancing cent nor can i imagine one. My numbers are all strings "5" is "5" forever. If one somehow ends up storing it as 4.99 why would the other entry be correct?

Your second sentence tells us your first sentence was a lie. You can clearly imagine one which is why you specified which data type you use for money. You know a floating point issue is an issue.

Now let's say your price is "0.00023" per unit and someone uses "213.34" units. Can you imagine it now?

Question was about double entry. Does recording it twice solve floating point issues?

Yes -- or rather it is designed to identify the issue that floating point causes.

The main idea behind double ledger accounting is that if you add a long list of the same numbers twice (as totally independent operations), then if you have added both lists correctly, you the two independent results will be the same. If you made at least one mistake in adding either or both of the lists, then it is possible, but unlikely that the results will match.

It's easy to think that computers don't make mistakes like humans when adding numbers, however floating point addition for long sequences of numbers is non deterministic (and error prone if you don't sort the numbers first and start with the small ones).

While double ledger systems won't fix this problem, they will identify if you have a problem with your addition (particularly a non-deterministic one like you would find with floating point addition) when you go to reconcile your books and find that the numbers across the various accounts don't add up.

> Does recording it twice

Double entry is (confusingly) not about recording it twice, it's about using a transaction model where the state of N accounts has to be changed in compensating directions within a transaction for it to be valid, N being >= 2.

So depending on how your transaction schema is defined, a double-entry transaction can be written without ever even repeating the amount, e.g.

    {"debit": "cash:main", "credit": "giftcards:1234", "amount": 100}

Making it effectively impossible to represent invalid state change. Things get trickier when N > 2 as classical double-entry tend to _not_ directly relate tuples of accounts directly to one-another and instead relying on a aggregated balancing of the changes to the N accounts at the transaction level, though YMMV between different ledgering systems.

There was no question, that I was answering. Simple someone claiming they have no idea how it would even be possible for a $5 end up being $4.98, while literally stating they know about floating point issues.

This is more a problem with ill-specified contracts though. It was a constant source of annoyance when I was sole-trading for a bit, because what happened was something like this:

I'd be quoted a day-rate. That was what I was actually going to get paid, one day. But then I'd be told to bill it as an hourly rate. And then actually to bill it as 7.5 hours.

But I wasn't told what the hourly was - the hourly was whatever my day rate was, divided by 7.5. So this led to the problem that it produced an irrational number as a result.

Technically this should've been fine...except no one I dealt with knew or cared about this concept - they all used Excel. So if I rounded the irrational to nearest upper cent (since that's the smallest unit which could be paid) they complained it didn't add up. If I added a "correction item" to track summing up partial cents, they complained it wasn't part of the hourly.

In the end I just send python decimal.Decimal to maximum precision, flowed through invoices with like 8 digits of precision on the hourly rate, and this seemed to make Excel happy enough. Of course it was completely useless for tracking purposes - i.e. no one would ever be able to pay or unpay 0.666666666667 cents.

Because what's not in employment contracts that really should be? Any discussion on how numbers are to be rounded in the event of uneven division. You just get to sort of guess what accounting may or may not be doing. In my case of course it didn't matter - no one was ever going to hold me to anything other the day rate, just for some reason they wanted to input tiny fractions of a cent which they actually couldn't track.

And it's not an idle problem either: i.e. in the case of rounding when it comes to wages, should it be against the employee? It's fractions of a cent in practice, but we're not going to define it at all?

The example of the price and units is actually a real-world example. If you look at how much you pay for electricity you'll see you're paying something like 0.321 per KwH and you're not billed in full units.

Your issue is just people being lazy and forcing a day rate into an hourly employment system.

> the hourly was whatever my day rate was, divided by 7.5. So this led to the problem that it produced an irrational number as a result.

The only way for this to be true is if your day rate was irrational to begin with.

I mean i want to understand not that the problem doesnt exist.

If the price is "0.00023" per unit and someone uses "213.34" units I feed those strings into a multiplication function that returns the correct string 100% of the time.

That much i understand. I dont get how that category of problems is addressed by the solution described.

What i also understand is that you inevtably get to deal with accountants or other finance people. Working in a format they understand is not optional. They will have you transform whatever you have into that anyway.

I learn not to wonder why but maybe i should.

> If the price is "0.00023" per unit and someone uses "213.34" units I feed those strings into a multiplication function that returns the correct string 100% of the time.

But you're not coming up with a valid monetary amount.

Seems like your definition of units is wrong. How can I sell .34 of an HVAC unit?

Btw for those wondering what to do instead is to simply use https://tigerbeetle.com/

I found this paragraph

> Prerequisites: TigerBeetle makes use of certain fairly new technologies, such as io_uring or advanced CPU instructions for cryptography. As such, it requires a fairly modern kernel (≥ 5.6) and CPU. While at the moment only Linux is supported for production deployments, TigerBeetle also works on Windows and MacOS.

wha— what? why?? they must be solving some kind of scaling problem that I have never seen

Thank you for posting this.

I’m trying to get my head around how to build a fairly complex ledger system (for managing the cost schedules in large apartment buildings where everyone might pay a different proportion and groups of apartments contribute towards differing collections of costs) and you’ve just massively accelerated my thinking. And possibly given me an immediate solution.

Have you used tigerbeetle in production?

I mean, tigerbeetle looks extremely cool (I've watched the project develop since its inception), and I trust them to be rock-solid. But saying "just use this project that is very new and unproven, written in a new and unproven programming language" is just pretty unserious. At least talk about pros, cons, risks, and tradeoffs.

>very new and unproven, written in a new and unproven programming language

while i'd generally agree with this, in the case of TigerBeetle i think it's safe to trust their tech based on their test-suite/practices - one of the most impressive things I've seen in the past 5 years.

they extend the testing practices of FoundationDB (now spun out into Antithesis[0]), going a layer deeper to test for data integrity when there is disk io corruption. check out from ~20:30 in this demo:

https://youtu.be/sC1B3d9C_sI?feature=shared&t=1228/

[0] https://antithesis.com

Fintech is safe heaven for scam.

> A double-entry system is an accounting method that tracks money at both its source and destination.

Nope. Double entry bookkeeping means every transaction is recorded in (at least) two accounts.

To illustrate the difference (hopefully I get this right): A transaction needs to balance, left vs right.

For example, you receive a sales order for 1000€ of widgets, which cost you 600€. You ship them. You invoice for 1000€.

No money moved (hopefully you do get paid at some point though). However, you need to do some bookkeeping.

On the left side of the ledger (debits): Accounts receivable 1000€. Cost of goods sold 600€.

On the right side of the ledger (credits): Revenue goes up by 1000€. Inventory goes down by 600€.

These completely match. No money has moved, but the books are now up to date and balance.

Any transaction that does not balance should be rejected.

Double-entry accounting is like strongly typed programming languages.

Yeah it's a real pain to get started because you need to understand the core concepts first then fight to balance the transactions ("compile errors") before you have anything useful. And when your results are wrong, you know that at least it's not because of the basic stuff.

SV Tech Bros roasting themselves on HN.

Every engineer in London is laughing at this right now.

Wow double entry! Ledgers are hard! Wow. So true.

> And yet, I used to work for a startup that, on every transaction, simply lost track of a couple of cents. As if they fell from our pockets every time we pulled out our wallets.

is this not the literal plot of Office Space? did you check for a Michael Bolton employee?

Great post!

The blockchain industry is basically founded on incompetence. If you go back to early Silk Road and Mtgox you can see some shocking things. Like the founders were posting questions on Stackoverflow about database basics. They then ended up building systems with race conditions (withdrawal code for both systems.)

See, if you're an engineer (from a web background) you might think that using a regular backend language with a relational DB would be fine. But this service is typically optimized for concurrency. That might sound amazing. But what you really want is a simple queue. This ensures everything is applied in order and makes it impossible to accidentally have transactions execute on stale state. Plus, a queue can be executed amazingly fast -- we're talking Nasdaq scales. If you use a standard DB you'll end up doing many horrible broken hacks to simulate what would be trivial engineering had you used the right design from the start.

You've got other things to worry about, too. Financial code needs to use integer math for everything. Floating point and 'decimals' lead to unexpected results. The biggest complexity with running large-scale blockchain services is having to protect 'hot wallets.' Whenever an exchange or payment system is hacked the target is always the funds that sit on the server. When the industry began there were still many ways to protect the security of these hot wallets. The trouble is: it required effort to implement and these protocols weren't widely known. So you would get drive-by exchanges that handled millions of dollars with private keys sitting on servers ready to be stolen...

Today there are many improvements for security. New cryptographic constructs like threshold ECDSA, hardware wallets, hardware key management (like enclaves), smart contract systems (decentralized secret sharing... and even multi-sig can go a long way), and designs that are made to be decentralized that remove the need for a centralized deposit system (still needs some level of centralization when trading to a stable coin but its better than nothing.)

I would say the era where people 'build' ledgers though is kind of over. To me it appears that we're organizing around a super-node structure where all the large 'apps' handle their own changes off-chain (or alternatively based on regular trust.) The bottom layer will still support payments but it will be used less often. With more transaction activity happening on layers above. I think its still important to scale the chain and make it as secure as possible. Bitcoin has a unique focus here on security above everything else. Making it ideal for long-term hedges. For every day stuff I can't think of any chains that have more credible R & D than Ethereum. They seem to even have been doing research on the P2P layer now which traditionally no one has cared about.

> Financial code needs to use integer math for everything. Floating point and 'decimals' lead to unexpected results.

Decimals are integers. There's no difference between integer math and decimal math, only in the meaning of the bit pattern afterwards.

> Decimals are integers

No, most decimal types are decimal floating-point. (So, sure, technically redundant, but a lot of people say "floating point" to mean IEEE 754 specifically, even if that's technically incorrect).

The whole "move fast and do a shit job" is laughably stupid. It's pure incompetence and egotism and a product of excess salary in the market.

Doing it the right way doesn't take any longer than doing it the shitty way. Successful startups focus on extreme minimalism, not focus on doing the worst possible job.

Ok, but please don't fulminate on HN. This is in the site guidelines: "Please don't fulminate."

https://news.ycombinator.com/newsguidelines.html.

Except if you don't actually know what to do, or even what concepts to apply. This is where I see many start ups fall down.

OK but apparently they did get to make the startup mistakes? They built the quick thing that worked well enough, got some customer traction, and then when they had bugs they were able to rework it and continue.

Frankly I'm not even convinced that double-entry is the sole right answer in this space. There are things you need to be able to represent, but the original reasons for doing double-entry (making errors when subtracting figures) no longer apply.

(I've worked at investment banks and fintech startups)

Hmm, can you elaborate on the original reasons for double entry and them not applying any.kre? I'm not in that space and double entry always seemed an extremely weird, arbitrary requirement / paradigm. Thanks!

The main point of double-entry account keeping is the notion that money never vanishes, it's always tracked as going somewhere.

I think this tends to get misrepresented by people trying to literally keep two entries like we were working with pen and paper book-keeping though.

Because if I have a simple list of transactions, I can easily use a SQL query to recreate a view of double-entry book-keeping. It's perfectly fine to just record the amount of money, and two entity names in a table.

Coz the whole point of the system is that you can explain where the money went at any given time, referenced against something which should be independently auditable: i.e. a vendor receipt which in turn says "there's a physical item of X we gave you".

The "double entry" system's actual merit is when there's multiple people keeping the books. i.e. if you're in a business and your department is spending money on goods, then when you're doing that those transactions should be getting reflected in the shipping departments books (owned by a different person) and maybe your accounting department or the like. The point is that there have to be actual independent people involved since it makes fraud and mistakes hard - if you say you bought a bunch of stuff, but no one else in the business received it or the money to pay for it, then hey, you now need to explain where the money actually is (i.e. not in your pocket).

Cool post, wish it existed 2 years ago when we started building Pave Bank, or 10 years ago when we started building Monzo.

If you're starting a bank or need a ledger these days (and aren't using a core banking provider that has one), then i usually recommend Tiger Beetle.

Reply on news.ycombinator.com