There are many systems that take a native data structure in your favorite language and, using some sort of reflection, makes an on-disk structure that resembles it. Python pickles and Java’s serialization system are infamous examples, and rkyv is a less alarming one.
I am quite strongly of the opinion that one should essentially never use these for anything that needs to work well at any scale. If you need an industrial strength on-disk format, start with a tool for defining on-disk formats, and map back to your language. This gives you far better safety, portability across languages, and often performance as well.
Depending on your needs, the right tool might be Parquet or Arrow or protobuf or Cap’n Proto or even JSON or XML or ASN.1. Note that there are zero programming languages in that list. The right choice is probably not C structs or pickles or some other language’s idea of pickles or even a really cool library that makes Rust do this.
(OMG I just discovered rkyv_dyn. boggle. Did someone really attempt to reproduce the security catastrophe that is Java deserialization in Rust? Hint: Java is also memory-safe, and that has not saved users of Java deserialization from all the extremely high severity security holes that have shown up over the years. You can shoot yourself in the foot just fine when you point a cannon at your foot, even if the cannon has no undefined behavior.)
I have zero doubt that you’re on some ‘no true Scotsman’-style “you’re not doing Real Development if you are using these technologies to solve these problems” thing. Let’s just drop that. There are myriad ‘real man webscale development’ scenarios where these are more than acceptable.
Actually, it's you who is giving that impression with an ultra vague "doesn't solve the problems described".
The only problem in the blog post is efficient coding of optional fields and all they was introduce a bitmap. From that perspective, JSON and XML solve the optional fields problem to perfection, since an absent field costs exactly nothing.
> Protobufs definitely doesn’t solve the problems described. Capnproto may solve it but I’m not 100% sure. JSON/XML/ASN.1 definitely don’t.
I'm not sure you are serious. What open problem do you have in mind? Support for persisting and deserializing optional fields? Mapping across data types? I mean, some JSON deserializers support deserializing sparse objects even to dictionaries. In .NET you can even deserialize random JSON objects to a dynamic type.
Can you be a little more specific about your assertion?
> Depending on your needs, the right tool might be Parquet or Arrow or protobuf or Cap’n Proto
I think parquet and arrow are great formats, but ultimately they have to solve a similar problem that rkyv solves: for any given type that they support, what does the bit pattern look like in serialized form and in deserialized form (and how do I convert between the two).
However, it is useful to point out that parquet/arrow on top of that solve many more problems needed to store data 'at scale' than rkyv (which is just a serialization framework after all): well defined data and file format, backward compatibility, bloom filters, run length encoding, compression, indexes, interoperability between languages, etc. etc.
But if you use complicated serialisation formats you can't mmap a file into memory and use it directly. Which is quite convenient if you don't want to parse the whole file and allocate it to memory because it's too large compared to the amount of memory or time you have.
Fully agreed. rkyv looks like something that is hyper optimizing for a very niche case, but doesn't actually admit that it is doing so. The use case here is transient data akin to swapping in-memory data to disk.
"However, while the former have external schemas and heavily restricted data types, rkyv allows all serialized types to be defined in code and can serialize a wide variety of types that the others cannot."
At a first glance, it might sound like rkyv is better, after all, it has less restrictions and external schemas are annoying, but it doesn't actually solve the schema issue by having a self describing format like JSON or CBOR. You won't be able to use the data outside of Rust and you're probably tied to a specific Rust version.
Sure, if your structure doesn't contain any pointers and you only ever want to support one endianness and you trust your compiler to fix the machine layout of the struct forever.
> Sometimes the best optimization is not a clever algorithm. Sometimes it is just changing the shape of the data.
This is basically Rob Pike's Rule 5: If you've chosen the right data structures and organized things well, the algorithms will almost always be self-evident.(https://users.ece.utexas.edu/~adnan/pike.html)
I wouldn't give too much credit to rules like this. Data structures are often created with an approach in mind. You can't design a data structure without knowing how you will use it.
If anything it's the other way round, if you're not talking about business domain modeling (where data structures first is a valid approach).
> If anything it's the other way round, if you're not talking about business domain modeling (where data structures first is a valid approach).
And even there, the data models usually come about to make specific business processes easier (or even possible). An Order Summary is structured a specific way to allow both the Fulfilment and Invoicing processes possible, which feed down into Payment and Collections processes (and related artefacts).
To elaborate on @jeswin's point above (IDK why it got downvoted)... a data structure is basically like a cache for the processing algorithm. The business logic and algorithm needs will dictate what details can be computed on-the-fly -vs- pre-generated and stored (be it RAM or disk). Eg: if you're going to be searching a lot then it makes sense to augment the database with some kind of "index" for fast lookup. Or if you are repeatedly going to be pllotting some derived quantity then maybe it makes sense to derive that once and store with the struct.
It's not enough for a data structure to represent the "fundamental" degrees of freedom needed to model the situation; the algorithmic needs (vis-a-vis the available resources) most definitely matter a lot.
I'm saying that if you care about performance, data structures should be designed with approach specific tradeoffs in mind. And like I've said above, in typical business apps, it's ok to start with data structures because (a) performance is usually not a problem, (b) staying close to the domain is cleaner.
You said: "You can't design a data structure without knowing how you will use it."
But the whole discussion involves knowing how you will use it; the advocacy is for careful consideration of data structures (based on how you will use them) resulting in less pain when designing/choosing algorithms.
"Show me your flowcharts and conceal your tables, and I shall continue to be mystified. Show me your tables, and I won’t usually need your flowcharts; they’ll be obvious."
I (deep, deep in embedded systems) have seen this too often, that code is incredibly complex and impossible to reason around because it needs to reach into some data structure multiple times from different angles to answer what should be rather simple questions about next step to take.
Fix that structure, and the code simplifies automagically.
> But SQL schemas often look like this. Columns are nullable by default, and wide tables are common.
Hard disagree. That database table was a waving red flag. I don't know enough/any rust so don't really understand the rest of the article but I have never in my life worked with a database table that had 700 columns. Or even 100.
I saw tables with more than a thousand columns. It was a law firm home-grown FileMaker tool. Didn't inspect it too closely, so don't know what was inside
I remember a phrase from one of C. J. Date's books: every record is a logical statement. It really stood out for me and I keep returning to it. Such an understanding implies a rather small number of fields or the logical complexity will go through the roof.
The blog post is an entertaining read, but I was left with the impression the author might have tried do embellish, particularly in it's disbelief angle.
Take this passage:
> The app relied on a SOAP service, not to do any servicey things. No, the service was a pure function. It was the client that did all the side effects. In that client, I discovered a massive class hierarchy. 120 classes each with various methods, inheritance going 10 levels deep. The only problem? ALL THE METHODS WERE EMPTY. I do not exaggerate here. Not mostly empty. Empty.
> That one stumped me for a while. Eventually, I learned this was in service of building a structure he could then use reflection on. That reflection would let him create a pipe-delimited string (whose structure was completely database-driven, but entirely static) that he would send over a socket.
Classes with empty methods? Used reflection to create a pipe-delimited string? The string was sent over the wire?
Why congratulations, you just rediscovered data transfer objects, specifically API models.
As to your hard disagree, I guess it depends... While this particular user is on the higher end (in terms of columns), it's not our only user where column counts are huge. We see tables with 100+ columns on a fairly regular basis especially when dealing with larger enterprises.
Can you clarify which knowledge domains those enterprises fall under with examples of what problems they were trying to solve?
If it's not obvious, I agree with the hard disagree. Every time I see a table with that many columns, I have a hard time believing there isn't some normalization possible.
Schemas that stubbornly stick to high-level concepts and refuse to dig into the subfeatures of the data are often seen from inexperienced devs or dysfunctional/disorganized places too inflexible to care much. This isn't really negotiable. There will be issues with such a schema if it's meant to scale up or be migrated or maintained long term.
Normalization is possible but not practical in a lot of cases: nearly every “legacy” database I’ve seen has at least one table that just accumulates columns because that was the quickest way to ship something.
Also, normalization solves a problem that’s present in OLTP applications: OLAP/Big Data applications generally have problems that are solved by denormalization.
We have many large enterprises from wildly different domains use feldera and from what I can tell there is no correlation between the domain and the amount of columns.
As fiddlerwoaroof says, it seems to be more a function of how mature/big the company is and how much time it had to 'accumulate things' in their data model.
And there might be very good reasons to design things the way they did, it's very hard to question it without being a domain expert in their field, I wouldn't dare :).
> I can tell there is no correlation between the domain and the amount of columns.
This is unbelievable. In purely architectural terms that would require your database design to be an amorphous big ball of everything, with no discernible design or modelling involved. This is completely unrealistic. Are queries done at random?
In practical terms, your assertion is irrelevant. Look at the sparse columns. Figure out those with sparse rows.
Then move half of the columns to a new table and keep the other half in the original table. Congratulations, you just cut down your column count by half, and sped up your queries.
Even better: discover how your data is being used. Look at queries and check what fields are used in each case. Odds are, that's your table right there.
Let's face it. There is absolutely no technical or architectural reason to reach this point. This problem is really not about structs.
Feldera speak from lived experience when they say 100+ column tables are common in their customer base. They speak from lived experience when they say there's no correlation in their customer base.
Feldera provides a service. They did not design these schemas. Their customers did, and probably over such long time periods that those schemas cannot be referred to as designed anymore -- they just happened.
IIUC Feldera works in OLAP primarily, where I have no trouble believing these schemas are common. At my $JOB they are, because it works well for the type of data we process. Some OLAP DBs might not even support JOINs.
Feldera folks are simply reporting on their experience, and people are saying they're... wrong?
I remember the first time I encountered this thing called TPC-H back when I was a student. I thought "wow surely SQL can't get more complicated than that".
Turns out I was very wrong about that. So it's all about perspective.
> Normalization is possible but not practical in a lot of cases: nearly every “legacy” database I’ve seen has at least one table that just accumulates columns because that was the quickest way to ship something.
Strong disagree. I'll explain.
Your argument would support the idea of adding a few columns to a table to get to a short time to market. That's ok.
Your comment does not come close to justify why you would keep the columns in. Not the slightest.
Tables with many columns create all sorts of problems and inefficiencies. Over fetching is a problem all on itself. Even the code gets brittle, where each and every single tweak risks beijg a major regression.
Creating a new table is not hard. Add a foreign key, add the columns, do a standard parallel write migration. Done. How on earth is this not practical?
There are sometimes reasons this is harder in practice, for example let’s say the business or even third parties have access to this db directly and have hundreds of separate apps/services relying on this db (also an anti-pattern of course but not uncommon), that makes changing the db significantly harder.
Mistakes made early on and not corrected can snowball and lead to this kind of mess, which is very hard to back out of.
Fine, but you still need to read in those 100+ fields. So now you gotta contend with 20+ joins just to pull in one record. Not more practical than a single SELECT in my opinion.
You don't need to join what you don't actually need. You also need to be careful writing your queries, not just the schema. The most common ones should be wrapped in views or functions to avoid the problem of everyone rolling their own later.
Performance generally isn't an issue for an arbitrary number of joins as long as your indices are set up correctly.
If you really do need a bulk read like that I think you want json columns, or to just go all in with a nosql database. Even then, the above regarding indexing is still true.
I think you believe the average developer, especially on enterprise software where you see this sort of shit, is far more competent or ambitious than they actually are. Many would be horrified to see the number of monkeys banging out nasty DDL in Hibernate or whatever C# uses that have no idea what "normal forms" or "relational algebra" are and are actively resistant to even attempting to learn.
> it sounds like helping customers with databases full of red flags is their bread and butter
Yes that captures it well. Feldera is an incremental query engine. Loosely speaking: it computes answers to any of your SQL queries by doing work proportional to the incoming changes for your data (rather than the entire state of your database tables).
If you have queries that take hours to compute in a traditional database like Spark/PostgreSQL/Snowflake (because of their complexity, or data size) and you want to always have the most up-to-date answer for your queries, feldera will give you that answer 'instantly' whenever your data changes (after you've back-filled your existing dataset into it).
I’m working on migrating an IBM Maximo database from the late 90s to a SQL Server deployment on my current project. Also charged with updating the schema to a more maintainable and extensible design. Manufacturing and refurbishing domain - 200+ column tables is the norm. Very demoralizing.
Data from measurement tools. Everything about the tool configuration, time of measurement, operator ID, usually a bunch of electrical data (we make laser diodes) like current, potential, power, and a bunch of emission related data.
Sounds like a generic form of single table inheritance. I don't honestly see any other way to do it (punting to a JSON field is effectively the same thing) when you have potentially thousands of parts all with their own super specific relevant attributes.
I've worked on multiple products that have had a concept of "custom fields" who did it this way too.
If lots of columns are a red flag then red flags are quite common in many businesses. I’ve seen tables with tens of thousands of columns. Naturally those are not used by humans writing sql by hand, but there are many tools that have crazy data layouts and generate crazy sql to work with it.
> I have never in my life worked with a database table that had 700 columns
Main table at work is about 600, though I suspect only 300-400 are actively used these days. A lot come from name and address fields, we have about 10 sets of those in the main table, and around 14 fields per.
Back when this was created some 20+ years ago it was faster and easier to have it all in one row rather than to do 20+ joins.
We probably would segment it a bit more if we did it from scratch, but only some.
771 columns (and I've read the definitions for them all, plus about 50 more that have been retired). In the database, these are split across at least 3 tables (registry, patient, tumor). But when working with the records, it's common to use one joined table. Luckily, even that usually fits in RAM.
Some businesses are genuinely this complicated. Splitting those facts into additional tables isn't going to help very much unless it actually mirrors the shape of the business. If it doesn't align, you are forcing a lot of downstream joins for no good reason.
It is very common to find tables with 1000+ columns in machine learning training sets at e-commerce companies. The largest I have seen had over 10000 columns.
It's OLAP, it very common for analytical tables to be denormalized. As an example, each UserAction row can include every field from Device and User to maximize the speed at which fraud detection works. You might even want to store multiple Devices in a single row: current, common 1, 2 and 3.
Not everyone understands normal form, much less 3rd normal form. I’ve seen people do worse with excel files where they ran out of columns and had to link across spreadsheets.
That statement jumped out at me as well. I've worked as a DBA on tons of databases backing a wide variety of ERPs, web apps, analytics, data warehouses...700 columns?!? No.
You've never seen an SAP database where the business object had a couple hundred fields? Its pretty much required if you're touching international data.
I have seen tables (SQL and parquet, too) that have at least high hundreds of optional columns, but this was always understood to be a terrible hack, in those cases.
> Hard disagree. That database table was a waving red flag.
Exactly this.
This article is not about structs or Rust. This article is about poor design of the whole persistence layer. I mean, hundreds of columns? Almost all of them optional? This is the kind of design that gets candidates to junior engineer positions kicked off a hiring round.
Nobody gets fired for using a struct? If it's an organization that tolerates database tables with nearly 1k optional rows then that comes at no surprise.
Some folks pointed out that no one should design a SQL schema like this and I agree. We deal with large enterprise customers, and don't control the schemas that come our way. Trust me, we often ask customers if they have any leeway with changing their SQL and their hands are often tied. We're a query engine, so have to be able to ingest data from existing data sources (warehouse, lakehouse, kafka, etc.), so we have to be able to work with existing schemas.
So what then follows is a big part of the value we add: which is, take your hideous SQL schema and queries, warts and all, run it on Feldera, and you'll get fully incremental execution at low latency and low cost.
700 isn't even the worst number that's come our way. A hyperscale prospect asked about supporting 4000 column schemas. I don't know what's in that table either. :)
Here is an article I wrote this week with a section on Feldera - how it uses its incremental compute engine to compute "rolling aggregates" (the most important real-time feature for detecting changes in user behavior/pricing/anamalies).
Strictly speaking, Isn't there still a way to express at least one Illegal string in ArchivedString? I'm not sure how to hint to the Rust compiler which values are illegal, but if the inline length (at most 15 characers) is aliased to the pointer string length (assume little-endian), wouldnt {ptr: null, len: 16} and {inline_data: {0...}, len: 16} both technically be an illegal value?
I'm not saying this is better than your solution, just curious :^)
I feel like I'm missing something, but the article started by talking about SQL tables, and then in-memory representations, and then on-disk representation, but...isn't storing it on a disk already what a SQL database is doing? It sounds like data is being read from a disk into memory in one format and then written back to a disk (maybe a different one?) in another format, and the second format was not as efficient as the first. I'm not sure I understand why a third format was even introduced in the first place.
Feldera is an incremental query engine, you can think of it as a specialized database. If you have a set of question you can express in SQL it will ingest all your data and build many sophisticated indexes for it (these get stored on disk). Whenever new data arrives feldera can instantly update the answers to all your questions. This is mostly useful when the data is much larger than what fits in memory because then the questions will be especially expensive to answer with a regular (batch) database.
I would assume because then the shape of the data would be too different? SOAs is super effective when it suits the shape of the data. Here, the difference would be the difference between an OLTP and OLAP DB. And you wouldn't use an OLAP for an OLTP workload?
> This struct we saw earlier had 700+ of optional fields. In Rust you would never design a struct like this. You would pick a different layout long before reaching 700 Options. But SQL schemas often look like this.
Really? I've never had to do any serious db work in my career, but this is a surprise to me.
I wasn't sure about writing the article in the first place because of that, but I figured it may be interesting anyways because I was kind of happy with how simple it was to write this optimization when it was all done (when I started out with the task I wasn't sure if it would be hard because of how our code is structured, the libraries we use etc.). I originally posted this in the rust community, and it seems people enjoyed the post.
Just cus structs and classes work differently, and classes are much more common. I tend to make everything a class, unless there is a really good reason to make it a struct.
The most common programming language where "struct" and "class" are both kinds of user defined type is C++. In C++ they only "work differently" in that the default accessibility is different, a "struct" is the same as a "class" if you change the accessibility to public at the top with an access specifier.
If you think you saw a bigger difference you're wrong.
Classes are a safe default even if you expect things to go very, very fast.
The overhead of screwing up NUMA concerns vastly outstrips any kind of class vs struct differences. It's really one of the very last things you should be worrying about.
Allocating an array of a class vs an array of struct might seem like you're getting a wildly different memory arrangement, but from the perspective of space & time this distinction is mostly pointless. Where the information resides at any given moment is the most important thing (L1/L2/L3/DRAM/SSD/GPU/AWS). Its shape is largely irrelevant.
If I understand this problem was in rkyv, and solution is using rkyv with glue code. I hope they could integrate some sort of official derive macro `rkyv::Sparse` for this if it can't be done automatically in rkyv.
Why is rust allowed to reorder fields? If I know that fields are going to be generally accessed together, this prevents me from ordering them so they fit in cache lines.
You can choose in Rust to explain the representation you want for your data type. Unlike C or C++ that's not a non-portable vendor extension it's just part of the language, look at the repr documentation: https://doc.rust-lang.org/nomicon/other-reprs.html
So if you want "what C does" you can just repr(C) and that's what you get. For most people that's not a good trade unless they're doing FFI with a language that shares this representational choice.
There are many systems that take a native data structure in your favorite language and, using some sort of reflection, makes an on-disk structure that resembles it. Python pickles and Java’s serialization system are infamous examples, and rkyv is a less alarming one.
I am quite strongly of the opinion that one should essentially never use these for anything that needs to work well at any scale. If you need an industrial strength on-disk format, start with a tool for defining on-disk formats, and map back to your language. This gives you far better safety, portability across languages, and often performance as well.
Depending on your needs, the right tool might be Parquet or Arrow or protobuf or Cap’n Proto or even JSON or XML or ASN.1. Note that there are zero programming languages in that list. The right choice is probably not C structs or pickles or some other language’s idea of pickles or even a really cool library that makes Rust do this.
(OMG I just discovered rkyv_dyn. boggle. Did someone really attempt to reproduce the security catastrophe that is Java deserialization in Rust? Hint: Java is also memory-safe, and that has not saved users of Java deserialization from all the extremely high severity security holes that have shown up over the years. You can shoot yourself in the foot just fine when you point a cannon at your foot, even if the cannon has no undefined behavior.)
Protobufs definitely doesn’t solve the problems described. Capnproto may solve it but I’m not 100% sure. JSON/XML/ASN.1 definitely don’t.
It’s like you listed a bunch of serialization technologies without grokking the problem outlined in the post doesn’t have much to do with rkyv itself.
I have zero doubt that you’re on some ‘no true Scotsman’-style “you’re not doing Real Development if you are using these technologies to solve these problems” thing. Let’s just drop that. There are myriad ‘real man webscale development’ scenarios where these are more than acceptable.
Actually, it's you who is giving that impression with an ultra vague "doesn't solve the problems described".
The only problem in the blog post is efficient coding of optional fields and all they was introduce a bitmap. From that perspective, JSON and XML solve the optional fields problem to perfection, since an absent field costs exactly nothing.
> Protobufs definitely doesn’t solve the problems described. Capnproto may solve it but I’m not 100% sure. JSON/XML/ASN.1 definitely don’t.
I'm not sure you are serious. What open problem do you have in mind? Support for persisting and deserializing optional fields? Mapping across data types? I mean, some JSON deserializers support deserializing sparse objects even to dictionaries. In .NET you can even deserialize random JSON objects to a dynamic type.
Can you be a little more specific about your assertion?
> Depending on your needs, the right tool might be Parquet or Arrow or protobuf or Cap’n Proto
I think parquet and arrow are great formats, but ultimately they have to solve a similar problem that rkyv solves: for any given type that they support, what does the bit pattern look like in serialized form and in deserialized form (and how do I convert between the two).
However, it is useful to point out that parquet/arrow on top of that solve many more problems needed to store data 'at scale' than rkyv (which is just a serialization framework after all): well defined data and file format, backward compatibility, bloom filters, run length encoding, compression, indexes, interoperability between languages, etc. etc.
Don't forget PHP's serialize/unserialize, it's also sketchy. Looks like they at least put up a big warning in their docs: https://www.php.net/manual/en/function.unserialize.php
Not hating on PHP, to be clear. It has its warts, but it has served me well.
Dotnet used to have BinaryFormatter, which they had to kill for this reason, and they basically recommend that exact set of serializers as a replacement: https://learn.microsoft.com/en-us/dotnet/standard/serializat...
Delightful metaphor, I'll be looking everywhere for a chance to use that now!
But if you use complicated serialisation formats you can't mmap a file into memory and use it directly. Which is quite convenient if you don't want to parse the whole file and allocate it to memory because it's too large compared to the amount of memory or time you have.
Fully agreed. rkyv looks like something that is hyper optimizing for a very niche case, but doesn't actually admit that it is doing so. The use case here is transient data akin to swapping in-memory data to disk.
"However, while the former have external schemas and heavily restricted data types, rkyv allows all serialized types to be defined in code and can serialize a wide variety of types that the others cannot."
At a first glance, it might sound like rkyv is better, after all, it has less restrictions and external schemas are annoying, but it doesn't actually solve the schema issue by having a self describing format like JSON or CBOR. You won't be able to use the data outside of Rust and you're probably tied to a specific Rust version.
and often performance as well
BS. Nothing can be faster than a read()/write() (or even mmap()) into a struct, because everything else would need to do more work.
Sure, if your structure doesn't contain any pointers and you only ever want to support one endianness and you trust your compiler to fix the machine layout of the struct forever.
> Sometimes the best optimization is not a clever algorithm. Sometimes it is just changing the shape of the data.
This is basically Rob Pike's Rule 5: If you've chosen the right data structures and organized things well, the algorithms will almost always be self-evident.(https://users.ece.utexas.edu/~adnan/pike.html)
I wouldn't give too much credit to rules like this. Data structures are often created with an approach in mind. You can't design a data structure without knowing how you will use it.
If anything it's the other way round, if you're not talking about business domain modeling (where data structures first is a valid approach).
> If anything it's the other way round, if you're not talking about business domain modeling (where data structures first is a valid approach).
And even there, the data models usually come about to make specific business processes easier (or even possible). An Order Summary is structured a specific way to allow both the Fulfilment and Invoicing processes possible, which feed down into Payment and Collections processes (and related artefacts).
To elaborate on @jeswin's point above (IDK why it got downvoted)... a data structure is basically like a cache for the processing algorithm. The business logic and algorithm needs will dictate what details can be computed on-the-fly -vs- pre-generated and stored (be it RAM or disk). Eg: if you're going to be searching a lot then it makes sense to augment the database with some kind of "index" for fast lookup. Or if you are repeatedly going to be pllotting some derived quantity then maybe it makes sense to derive that once and store with the struct.
It's not enough for a data structure to represent the "fundamental" degrees of freedom needed to model the situation; the algorithmic needs (vis-a-vis the available resources) most definitely matter a lot.
If you don't know enough to design a data structure, requirements are missing and someone talking to the client is dropping the ball big time.
Where did I say any of that?
I'm saying that if you care about performance, data structures should be designed with approach specific tradeoffs in mind. And like I've said above, in typical business apps, it's ok to start with data structures because (a) performance is usually not a problem, (b) staying close to the domain is cleaner.
You said: "You can't design a data structure without knowing how you will use it."
But the whole discussion involves knowing how you will use it; the advocacy is for careful consideration of data structures (based on how you will use them) resulting in less pain when designing/choosing algorithms.
My point is that one doesn't follow the other. To design good data structures, you need to know how it'll get used (the algorithm).
> If you've chosen the right data structures and organized things well, the algorithms will almost always be self-evident.
This is what I was responding to.
See also:
"Show me your flowcharts and conceal your tables, and I shall continue to be mystified. Show me your tables, and I won’t usually need your flowcharts; they’ll be obvious."
https://en.wikiquote.org/wiki/Fred_Brooks
Full agree on this.
I (deep, deep in embedded systems) have seen this too often, that code is incredibly complex and impossible to reason around because it needs to reach into some data structure multiple times from different angles to answer what should be rather simple questions about next step to take.
Fix that structure, and the code simplifies automagically.
> But SQL schemas often look like this. Columns are nullable by default, and wide tables are common.
Hard disagree. That database table was a waving red flag. I don't know enough/any rust so don't really understand the rest of the article but I have never in my life worked with a database table that had 700 columns. Or even 100.
I saw tables with more than a thousand columns. It was a law firm home-grown FileMaker tool. Didn't inspect it too closely, so don't know what was inside
I remember a phrase from one of C. J. Date's books: every record is a logical statement. It really stood out for me and I keep returning to it. Such an understanding implies a rather small number of fields or the logical complexity will go through the roof.
https://jimmyhmiller.com/ugliest-beautiful-codebase
I kinda love this. That sounds like an incredibly entertaining place to work for between 1 and 2 years in your late 20s and not a second longer.
If you enjoyed this, you'd probably enjoy thedailywtf.com, which is full of stories like that.
With AI "programmers", this will be the future: bugs galore and the things that do work, work by accident.
I think this company was ahead of the curve.
This is awesome. Got completely lost reading this and was struggling to figure out where I got this link from. Amazing story.
This is like the functional ugly tier of buildings from "how buildings learn". Excellent stuff
The blog post is an entertaining read, but I was left with the impression the author might have tried do embellish, particularly in it's disbelief angle.
Take this passage:
> The app relied on a SOAP service, not to do any servicey things. No, the service was a pure function. It was the client that did all the side effects. In that client, I discovered a massive class hierarchy. 120 classes each with various methods, inheritance going 10 levels deep. The only problem? ALL THE METHODS WERE EMPTY. I do not exaggerate here. Not mostly empty. Empty.
> That one stumped me for a while. Eventually, I learned this was in service of building a structure he could then use reflection on. That reflection would let him create a pipe-delimited string (whose structure was completely database-driven, but entirely static) that he would send over a socket.
Classes with empty methods? Used reflection to create a pipe-delimited string? The string was sent over the wire?
Why congratulations, you just rediscovered data transfer objects, specifically API models.
Hi, I'm the author of the article.
As to your hard disagree, I guess it depends... While this particular user is on the higher end (in terms of columns), it's not our only user where column counts are huge. We see tables with 100+ columns on a fairly regular basis especially when dealing with larger enterprises.
Can you clarify which knowledge domains those enterprises fall under with examples of what problems they were trying to solve?
If it's not obvious, I agree with the hard disagree. Every time I see a table with that many columns, I have a hard time believing there isn't some normalization possible.
Schemas that stubbornly stick to high-level concepts and refuse to dig into the subfeatures of the data are often seen from inexperienced devs or dysfunctional/disorganized places too inflexible to care much. This isn't really negotiable. There will be issues with such a schema if it's meant to scale up or be migrated or maintained long term.
Normalization is possible but not practical in a lot of cases: nearly every “legacy” database I’ve seen has at least one table that just accumulates columns because that was the quickest way to ship something.
Also, normalization solves a problem that’s present in OLTP applications: OLAP/Big Data applications generally have problems that are solved by denormalization.
Yep, this comment sums it up well.
We have many large enterprises from wildly different domains use feldera and from what I can tell there is no correlation between the domain and the amount of columns. As fiddlerwoaroof says, it seems to be more a function of how mature/big the company is and how much time it had to 'accumulate things' in their data model. And there might be very good reasons to design things the way they did, it's very hard to question it without being a domain expert in their field, I wouldn't dare :).
> I can tell there is no correlation between the domain and the amount of columns.
This is unbelievable. In purely architectural terms that would require your database design to be an amorphous big ball of everything, with no discernible design or modelling involved. This is completely unrealistic. Are queries done at random?
In practical terms, your assertion is irrelevant. Look at the sparse columns. Figure out those with sparse rows. Then move half of the columns to a new table and keep the other half in the original table. Congratulations, you just cut down your column count by half, and sped up your queries.
Even better: discover how your data is being used. Look at queries and check what fields are used in each case. Odds are, that's your table right there.
Let's face it. There is absolutely no technical or architectural reason to reach this point. This problem is really not about structs.
Feldera speak from lived experience when they say 100+ column tables are common in their customer base. They speak from lived experience when they say there's no correlation in their customer base.
Feldera provides a service. They did not design these schemas. Their customers did, and probably over such long time periods that those schemas cannot be referred to as designed anymore -- they just happened.
IIUC Feldera works in OLAP primarily, where I have no trouble believing these schemas are common. At my $JOB they are, because it works well for the type of data we process. Some OLAP DBs might not even support JOINs.
Feldera folks are simply reporting on their experience, and people are saying they're... wrong?
Haha, looks like it.
I remember the first time I encountered this thing called TPC-H back when I was a student. I thought "wow surely SQL can't get more complicated than that".
Turns out I was very wrong about that. So it's all about perspective.
We wrote another blog post about this topic a while ago; I find it much more impressive because this is about the actual queries some people are running: https://www.feldera.com/blog/can-your-incremental-compute-en...
> Normalization is possible but not practical in a lot of cases: nearly every “legacy” database I’ve seen has at least one table that just accumulates columns because that was the quickest way to ship something.
Strong disagree. I'll explain.
Your argument would support the idea of adding a few columns to a table to get to a short time to market. That's ok.
Your comment does not come close to justify why you would keep the columns in. Not the slightest.
Tables with many columns create all sorts of problems and inefficiencies. Over fetching is a problem all on itself. Even the code gets brittle, where each and every single tweak risks beijg a major regression.
Creating a new table is not hard. Add a foreign key, add the columns, do a standard parallel write migration. Done. How on earth is this not practical?
There are sometimes reasons this is harder in practice, for example let’s say the business or even third parties have access to this db directly and have hundreds of separate apps/services relying on this db (also an anti-pattern of course but not uncommon), that makes changing the db significantly harder.
Mistakes made early on and not corrected can snowball and lead to this kind of mess, which is very hard to back out of.
> How on earth is this not practical?
Fine, but you still need to read in those 100+ fields. So now you gotta contend with 20+ joins just to pull in one record. Not more practical than a single SELECT in my opinion.
You don't need to join what you don't actually need. You also need to be careful writing your queries, not just the schema. The most common ones should be wrapped in views or functions to avoid the problem of everyone rolling their own later.
Performance generally isn't an issue for an arbitrary number of joins as long as your indices are set up correctly.
If you really do need a bulk read like that I think you want json columns, or to just go all in with a nosql database. Even then, the above regarding indexing is still true.
I think you believe the average developer, especially on enterprise software where you see this sort of shit, is far more competent or ambitious than they actually are. Many would be horrified to see the number of monkeys banging out nasty DDL in Hibernate or whatever C# uses that have no idea what "normal forms" or "relational algebra" are and are actively resistant to even attempting to learn.
No idea what these guys do exactly but their tagline says "Feldera's award-winning incremental compute engine runs SQL pipelines of any complexity"
So it sounds like helping customers with databases full of red flags is their bread and butter
> it sounds like helping customers with databases full of red flags is their bread and butter
Yes that captures it well. Feldera is an incremental query engine. Loosely speaking: it computes answers to any of your SQL queries by doing work proportional to the incoming changes for your data (rather than the entire state of your database tables).
If you have queries that take hours to compute in a traditional database like Spark/PostgreSQL/Snowflake (because of their complexity, or data size) and you want to always have the most up-to-date answer for your queries, feldera will give you that answer 'instantly' whenever your data changes (after you've back-filled your existing dataset into it).
There is some more information about how it works under the hood here: https://docs.feldera.com/literature/papers
It might not be common in typical software shops. I work in manufacturing and our database has multiple tables with hundreds of columns.
I’m working on migrating an IBM Maximo database from the late 90s to a SQL Server deployment on my current project. Also charged with updating the schema to a more maintainable and extensible design. Manufacturing and refurbishing domain - 200+ column tables is the norm. Very demoralizing.
What's in them?
Data from measurement tools. Everything about the tool configuration, time of measurement, operator ID, usually a bunch of electrical data (we make laser diodes) like current, potential, power, and a bunch of emission related data.
I think I'd rather work with very wide tables than the entity-attribute-value approach - which seems like a good idea but rapidly becomes a mess...
Property1 to 20 or more is an example. There are better ways to do it but I have seen columns for storing ‘anything’
Sounds like a generic form of single table inheritance. I don't honestly see any other way to do it (punting to a JSON field is effectively the same thing) when you have potentially thousands of parts all with their own super specific relevant attributes.
I've worked on multiple products that have had a concept of "custom fields" who did it this way too.
If lots of columns are a red flag then red flags are quite common in many businesses. I’ve seen tables with tens of thousands of columns. Naturally those are not used by humans writing sql by hand, but there are many tools that have crazy data layouts and generate crazy sql to work with it.
> I have never in my life worked with a database table that had 700 columns
Main table at work is about 600, though I suspect only 300-400 are actively used these days. A lot come from name and address fields, we have about 10 sets of those in the main table, and around 14 fields per.
Back when this was created some 20+ years ago it was faster and easier to have it all in one row rather than to do 20+ joins.
We probably would segment it a bit more if we did it from scratch, but only some.
https://apps.naaccr.org/data-dictionary/data-dictionary/vers...
771 columns (and I've read the definitions for them all, plus about 50 more that have been retired). In the database, these are split across at least 3 tables (registry, patient, tumor). But when working with the records, it's common to use one joined table. Luckily, even that usually fits in RAM.
Some businesses are genuinely this complicated. Splitting those facts into additional tables isn't going to help very much unless it actually mirrors the shape of the business. If it doesn't align, you are forcing a lot of downstream joins for no good reason.
It is very common to find tables with 1000+ columns in machine learning training sets at e-commerce companies. The largest I have seen had over 10000 columns.
It's OLAP, it very common for analytical tables to be denormalized. As an example, each UserAction row can include every field from Device and User to maximize the speed at which fraud detection works. You might even want to store multiple Devices in a single row: current, common 1, 2 and 3.
OLTP tables typically are normalized.
But OLAP tables (data lake/warehouse stuff), for speed purposes, are intentionally denormalized and yes, you can have 100+ columns of nullable stuff.
Not everyone understands normal form, much less 3rd normal form. I’ve seen people do worse with excel files where they ran out of columns and had to link across spreadsheets.
Salesforce by default comes with some where your tables have 50 columns before you start tweaking anything.
100s is not unusual. Thousands happens before you realise.
That statement jumped out at me as well. I've worked as a DBA on tons of databases backing a wide variety of ERPs, web apps, analytics, data warehouses...700 columns?!? No.
You've never seen an SAP database where the business object had a couple hundred fields? Its pretty much required if you're touching international data.
I have seen tables (SQL and parquet, too) that have at least high hundreds of optional columns, but this was always understood to be a terrible hack, in those cases.
> Hard disagree. That database table was a waving red flag.
Exactly this.
This article is not about structs or Rust. This article is about poor design of the whole persistence layer. I mean, hundreds of columns? Almost all of them optional? This is the kind of design that gets candidates to junior engineer positions kicked off a hiring round.
Nobody gets fired for using a struct? If it's an organization that tolerates database tables with nearly 1k optional rows then that comes at no surprise.
The database table is someone else’s data. That’s why this company exists and is explained in the article.
They don’t have the option to clean up the data.
Yep pretty much, feldera is the engine we don't control what SQL people throw at us.
Feldera co-founder here. Great discussions here.
Some folks pointed out that no one should design a SQL schema like this and I agree. We deal with large enterprise customers, and don't control the schemas that come our way. Trust me, we often ask customers if they have any leeway with changing their SQL and their hands are often tied. We're a query engine, so have to be able to ingest data from existing data sources (warehouse, lakehouse, kafka, etc.), so we have to be able to work with existing schemas.
So what then follows is a big part of the value we add: which is, take your hideous SQL schema and queries, warts and all, run it on Feldera, and you'll get fully incremental execution at low latency and low cost.
700 isn't even the worst number that's come our way. A hyperscale prospect asked about supporting 4000 column schemas. I don't know what's in that table either. :)
I have mixed feelings about it, but I'm going to fire somebody tomorrow for using a struct just to prove a point to the author.
Point them to us https://github.com/feldera/feldera -- we are hiring ;)
You folks have too many structs already! I just finished reading about it!
Here is an article I wrote this week with a section on Feldera - how it uses its incremental compute engine to compute "rolling aggregates" (the most important real-time feature for detecting changes in user behavior/pricing/anamalies).
https://www.hopsworks.ai/post/rolling-aggregations-for-real-...
Strictly speaking, Isn't there still a way to express at least one Illegal string in ArchivedString? I'm not sure how to hint to the Rust compiler which values are illegal, but if the inline length (at most 15 characers) is aliased to the pointer string length (assume little-endian), wouldnt {ptr: null, len: 16} and {inline_data: {0...}, len: 16} both technically be an illegal value?
I'm not saying this is better than your solution, just curious :^)
> Isn't there still a way to express at least one Illegal string in ArchivedString?
There may be good reasons (I don't know any) why it wasn't done like this, but from a high-level it looks possible to me too yes.
I feel like I'm missing something, but the article started by talking about SQL tables, and then in-memory representations, and then on-disk representation, but...isn't storing it on a disk already what a SQL database is doing? It sounds like data is being read from a disk into memory in one format and then written back to a disk (maybe a different one?) in another format, and the second format was not as efficient as the first. I'm not sure I understand why a third format was even introduced in the first place.
Feldera is an incremental query engine, you can think of it as a specialized database. If you have a set of question you can express in SQL it will ingest all your data and build many sophisticated indexes for it (these get stored on disk). Whenever new data arrives feldera can instantly update the answers to all your questions. This is mostly useful when the data is much larger than what fits in memory because then the questions will be especially expensive to answer with a regular (batch) database.
Feel free to try it out, it's open source: https://github.com/feldera/feldera/
I did read the rest, but I'm stuck on the first part where their SQL table has almost a thousand cols. Why so many?
Why not use a struct of arrays?
https://en.wikipedia.org/wiki/Data-oriented_design
> Why not use a struct of arrays?
I would assume because then the shape of the data would be too different? SOAs is super effective when it suits the shape of the data. Here, the difference would be the difference between an OLTP and OLAP DB. And you wouldn't use an OLAP for an OLTP workload?
> This struct we saw earlier had 700+ of optional fields. In Rust you would never design a struct like this. You would pick a different layout long before reaching 700 Options. But SQL schemas often look like this.
Really? I've never had to do any serious db work in my career, but this is a surprise to me.
Looks like they just recreated a tuple layout in rust with null bit map and everything, next up would be storing them in pages and memmap the pages.
https://www.postgresql.org/docs/current/storage-page-layout....
Absolutely, it's a very common technique :)
I wasn't sure about writing the article in the first place because of that, but I figured it may be interesting anyways because I was kind of happy with how simple it was to write this optimization when it was all done (when I started out with the task I wasn't sure if it would be hard because of how our code is structured, the libraries we use etc.). I originally posted this in the rust community, and it seems people enjoyed the post.
Just cus structs and classes work differently, and classes are much more common. I tend to make everything a class, unless there is a really good reason to make it a struct.
> Just cus structs and classes work differently
The most common programming language where "struct" and "class" are both kinds of user defined type is C++. In C++ they only "work differently" in that the default accessibility is different, a "struct" is the same as a "class" if you change the accessibility to public at the top with an access specifier.
If you think you saw a bigger difference you're wrong.
Classes are a safe default even if you expect things to go very, very fast.
The overhead of screwing up NUMA concerns vastly outstrips any kind of class vs struct differences. It's really one of the very last things you should be worrying about.
Allocating an array of a class vs an array of struct might seem like you're getting a wildly different memory arrangement, but from the perspective of space & time this distinction is mostly pointless. Where the information resides at any given moment is the most important thing (L1/L2/L3/DRAM/SSD/GPU/AWS). Its shape is largely irrelevant.
If I understand this problem was in rkyv, and solution is using rkyv with glue code. I hope they could integrate some sort of official derive macro `rkyv::Sparse` for this if it can't be done automatically in rkyv.
No one has written a struct in 10 years.
They’re pretty popular in Go?
And pervasive in Rust.
Why is rust allowed to reorder fields? If I know that fields are going to be generally accessed together, this prevents me from ordering them so they fit in cache lines.
You can choose in Rust to explain the representation you want for your data type. Unlike C or C++ that's not a non-portable vendor extension it's just part of the language, look at the repr documentation: https://doc.rust-lang.org/nomicon/other-reprs.html
So if you want "what C does" you can just repr(C) and that's what you get. For most people that's not a good trade unless they're doing FFI with a language that shares this representational choice.
It's allowed as an optimization, the order it uses will limit the space lost to field alignment.
You can tell it not to reorder them if you want but it's not the default.