For example, you write S-expressions to files like Hacker News does. This is clever, because the file system has some of the features of a database system, and files and S-expressions are abstractions that already exist. You do have to manage what data is in memory and what data is on disk at any given time, but the complexity and amount of code are low.

The idea that "event sourcing" somehow keeps you from needing a DB is ridiculous. By the time you've defined the event format, and written the software to replay the logs, etc., which if you're smart will be fairly general and modular, congrats, you've just written a database. At best, you keep complexity low, and it's another example of a small custom DB for a case where you don't need a fancy off-the-shelf DB. Maybe it's the perfect solution for your app, but it's still a database.

"Memory images," as a completely separate matter, are an abstraction that saves you some of the work of making a DB. Just as S-expressions can save you from defining a data model format, and files can save you from a custom key-value store, memory images as in Smalltalk could save you from having to deal with persistence. And if your language has transactions built in, maybe that saves you from writing your own transaction system. In general, though, it's very hard to get the DB to disappear, as there is a constellation of features important to data integrity that you need one way or another. It's usually pretty clear that you're using a DB, writing a DB, or using a system that already has a DB built in. If you think there's no DB, there's a high chance you're writing one. Again, that could be fine if you don't need all the features and properties of a robust general-purpose DB.

Funnily enough, in EtherPad's case, we had full versioned history of documents, and did pretty much everything in RAM and application logic -- a pretty good example of what the article is talking about -- and yet we used MySQL as a "dumb" back-end datastore for persistence. Believe me, we tried not to; we spent weeks trying alternatives, and trying to write alternatives. Perhaps if every last aspect of the data model had been event-based, we could have just logged the events to a text file and avoided SQL. More likely, I think, we would use something like Redis now.

It sounds that SQL itself wasn't the problem. Were you looking for versioning alternatives in SQL that weren't up to par?

For example, it's possible that logging all data model changes to a text file would have given us the persistence we were looking for without bridging all our data to SQL. Cutting out SQL from our production set-up would have been an inherent win -- one less process to manage, one less black-box source of complexity, etc.

While I'm no fan of databases, one obvious advantage is that they provide direct access to the data in a standard way that is decoupled from the specific application code. This makes it easy to perform migrations, backups etc. It also increases one's confidence in the data integrity. Any solution that aims to replace databases altogether must address these concerns. I think that intimately coupling data with the application state, as suggested in the article, does not achieve this.

In my view, direct access to data decoupled from application code is a bug, not a feature. With multiple code bases touching the same data, schema improvements become nearly impossible.

I also think data integrity is easier to maintain with a system like this. SQL constraints don't allow me to express nearly as much about data integrity as I can in code. Sure, I could use stored procedures, but if I'm going to write code somewhere, I'd rather it be in my app.

If you are in the middle of a transaction and you realize that some constraint is being violated, how do you roll it back without interfering with the other transactions?

Uh, OK. So, you're happy with single-core boxes then, I take it?

Actually, you're regressing to even before that, when there was no pre-emptive multitasking. When the program is done doing something, it yields control to some other task.

Also, I'd like to point out that just because you aren't explicitly doing I/O doesn't mean that you aren't doing I/O. The OS might have paged out some stale data (quite likely, since you aren't managing I/O yourself), and you might be holding the giant lock while it's paging it in.

"I can't speak to all systems like this, but the Prevayler approach is pretty straightforward."

I just want to clarify: so when you encounter a problem, you do some rollback, which automatically moves the state to the last snapshot and rolls forward to the previous transaction, right? No manual steps?

I hope you have a recent snapshot or that will be a long wait (while holding the big lock, I might add).

Not at all. You use only one core for the core execution of write transactions, but that's a small part of any real system. All cores can read simultaneously. All cores can also do all sorts of other work, including preparing transactions to execute, deserialization of requests, rendering responses, logging, and anything else your app needs to get up to.

The limit is also one core per transactional domain. If you can split your data up into lumps between which you never need transactions, you can happily run one core on each.

> Also, I'd like to point out that just because you aren't explicitly doing I/O doesn't mean that you aren't doing I/O.

Actually, it does explicitly do I/O. You do it just before every command executes.

> The OS might have paged out some stale data.

I guess that's possible, which would indeed cause a momentary pause, but this approach is typically used with dedicated servers and plenty of RAM, so it's never been a problem in practice for me.

> I just want to clarify: so when you encounter a problem, you do some rollback, which automatically moves the state to the last snapshot and rolls forward to the previous transaction, right?

You mean a bug in our code that causes a problem? Depends, on the system, I suppose. Prevayler had an automatic rollback. It just kept two copies of the data model in RAM; if a transaction blew up it would throw out the possibly tainted one. But there are a number of ways to solve this, so I don't advocate anything in particular. Other than heavy unit testing, so that things don't blow up much.

Unless someone is writing, of course, in which case you have to worry about isolation. So a single writer would block all readers, right?

"You mean a bug in our code that causes a problem?"

No, I mean like "I already wrote some data, but now a constraint has been violated so I need to undo it".

Correct. For the fraction of a millisecond the transaction is executing, anyhow. Since transactions only deal with data hot in RAM, transactions are very fast.

> No, I mean like "I already wrote some data, but now a constraint has been violated so I need to undo it".

That shouldn't happen, and I've used two approaches to make sure. One is do all your checking before you change anything. The other is to make in-command reversion easy, which is basically the same way you'd make commands undoable.

Basically, instead of solving the problem with very complicated technology (arbitrary rollback), you solve it with some modest changes in coding style. Since you never have to worry about threading issues, I've found it pretty easy.

Transactions don't just read and write. They sometimes compute things, like joins, which can take several milliseconds. These computations often must run within the transaction and would thus need to acquire the lock for several milliseconds.

Joins haven't been a problem for me, mainly because this approach doesn't constrain you to a tables-and-joins model of the world. With Prevayler, for example, you treat things as a big object graph, so there are no splits to join back up.

Of course, it could be that some problem is just computationally intense, but I can think of a number of approaches to lessen the impact of that in a NoDB system.

Surely I could just as easily say 'Sure, I could use a data access layer in my app but if I'm writing a multi-app database, I'd rather the database enforced its own integrity.'?

Nothing's the perfect tool for every job, but I certainly think stored procedures have their place and have the power to handle the bulk of tasks.

Personally, though, I'd much rather do my important coding in a real programming language. Better tools, more libraries, bigger communities, and no vendor lock-in.

Modern SQL dialects are Turing complete and frankly pretty rich dialects. I know MS SQL Server best so can't speak in detail for others, but the community around that is certainly very substantial. Library support, well, doesn't work quite the same way (yet!) but there's plenty of libraries of code samples available for adapting. Vendor lock-in is a tricky one; by the time you've got to a certain scale of application I tend to think you're programming as much to the API (whether it's the provider's standard API or your own specific API layered over the underlying platform) as to the official 'language'; lock-in can creep up surprisingly easily. Facebook avoided vendor lock-in by writing in open PHP and have since had to write their own PHP compiler to get the performance they needed from the solution they were locked in to.

A former employer used to bulk process EDI order lines in very large quantities. Deduplicating them, dynamically rebatching them according to what was available and what wasn't, updating orders with newer product where customer had specified 'this or better', cross-referencing against multi million row datasets of cataloguing and tagging information to identify how to handle the item. It was a monster; I hate to think about the volumes of data that touched each batch, and with processing orders it absolutely had to have transactional integrity. And yet, written in SQL and running on a very average commodity server, it was actually very fast. The data never left the server until it was ready to do so and all the loads stayed happily internal. The implications of trying to implement it on a NoDB solution - the volumes of data being passed around, the amount of data specific library code the DBMS provides but the underlying language doesn't which would need reading..... It's not pretty.

I don't maintain SQL is the perfect language for everything, that's patently silly. But I do maintain it's a lot more powerful (and with good performance and reliability) than it's given credit for on some very complex operations, and that a lot the reasons people prefer to work in alternatives boil down to lack of understanding. A little learning of what a modern DBMS is capable of can reap huge rewards of saving work in the 'real programming language', as you put it.

I'm sure the developer community for MS SQL Server is reasonably large, but it is much, much smaller than the Ruby community or the Java community. The same is true for library code for each environment.

I think the Facebook example cuts the other way. If they had implemented all their core application logic in MS SQL Server, they would have been well and truly screwed if the performance wasn't enough. With PHP, at least they could write their own compiler; trying to reverse-engineer MS SQL Server is orders of magnitude harder.

Regarding the "volumes of data being passed around" part, that works well with a hot-in-RAM system; no data is passed anywhere. As with doing it all in stored procedures, no data leaves the server.

I do agree that a lot of people don't get full value out of their database. Sometimes that's a very reasonable business choice: vendor lock-in is extremely expensive at scale. But it does often come from ignorance. On the other hand, almost every developer has written a few database-backed applications, but very few have written anything in the NoDB style. Many can't even conceive of coding something without a database. I'd love to see both sorts of ignorance reduced.

Edit: Also, not all databases follow the standard very closely

My interpretation of the parent post was that it was a response to a comment about vendor lock in. I was only trying to point out that it is not always possible to ensure compatibility between databases by writing strict ANSI SQL.

It stores data just fine and is not vendor specific.

ANSI SQL is not vendor specific but it is just a standard not an implementation. As a result you have to rely on vendors to implement the language. Many vendors deviate from the standard. This means that you cannot just write ANSI SQL and expect it to work on all databases.

http://en.wikipedia.org/wiki/SQL#Procedural_extensions

It looks like a lot of vendors still only support propriatary extensions.

It's unclear that multiple applications with direct access to said data make schema improvements any easier.

The obvious solution, copying the data for applications that are using the new schema, pretty much guarantees that one or more of the copies are wrong.

> I also think data integrity is easier to maintain with a system like this. SQL constraints don't allow me to express nearly as much about data integrity as I can in code. Sure, I could use stored procedures, but if I'm going to write code somewhere, I'd rather it be in my app.

How do you guarantee that all of the apps that touch that data use the current version of said code?

Code normalization is as important as data normalization.

> How do you guarantee that all of the apps that touch that data use the current version of said code?

An approach that may be worth considering is to not require all apps to use the current version: allow multiple versions. For some cases this would work, say if the semantics of the newer version are backwards compatible with the semantics of the older version. If the data semantics are preservable, transforming a schema could happen while each data access request to the schema is actively transformed.

But it clearly wouldn't work in all cases. More work to handle that would be needed.

> Code normalization is as important as data normalization.

True, this is the near show-stopper really. In that case, the best one can hope for is preparing the state of the new version (new data in new schema) and carefully coordinating a quick restart of the old version for the new version.

I would love to hear your thoughts on this. We have been working towards that direction with ChronicDB (http://chronicdb.com) and would welcome feedback.

An e-mail address in your profile would have made that possible. (Chronicdb looks interesting and complements something that I've been thinking about. I suspect that you've implemented many of the relevant mechanisms.)

In a NoDB app? It's very easy: only one code base ever directly touches the data, because the data lives in the RAM allocated to the app. You give external access via an API, so integrity is very easy to enforce.

But one has to wonder, what happens when you need to upgrade the app? Shutting down the process and destroying the memory image doesn't seem like the best option:

- First, it disrupts connected applications since the process is killed, introducing downtime.

- Second, when starting up again in say version 2, the data that will be loaded in memory still needs to be transformed in the format expected by version 2. This transformation can take time on large data, introducing further downtime.

The challenge would be to eliminate this downtime by combining a solution for both client disruption and state transfer. A data abstraction like a database using SQL can simplify such a solution.

We have been working on building what we hope are the procedures for this with ChronicDB (http://chronincdb.com). But it turned out harder than it seems, and we are not sure it will quite work out. We'd welcome feedback.

By basing your system on a journaled event stream, you start with a foundation of complete history retention, and you can build exactly the sort of reporting views you need at any time (say, by creating a SQL database for other applications to query).

FTFY...

It's not hard to build history into a SQL table design. You can even store events in a...wait for it... SQL database. I have built numerous systems backed by SQL databases that have complete history retention. Answering questions like 'who had id 'X' on this date 3 years ago' are easily solvable with basic standard sql.

I certainly don't believe SQL databases are perfect or the tool for every job, but in many cases they work just fine until you get into very large datasets. Admittedly, I only deal with databases in the 100s of GB range so I have yet to personally run into the scaling problems that a Google or Facebook have and the SQL backed systems I have built work just fine.

If your system operates in this journaled/event-sourcing way at the most basic level then you have the ultimate future-proof storage layer. You could decide to completely change the way the data is stored and represented (in-memory or otherwise) at any time, as long as you have that raw history.

For example if you only want customers to create orders you can set a foreign key between customer and order. If you don't delete orders and customers, then the database will allow you to insert new orders against old customers, unless you apply a more complex constraint.

That leaves tracking the remaining inserts and updates, which is a well understood problem. It's called Auditing. Here is a simple script that will auto-audit a SQL Server database... variations in other SQL dialects are likely just as straighforward.

http://www.geekzilla.co.uk/ViewECBC0CC3-1C7E-4E7E-B243-F2F25...

But I wouldn't write off the noDB approach for various transitional data, or data that isn't mean to live long anyway, like tweets.

There are a lot of questions that make the choice of data store a difficult one, but I'm not sure that plays into it at all.

I suppose it would be equally possible to think of the main data tables as a "reporting view" in the sense you use here, except that the application was 1000:1 or more in update frequency to read frequency, and all the reads were performed on the main data tables, so that's kind of a "tail wagging the dog" view of the app.

For an application with different requirements, your view of things might be quite useful, of course.

http://wiki.postgresql.org/wiki/MVCC

"The key element to a memory image is using event sourcing, which essentially means that every change to the application's state is captured in an event which is logged into a persistent store."

That is a key element of a database. It's called a logical log.

"Furthermore it means that you can rebuild the full application state by replaying these events."

Yup, logical log.

"Using a memory image allows you to get high performance, since everything is being done in-memory with no IO or remote calls to database systems. "

This is _exactly_ what sophisticated old-school databases do. You can have them require to write to the DB on commit, or just to memory, and have a thread take care of IO in the background.

"Databases also provide transactional concurrency as well as persistence, so you have to figure out what you are going to do about concurrency."

Righty-ho.

"Another, rather obvious, limitation is that you have to have more memory than data you need to keep in it. As memory sizes steadily increase, that's becoming much less of a limitation than it used to be."

So why not store your old-school DB in memory?

I can understand the argument that you don't want to lock into a big DB vendor's license path, but the technical arguments here look distinctly weak to me.

Maybe old-fashioned DBs are hipper than people think?

Also it's important to keep a good decoupling between the events and the model structure itself. It may be tempting to come up with some automatic mapping system that retrospects on the event data and the model, but this couples the events and model together which makes it difficult to migrate the model and still process old events.

So, you're right that he doesn't envision a translation to a relational model but it's not just object structures either.

With such an application I'd probably still be writing the events themselves to a relational database for archiving and potentially sending out report-oriented data as well. I'm not sure how all of that would work out re: ultimately the whole app needs to be stored in an RDBMS anyway for various reasons but it seems interesting to try.

Since the points I've highlighted argue that he's talking about database features I'd call that a database without impedance mismatch, like erlang's Mnesia.

The interesting thing that Prevayler and such things did was that they expanded the use of logical logs beyond simple relational tables to much richer sets of state.

Not sure that justifies dumping DBs altogether, but it's still an interesting advantage.

The idea that you can reliably use event sourcing in memory to persist your data is as foreign to me as it is impressive. Is anyone familiar with major applications (web apps, ideally) that use this method for their data persistence?

Financial trading is another area where databases are too slow. I know of one place that uses this approach to keep pricing data hot in RAM for their financial models. And Fowler previously documented using this for a financial exchange:

http://martinfowler.com/articles/lmax.html

Of course, Fowler talks about 'snapshotting' the data, which, again, makes me wonder if playing with all of this resident memory and the systems needed to make that happen haven't already been solved by...um...databases.

Expand your thinking in the abstract about what a database is. As 71104 mentions, a file system is also a database. What you are thinking of as "a database" is really a specific type of key-value store that is located on disk. But the fact most DBs are on disk has nothing to do with the concept itself.

The big players are backing their services with custom data stores, though.

To the extent that you could actually write your program as a pure function of its past input history — ideally, one whose only O(N) part (where N was the length of the history) was a fold, so the system could update it incrementally as new events were added — you could get schema upgrade and decentralization "for free". However, to get schema upgrade and decentralization, your program would need to be able to cope with "impossible" input histories — e.g. the same blog post getting deleted twice, or someone commenting on a post they weren't authorized to read — because of changes in the code over the years and because of distribution.

I called this "rumor-oriented programming", because the propagation of past input events among the nodes resembles the propagation of rumors among people: http://lists.canonical.org/pipermail/kragen-tol/2004-January...

I wrote a bit more on a possible way of structuring web sites as lazily-computed functions of sets of REST resources, which might or might not be past input events: http://lists.canonical.org/pipermail/kragen-tol/2005-Novembe...

John McCarthy's 1998 proposal, "Elephant", takes the idea of writing your program as a pure function of its input history to real-time transaction processing applications: http://www-formal.stanford.edu/jmc/elephant/elephant.html

The most advanced work in writing interactive programs as pure functions of their input history is "functional reactive programming", which unfortunately I don't understand properly. The Fran paper http://conal.net/papers/icfp97/ is a particularly influential, and there's a page on HaskellWiki about FRP: http://www.haskell.org/haskellwiki/Functional_Reactive_Progr...

- Long startup times as the entire image needs to be loaded and prepared.

- It would be hard to distribute the state across multiple nodes

- What happens in case of a crash? How fault tolerant would this be?

- Does this architecture essentially amount to building in a sort-of-kind-of datastore into your already complex application? Without a well-defined well-tested existing code base, is this just re-inventing the wheel for each new project?

- How do you enforce constraints on the data?

- How do transactions work (debit one account, [crash], credit another account?

- How do you allow different components (say web user interface, admin system, reporting system, external data sources) to share this state?

Just curious.

EDIT:

- Isn't this going to lead to you writing code that almost always has side-effects, causing it to be really hard to test? How would you implement this system in Haskell?

- Mirroring state is easy; you just pipe the serialized commands to multiple boxes.

- It's very fault tolerant. Because every change is logged before being applied, you just load the last snapshot and replay the log.

- It didn't seem that way to me.

- In code. In the system I built, each mutation was packaged as a command, and the commands enforced integrity.

- Each command is a transaction. As with DB transactions, you do have to be careful about where you draw your transaction boundaries.

- Via API. Which I like better, as it allows you to enforce more integrity than you can with DB constraints.

> - The startup times can be a problem if you have a lot of data. Modern disks are pretty fast for streaming reads, though, and you can split the deserialization load across multiple processors.

Reading data, at even a GB/second from disk (which is currently not possible) is going to mean a second spent of a GB of data, just to read, let alone deserialize. That's with reading a snapshot, not replaying old transactions.

> - Mirroring state is easy; you just pipe the serialized commands to multiple boxes.

That's not distributing the load. I'm talking about having more data than fits in an reasonable amount of RAM (say 1TB). Also mirroring is nice for when you want read-only access to your data. You'll have the same problem as any other data store when you want multiple writers. Also, is replication synchronous or asynchronous (which end of CAP do you fall on)?

>- It's very fault tolerant. Because every change is logged before being applied, you just load the last snapshot and replay the log.

So it's going to at the speed of the disk then (http://smackerelofopinion.blogspot.com/2009/07/l1-l2-ram-and...). Don't get me wrong, this is still faster than writing to the network, but then writes are way slower than reads.

My other question is how much of a pain in the ass is it to debug such a system? I suppose if you have a nice offline API to look at your data, change something, revert back, etc, it would work well, but if it's deep within your normal application, it could become nightmarish.

If that's just saying that startup time can be an issue, I agree. There are a variety of techniques to mitigate that, though. The simplest is to compress snapshots and/or put them on RAID, boosting read speed. The most complicated is just to have mirrored servers and only restart the one not in use right now.

> I'm talking about having more data than fits in an reasonable amount of RAM (say 1TB).

For something where you need transactions across all of that? This architecture's probably not a reasonable approach, then. The basic precondition is that everything fits in RAM. However, sharding is certainly possible if you can break your data into domains across which you don't require consistent transactions.

> So it's going to at the speed of the disk.

Sort of.

Because it's just writing to a log, mutations go at the speed of streaming writes, which is very fast on modern disks. And there are a variety of techniques for speeding that up, so I'm not aware of a NoDB system for which write speed is the major problem.

Regardless, it's a lot better for writes than the performance of an SQL database on the same hardware.

> My other question is how much of a pain in the ass is it to debug such a system?

It seemed fine. A big upside is that you have a full log of every change, so there's no more "how did X get like Y"; if you want to know you just replay the log until you see it change.

Last I did this we used BeanShell to let us rummage through the running system. It was basically like the Rails console.

What? No, that's ridiculous. That's how inconsistencies crop up. Unless you plan on locking the entire system during each command.

That seems ridiculous if you are thinking like databases do, in terms of taking away the pain of all those disk seeks needed to write something. But if everything is hot in RAM, executing a command is extremely fast. Much faster than a database.

If that still isn't fast enough, you can split your data graph into chunks that don't require simultaneous locking and have one lock per. For example, if you are making a stock exchange like the LMAX people were, you can have one set of data (and one lock) per stock.

The first "fix" to this that comes to mind is timestamping data, rolling back transactions which try to write having read outdated data. Do extant NoDB systems do this?

Imagine you have a graph of objects in RAM. Say, a Set of Counters. Imagine you want to change the graph by incrementing a counter. To do this create an object, the UpdateCounterCommand, with one method: execute. You hand the command to the system, and it puts it in a queue. When it gets to the top, it serializes the command to disk and then executes it. Exactly one write command runs at a time.

For a real-world example, check out Prevayler. It provides all the ACID guarantees that a database does, but in a very small amount of code.

> - Isn't this going to lead to you writing code that almost always has side-effects, causing it to be really hard to test? How would you implement this system in Haskell?

You'd have to figure out how to isolate the IO monad as much as possible, but this is no different than interacting with a database in Haskell. And Haskell would give you nice features like STM to address other concerns as well.

Sven Van Caekenberghe (the author of cl-prevalence) and I used this approach to power the back-end/cms of a concert hall back in 2003. A write-up of our experiences can be found at http://homepage.mac.com/svc/RebelWithACause/index.html

The combination of a long-running Lisp image with a remote REPL and the flexibility of the object prevalence made it a very enjoyable software development cycle. It's possibly even more applicable with the current memory prices.

I especially liked the fact that your mind never needs to step out of your object space. No fancy mapping or relationship tables, just query the objects and their relations directly. I guess that's what SmallTalk developers also like about their programming environment.

http://book.seaside.st/book/advanced/persistency/sandstone

I actually have a couple Clojure apps that rely on a hefty amount of in-memory data to do some computations. Even the cost of pulling the data from Redis would be too expensive. The in-memory data grows very slowly, so it's easy to maintain. Moving faster-growing data in-process would be trickier, but this article makes me want to try.

Maintenance: I can easily give a 10% raise to everyone with a single SQL statement. Fowler's method requires that I first create an entire infrastructure (transaction processing, ACID properties) in code for this particular application. And it had better be as reliable as the transaction processing available in modern relational databases (so says my boss) or I'll be looking for a new job.

Support: you get to teach the new guy how "Event Sourcing" works for this application A and also applications B, C, ....

That said, I _have_ done this with great success. But the work involved a single application (a minicomputer-based engineering layout system). The ease with which versioning could be included was a selling point.

And don't get me started on reporting or statistics.

As to creating the infrastructure and worries about reliability, there are a number of frameworks for this. E.g., Prevayler. It gives you all the ACID guarantees, but has about three orders of magnitude less code than a modern database.

Supporting it could definitely be a problem. That's true for anything novel, so I'd only do this where the (major) performance benefits outweigh the support cost.

Some kinds of statistics are easier with this. For example, if you want to keep a bunch of up-to-date stats on stocks (latest price, highs, lows, and moving averages for last hour, day, and week) it is almost trivially easy in a NoDB system, and much, much faster than with a typical SQL system.

For other stats and reporting, though, dumping to an SQL database is great. For many systems you don't want to use your main database for statistics anyhow, so a NoDB approach mainly means you start using some sort of data warehouse a little earlier.

Database engines are not just for storing - each is basically a "utility knife" of data retrieval - indexing, sorting and filtering are available via (relatively) simple SQL constructs. If your app uses an index right now, ditching the DB will mean re-implementing it manually. It's not hard, but it's extra code.

So basically, the DB engine might still be a necessary "library", at least for data retrieval. A middle-of-the-road take on this is e.g. using an in-memory Sqlite instance to perform indexing, etc - seeding it at run-time to help with data searches, but then still not using it for storing persistent information and discarding the data at the end.

The area where I still most needed SQL was for ad-hoc querying and reporting. I dreamed of building a relational-to-object mapper, but settled for a) XPath queries against our snapshot files, and b) dumping to an SQL database for reports.

If you're building a trading platform that handles 6M transactions/second, you have the money to handle this in the application layer and the load to justify the expense. But for many other tasks, you may be wasting money or putting data at risk.

Standard tools are useful because you can get to working code fast ... this is why LAMP is still such a powerful framework upon which to build. While it may make sense to consider adding a search indexer (Solr) or key-value cache (Redis), for almost every use case, rewriting data storage is a waste.

Also, to paraphrase Ted Dziuba, it probably doesn't matter if your product doesn't scale, because nobody cares, or will ever use it. So I think it is better to get something up and running quickly to see if anyone cares before you bother trying to optimize for the rare case where your product turns out to be the next Twitter.

The impedance mismatch between database and application is a lot of code too.

If, on the other side, I had to do it all myself, I'd most probably have lost my last hair.

Compared with a database-backed system, many operations are thousands of times faster. Some things that I was used to thinking of as impossible became easy, and vice versa. Coming to grips with why was very helpful.

However, I wouldn't call this "no-DB." Rather, it's "less-db." Ultimately, historical and statistical data needs to be stored and databases are great for that (and for a stats team).

Doing this in-memory in java via someAtomicLong.incrementAndGet() sounds appealing.

Just for fun, in Clojure:

    (def current-id (atom (long 0)))
    (defn get-id [] (swap! current-id inc))

Other example vector graphic editors: it replays vector graphics primitive instead of pixel bitmaps.

What about transaction isolation? How do you keep one transaction from seeing partial results from a concurrent transaction? Sounds like a recipe for a lot of subtle bugs.

And all of the assumptions you need to make for this no-DB approach to be feasible (e.g. fits easily in memory) might hold at the start of the project, but might not remain valid in a few months or years. Then what? You have the wrong architecture and no path to fix it.

And what's the benefit to all of this? It's not like DBMS do disk accesses just for fun. If your database fits easily in memory, a DBMS won't do I/O, either. They do I/O because either the database doesn't fit in memory or there is some better use for the memory (virtual memory doesn't necessarily solve this for you with the no-DB approach; you need to use structures which avoid unnecessary random access, like a DBMS does).

I think it makes more sense to work with the DBMS rather than constantly against it. Try making simple web apps without an ORM. You might be surprised at how simple things become, particularly changing requirements. Schema changes are easy unless you have a lot of data or a lot of varied applications accessing it (and even then, often not as bad as you might think) -- and if either of those things are true, no-DB doesn't look like a solution, either.

Basically instead of making a transaction between 2 entities, you send a message to the first reserving some data, a message to the second reserving the data and once you get confirmation from both (or however many entities are involved in the transaction) you send a commit to them.

These reservations can be revoked though. Your rollback has to be managed by an "activity".

Ex: Bank transfers. You have the activity called BankTransfer. It manages the communication between entities and the overall workflow. It starts by sending messages to entities Account#1 with 100$ in it and Account#2 also with 100$. To #1 it says debit 500$. To #2 it says credit 500$. #2 responds first and says Done. #1 responds second and says Insufficient Funds. BankTransfer sends another message to #2 saying Cancel event id 100 (the crediting).

Other activities that want to read the state of number 1 will see 100$ in it but the transfer (as yet unconfirmed) had been of 50 rather than 500$ and another debit of 75$ comes in it would respond insufficient funds. At this point it's the activity's job to decide what to do. Wait and try again? Fail entirely and notify any other entities relevant to the workflow? That's up to the business rules. Also, since the credit has not yet been confirmed, reading the balance on #2 would still say 100, not 600$.

Of course, depending on your use case you may want the read to return the balance with unconfirmed transactions. That's entirely up to the application code and business rules but the example should be explanatory as to how rollback is implemented.

Eventual consistency is the only scalable way to go for very large systems.

If an event causes the model to be in an invalid state, another event must be triggered to rectify the model into a valid state. (Simplistically speaking)

As far as I can see, they seem to be comparable to b+tree indexed key value stores. A key would e.g. be "/home/user/test.txt". Thanks to the B+Tree "indexation" you can do a prefix scan and list folders (e.g. "ls /home/user/"--> all keys starting with "/home/user/").

In the case of e.g. ReiserFS they actually use B+Trees. They have a caching layer managed by the OS. Most of them have journaling which would be the equivalent of a "write ahead log".

Map reduce based "view" generation can easily be done by pipes and utilities like grep. We might be even able to do some sort of simplistic filtering/views/relations using symlinks.

I guess the main difference is that they aren't optimized for this database-like behavior from a performance standpoint and that the network interfaces to them are SMB/AFP/NFS.

I am working on a team that is building an in-memory SQL database. It features a custom language that makes it possible to push the time-critical, data-processing parts of the application directly to the database, which allows for the same speed as this no-DB approach. But you don't have to build your own DB and do everything yourself (correct persistence, backup, transactions...)

[1]: http://www.infoq.com/presentations/Command-Query-Responsibil...

[2]: http://www.udidahan.com/?blog=true

[3]: http://blog.jonathanoliver.com/

[4]: https://github.com/joliver/EventStore/

http://codebetter.com/gregyoung/2010/02/16/cqrs-task-based-u...

"CQRS is not eventual consistency, it is not eventing, it is not messaging, it is not having separated models for reading and writing, nor is it using event sourcing."

It is gaining a lot of traction simply because it seems like a complicated and cool way to solve an uncommon problem. I fear that Event Sourcing will soon become a over-engineered hammer for the wrong nail.

If it were me, I would use CQRS principles, with a relational backend as my source model. Then when the need for scale arises, use ETL to either non-relational db or no-db for queries.

There are many ways get rid of the problem with getting rid of the database. For example putting all knowledge of the domain in the database via stored procedures, couchapps, and object stores.

why store passwords through SQL when a server application could just use a specific directory containing one file for each user, each file named with the username and containing his password (without any file format, just the password, possibly hashed or encrypted)? the operating system I/O cache should be able to handle that efficiently and the advantage would be the elimination of the dependence on another software, the DBMS.

More generally, for all but the very simplest apps I'm afraid can't agree with you (though, to clarify, this isn't really addressing the topic of the original post). Years ago I had a summer job maintaing a set of Perl scripts that persisted online store inventories to flat files. It was horrible. I've never found Postgres to add much complexity to a project, but on the storage side it provides me with the reassurance that smart people have thought hard about issues of consistency, stability and performance -- which millions have tested -- and on the query side, I can do various clever things at amazing speed if ever that's required. Why would I not go for that?

Because your advertisers want to know how many users signed up last month, last six months, and last year. When you only consider one use-case for your data, it's easy to consider using NoSQL or the file system to store your data but in doing so you fail to imagine all the other ways you might want that very same data.

I have written several systems that used that sort of approach. Pretty soon you realize you just reimplemented SELECT, and you did a buggy, half-baked job of it.

If you have money to burn on speed & reliability, dropping the database is a good idea. Otherwise, I simply have written too many half-baked hardwired select queries to recommend it.

At this point, I'd rather do some kind of in-memory SQLite with a persistent MySQL/PgSQL backend.

Also, see maildir and various things associated with qmail.

http://happstack.com/index.html

I've always thought that software-transactional memory and persistent distributed heaps would get us there. Unfortunately the nearest things have been Redis and Terracotta plugged into Clojure. It should be:

Insert? new an object. Delete? dispose an object. Look up? Hash table.

Solved problems that just require persistence.

The individual transactions could also easily be distributed to multiple nodes via a message queue.

One drawback to keep in mind is that LDAP is generally not meant for lots and lots of writes so it is by no means a substitute for DBs but it is great for looking up data and if that data somehow fits a sort-of "file card" paradigm anyway and there are way more reads than writes on that data and several different applications should be able to access it then all the better.

The major and most popular applications of LDAP, however, are certainly always somehow connected to authenticating users and that is also where it really, really shines and that was another reason I brought it up. If applicable, personally I would prefer managing users and their logins in an LDAP server over keeping all that in a database.

Luckily nowadays most (web) applications offer some sort of support for using LDAP anyway, however dodgy those implementations sometimes are. (One of my favorite examples here is netscape navigator/mozilla/thunderbird and the addressbook schema shenanigans...)

I just think it gets too little credit or news these days but that probably stems from the fact that it is a pretty stable system without lots of innovations and it has been around a looong time and it is not so "sexy" anymore and most HN hackers wouldn't have to deal with it most of the time anyway.

But I cannot recommend looking into LDAP and playing around with it and understanding how to get a directory going enough - it is a bit confusing (sometimes frustrating) at first because it is so different from typical databases but it is fun once you get the hang of it and learn to appreciate its simple and efficient beauty and some of the things you can do in huge directories with e.g. the Sun LDAP server are nothing short of amazing.