SQL Databases Are An Overapplied Solution
adam.blog.heroku.com
adam.blog.heroku.com
Similarly, with the e-commerce orders: what if you want to retrieve by the product ordered? It's not unreasonable to assume a situation where certain products are fulfilled by warehouse A and others by warehouse B. Well, you're in the same boat again doing a less efficient thing.
Plus, what is hard to scale when it comes to a database? The article is passing off SELECT statements as hard to scale. Now, I'm not saying that they're a piece of cake. You can get into a lot of trouble. However, reads are relatively easy to scale since you can just add more boxes. Writes are hard to scale because, unless you shard and do other hard things, you only get the power of one box for writes (since the writes have to be done on every box while a read only has to occur on one box).
So, even if you're using a document based store, you eventually have to shard. Now, when you never relate data, sharding can be a lot easier since it can be done based on a hash function. Systems like memcached do this automatically. So when you say get(1, 42, 64, 128) it will be able to hash those ids and send each request to the proper server for that item. But that means that you lose out on a lot of ease. And most of these alternative stores don't do that for you (and it's why memcached is such a useful tool alongside a RDBMS).
And SQL databases do scale a lot more than Heroku (in these two articles) seems to let on. Wikipedia, Facebook, Craigslist, and Flickr are all backed by MySQL. Now, not MySQL alone. Memcached is a big part of it for all/most of them (I don't remember what each site uses exactly). There's a reason why many of the largest sites use a SQL database and it's not because they're unaware of other storage engines.
It seems like Heroku might be getting a lot of complaints that the service isn't magically scaling. Computers aren't magic and document based databases aren't any more magic. CouchDB uses B-Tree indexes just like relational databases. The difference isn't so much that these data stores offer better performance for some lookups. It's more that they only offer the lookups that can have good performance.
I feel like I should offer some free advice to Heroku: your SQL databases would scale better if your dedicated SQL boxes were Amazon's high-RAM boxes rather than the high-CPU ones they opted for. RAM means more for database performance than CPU. Oh, also, offer some consulting for clients on their database woes. A lot of the time, people are doing lookups that should be using an index, but they haven't created it and so the database has to do a full table scan rather than an index lookup. And there's a big difference there. A full table scan of 1 million records will take 50,000 times longer than an index scan. Yeah, indexes are good. One of the reasons that CouchDB "scales so well" is that you can only do queries on things you've made indexes for.
I'm not saying that non RDBMS don't have their place. They do. However, we keep seeing articles posted about RDBMS not scaling and it doesn't seem like they quite know the purpose of scale. Scale doesn't have to be infinite. Nothing will do that. The questions is: will it scale enough to handle the traffic? And SQL databases will unless you're really, really big. Do you expect your site to become a top 500 site on the internet? Heck, WordPress.com at #20 is doing fine implementing a not very efficient SQL backed blog. Now, I'm sure they implement caching and such, but it's still SQL backed.
Basically: learn a lot about indexes. If you become one of the top sites on the internet, hire someone who can help you. In the meantime make your product and don't worry too much about the FUD.
We do this at the day job. It runs about $X00 an hour, with a minimum of Y hours, if our customers need it. Heroku, on the other hand, has a lot of customers who are wondering how far their $50 a month is going to get them. For these customers, many of whom are Rails types who don't quite grok indexes, it might be a better solution to say "Um, look, rather than us teaching you a core engineering skill that you're manifestly unwilling to pay for, how about we suggest a technology stack that makes this skill unnecessary".
Previously, one of the Rails hosts (Dreamhost or Heroku, can't remember) released stats saying something like 97% of customers create no indexes. I totally understand how that can happen, too -- you expect ActiveRecord to be magic, and with what it does it is very powerful magic, but it is not magic that totally obviates your need to think about database design. (Edited to add: My business runs on Rails, I consider myself to have low to intermediate SQL ability, and if you contact my day job to get consulting on your database woes you won't get handed off to me anytime soon.)
Anyway, it doesn't get much simpler than: add_index :users, :account_id
Near trivial example: what index or indices do you need to support the business requirement "I want to know how long users stay active after they sign up, and I want you to be able to slice that data by signup date and by whether they're paying customers or not."
So programmer Bob goes off and does this.
"Oh, Bob, the screen only lets me slice the data by signup date and by customer type, but I want to slice by both at once."
So Bob makes a one line tweak to his controller (to use both conditions, instead of one or the other)... and BOOM, down goes the poor server.
But yeah. It's complicated. "Software is hard." But the best thing you can do is become aware of your ignorance, then try to eliminate what you can.
This is not true of CouchDB. Indexing is done on the keys generated by arbitrary Javascript views. A view returning results keyed by interest is trivial.
That isn't true. Well, it is with MySQL, but not with Oracle, and hasn't been for over a decade. As usual, most complaints about "SQL databases" (no-one who actually does databases calls them that) are really complaints specific to MySQL.
The problem you are going to run into, I know a company they wanted a particular database, so they got a designer turned programmer who built a system for them in Perl using flat files. His architecture was to put one data field per line in a file and then put the one-to-many records, comments about the master record in the file starting at line x. Well, now they have fields for the master item all the way to line x minus 2 and they know they want like 5 more fields.
If the developer who built the system had started on an SQL database, then the solution to the problem of needing more fields would be really simple.
Now this problem is just one issue. Add to this the problem of needing to query the master records on particular fields. The programmer has to open up every single file and go to the particular line where the field exists and see what the value is. Of course as the file based database grows, it's really starting to impact performance.
The big problem here is that the original programmer didn't know SQL or how to manage a SQL database so he downloaded some Perl code (note, I absolutely LOVE perl, so don't be hating) that managed a dataset in perl and just went to town.
This is not the way to build a robust information based system.
SQL databases are a tool and files are a tool and CouchDB is a tool and a lot of people are going to read this biased stuff about noSQL and think, "hey, let's use couchdb, et al" and they only have 5 users on the system and, you know, maybe a million records and Access, MySQL, SQL Server Express, or any number of other free sql based systems would handle a problem like that just fine.
Instead, they're going to go write a ton of code to fit a key/value store to the problem and it's going to be a nightmare to support as the system grows and they need to implement security and manage more fields.
But the pro-noSQL people don't talk about all those issues that are going to arise later and focus solely on the scalability issue when in the vast majority of cases scalability is really the last thing you need to think about.
You gotta kinda wrap your mind around SQL though and if you don't have a support group helping you do that, you're probably going to struggle with it. So I can understand the desire to find a simpler alternative to solve information based systems, something between file based Perl and SQL.
Map/reduce, though, that will do your head in! It's like changing from procedural to functional programming. To me, it's much harder and far less intuitive. So I can't imagine how anyone could think using M/R is somehow easier than learning SQL. Anyone who thinks that is in for a nasty surprise. It's like claiming Erlang is easier to learn than Ruby.
I also remember when I first started with sql that I could read and understand (most of) it quite easily. Writing it was of course another story, especially if optimization mattered, but it wasn't hard to read and not too hard to learn at least the basics. At this point the market is filled with books like "SQL for Dummies" and "The Manga Guide to Databases" that provide very gentle introductions for complete beginners as well as numerous classes at practical trade schools.
Mastering SQL and RDBMS is definitely a work for years and creates dedicated specialists just like any deep field, but picking up the basics is relatively easy compared to many other tasks in this field.
I agree that relational databases are often applied to problems where other systems would be better, and we should work on studying those other systems, improving their implementations, and educating engineers about their use. But we should be "pro DHT" or "pro document store" or "pro column store", not "anti SQL."
And it's getting a little old.
No solution works for everything all the time. Can we just stipulate that?
To state that the applicability of relational databases is narrow is egregious. Any one of the list of reasons could be a reason to use a relational database, it doesn't require a significant portion at the author seems to imply. There are other reasons too... like how error prone it is to write raw map-reduce code to generate a query that would be trivial, fast and guaranteed correct in any relational DB. And I'm thoroughly sick of the boogeyman of scalability as well. You can go up to the scale of 99.99% of web applications with relatively little hassle.
The article acts as if data integrity is usually not important, but for the vast majority of applications that are out there in the wild and older than a 6-month startup the data is far more valuable than the code. For a huge subset of all data SQL databases represent it the most accurately with adequate performance.
I don't care if the pendulum has swung too far towards SQL--this type of article is intellectually dishonest and won't help anybody make better decisions, it's just more hype.
Now you have a bunch more stuff in the mix. Stuff to install, stuff to configure, secure, periodically patch or update, backup, etc.
SQL databases don't scale? Maybe not in the extreme, but for the majority (vast majority) of web applications they will do just fine if they are used intelligently.
1) Action caching. I stick it in memcached. This gets me free expiry-by-time, free blow-away-cache-when-deploying-new-version, and saves me having to create and maintain a table specific to caching in the database. (I really, really hate changing my schema, since I invariably manage to screw something up when migrating. Solution: don't, absent a clear need to.)
2. Caching expensive queries. Again, memcached. Much of my analytics data is publicly accessible, changes infrequently (how often am I going to add a new sale to March 2007, honestly?), and lets users calculate very expensive summary statistics. This screams "cache me!" rather than letting somebody burn my server to the ground just by opening up /stats and clicking like mad.
3. Short-term scratchpad. I have a couple of places where I need to remember little bits of information for a few minutes.
For example, a lot of customers have issues with entering registration keys, so I try to make this automatic for as many as possible. One way to do this is to have the application generate a random ID for itself, and submit this random ID to my web page when requesting the sales page. The sales page then passes this to my offsite payment processor, who will in turn pass it back with successful transactions. Meanwhile, the application polls the server periodically to see whether their ID has been associated with a purchase recently. Sure, I could stick that in the Customer table, but it feels unclean: logically, it isn't a business requirement of customers, it is just a little hack I use to make their lives easier, and it has no meaning outside of an interval which lasts, on average, about 30 seconds.
The database is for searching, indexing, and ensuring that transactions are atomic. The cacheing is a different layer.
This is what the author says about an e-commerce order: "But in practice, and the app almost always wants to retrieve the entire record all at one, and write the entire record all at once."
Right, because no business ever wants a total of their orders in the month?!? Or a business never wants to know who their best customers are? This is the worst example ever and proves that this "in practice" thought is entirely backwards.
Careful analyses are useful (and I've seen some well-reasoned debates in HN) but what strikes me is that clear-cut cases for using one tool vs another are rare, for web-dev back ends at least. Witness: the large-scale sites that use various imperfect db solutions. Yes, they sometimes involve some "hacks" to the canned db solutions, but who's found the perfect swiss-army-knife that needs no modification for any project.
What does strike me as significant is the mastery level of the db developer (whatever the db on which they have the greatest mastery/comfort) ...because projects sometimes "go south" or end up getting late-stage additions of unforeseen complexity.
Conclusion: I'm a db agnostic. Or maybe "poly-theist." There are several good competitors. None is perfect. Several are powerful, especially in the hands of an insightful and experienced db hacker. ( IMHO )
This is starting to look like an N queries type problem if I want to print out an receipt that includes the name of the product. I'll have to grab all of the product codes out of the order and iterate over the list to separately query for each of the n products to get its name. A join would be faster. And what happens when a product changes its code? Where's my surrogate primary key?
I also don't see where he gets the infrequently-written part of this: "If we restrict our use of SQL database to only what they are good at - small, infrequently-written, long-lived records with complex data relations - and start storing all our other data in more appropriate places, we won’t need to push our SQL databases to such high levels of performance or scale." A lot of frequently updated data is better off in a relational, or at least normalized, store. If the price of a stock is changing all of the time, do I want to keep updating it in everyone's portfolio document? If you denormalize, you turn a single update into a multiple document update, which turns into locking a bunch of documents or records which might be waiting for another multiple document update.
Now, if all of your writes are inserts, it's a little easier to live in a denormalized world.
Then he goes on to mention that e-commerce orders are examples of things where 100% data integrity isn't needed.
I find that shocking. An ACID-compliant DB engine is essential backup for when customers manage to break your application in ways you never even imagined.
While the article has an agenda, it seems useful to understand why a database might be needed and what the useful alternatives might be.
That's why a crap-load of web sites and apps can all use the same database software.
You will ALWAYS have much better performance if you write your own "database" that specifically does what it is you need doing, as in the case of Google, etc., that you mention.
And don't forget, Google didn't just write their own database, they wrote their own file system, etc., and even designed their own hardware (batteries on board to eliminate large UPS's, etc).
But they can afford to spend the money to hire the really smart people to build and maintain that stuff for them, and they have the scaling requirements that justify it.
Most other companies either can't justify it, or don't want to... they'd rather rely on the vendor to maintain the DB code, to security testing/updates, etc.
Really, I think a lot of the complaints about SQL (especially in the case of these recent articles) come from the fact that it's hard to design and implement a system that works the way they want it to, and there's no tool that just plugs into what they have that does what they need.
I see their issues as the exception rather than the rule, and not really applicable to most situations.
$0.02
The data set is pretty small, though - the id for this comment is just under 700,000, for instance. While it's nice to see that one server running an interpreter can hold up to so much traffic (and this is much more traffic than many people will ever get!), it's not really representative of the cases where worrying about scaling actually becomes necessary.
It just seems to me there is so much to worry about when working with the FS directly. Like what happens when the app crashes in the middle of a write? Maybe it is not an issue anymore with modern, journaling file systems, but there are probably some issues left?
I agree that a real database would still be useful, though at that size it would just be for ACID transactions. OTOH, writing everything using native data structures makes prototyping a lot easier. Lisp sexps, Lua tables, Python dicts, etc. are fantastic for prototyping.
Yes, it's less work. The introduction of a databases nearly always more than doubles the complexity of a program and the amount of code written. There's a reason developers like in memory prototypes and persistent hash tables, they require less code than any other solution and don't force you to tear apart your model and try and stuff it into tables and put it back together again on every page view.
(If you're going to the database and back for every page view, you could probably be using caching more effectively, e.g. memcached.)
The issue is whether your data fits well into tables and rows, or not, and whether it's worth cramming it into tables and rows and continually reconstructing it to get the benefits relational db's generally provide over the alternatives... which are enforced data integrity, language neutrality, and a generic fixed structure capable of allowing flexible querying. To get those, you trade speed and ease of development.
Have a nice weekend.
Of course transactions and certain types of constraints would become impossible (as key-value stores do not provide for atomicity), but that could be an acceptable trade off.
Atomicity: save a journal of the pre and post values of any rows, and on commit take all of your staged changes and write them and then delete the journal. If you're ever going to hit the DB again and you see a journal file that wasn't cleaned up, roll back to whatever pre values there are.
Isolation: use a semaphore to cause the start of a writing transaction to block all reads and writes, use a semaphore to cause the start of a reading transaction to block all writes, and escalate from reading transaction to writing transaction when necessary.
For example: http://sqlite.org/atomiccommit.html
Also, it will slow down the database a bit (particularly if added to the db as an afterthought!), which seems to really irritate people who would rather have potentially corrupt information as soon as possible rather than automatic guarantees of internal consistency.
I wish I could have read that sentence first, instead of last.
Relational databases are just persisted forms of cross-linked lists. As if they're going anywhere anytime soon.
| The database does not need round-the-clock availability (middle-of-the-night maintenance windows are no problem).
Lots of SQL databases are part of 24/7 critical infrastructure.
| Queries do not need to be particularly fast, they just need to return correct results eventually.
Queries can be very fast if you predefine the correct indexes to suit the particular query. Only ad-hoc queries that you didn't plan for will be slow (sometimes). As it would be on any type of database.
Not true. If that were so denormalization wouldn't be a standard technique to speed up queries that join many tables and have to be really fast. The fact is SQL databases force you to reconstruct your document from many tables in a query and that is costly, even with indexes.
It's simply much faster to not break the data up in the first place to fit it into a tables and rows data structure if speed is your primary concern. Storing a complex document, as is, will result in much faster retrieval.
RDBM's are a tool, a damn fine tool, but like any tool, they're aren't the answer to every solution and far too many people pull out the RDBMS as the answer to every problem.
Of course they aren't the answer to every problem, and I think unstructured or semi-structured data (such as HTML pages or text data with entity tags) is a good case for storing things in a more document oriented system.
Relational databases are always are good solution; however, they are not always the best solution for the problem at hand. The real problem is that once someone has their head wired to an RDBMS, it's often impossible to get them to even consider another solution. It becomes their golden hammer.
The articles main point, that "SQL Databases Are An Overapplied Solution", is absolutely correct.
That is a blatant myth. Joins are not slow.
Now older versions of MySQL, like 3.23, are still in use (unfortunately) and pretty crappy but it was released over 8 years ago. Most of these performance myths about MySQL are outdated.
Are there any improvements in this area, or is it still just the same old, same old?
The bill of goods that RDBM's sell is that data integrity trumps all else. The reality is that is doesn't. Data can be cleaned and integrity maintained by background processes if the performance you get is more important to you than absolute up to the second data integrity. There are lots of cases where available data is far more important than up to the second accurate data.
Eventual consistency is more than good enough for many many things common web applications do. In many cases it's perfectly fine that the web app is looking at 4 or 5 hour old data, it just doesn't matter, what matters is that it can look at it in milliseconds without having to rebuild it on every query.
In other words, it's premature optimization.
Most people aren't building banks, for most apps data corruption is not the big concern RDBMS fans make it out to be.
Yeah, this is idiotic. This kind of thinking is why it sometimes took three days to register for classes in college, and why I couldn't check my account balance in the evening after I got home from work/class, why I sometimes couldn't search the library catalog... etc., etc.
I can think of perhaps one of our clients who would be fine with their database being down for hours at a time at night. Maybe.
"Queries [in a relational database] do not need to be particularly fast, they just need to return correct results eventually."
I stopped reading at that point. The article is symptomatic of everything that is wrong with this profession.
ObRelational: For people who want to learn about the relational model, two of the better presentations I've read are:
1. The Definitive Guide to SQLite by Mike Owens -- short, non-academic, theoretical enough to be useful. Focuses on SQLite, which is nice, because that way you don't take MySQL's flaws for granted.
2. An Introduction to Database Systems by C. J. Date -- academically thorough; it's a textbook, so the previous edition is pennies on the dollar. Mostly covers the relational model itself, rather than SQL, which is somewhat like comparing the pure untyped lambda calculus to Scheme. Once you understand the big ideas, though, everything else is just syntax and performance tuning (which are often implementation-specific anyway).