SQL Databases Are An Overapplied Solution

http://adam.blog.heroku.com/past/2009/7/8/sql_databases_are_an_overapplied_solution_and_what_to_use_instead/

by fogus • 17 years ago

83 72 17 years ago

adam.blog.heroku.com

ams6110• 17 years ago

I'm not convinced by this. SQL databases have proven over the years (decades) to be a very good choice for storing data. The author ticks off a litany of other products (document databases, cloud storage, memory object caches, message queues...) to consider instead.

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.

brown9-2• 17 years ago

I don't think that the author's point has to do with performance, but that a SQL database isn't always the right tool for the job and shouldn't be the default answer all of the time.

patio113• 17 years ago

Right. I love SQL, don't get me wrong, but there are a couple things my site does which SQL is just a suboptimal solution for:

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.

wvenable3• 17 years ago

While a SQL database isn't always the right tool for the job, it most definitely is the best default answer most of the time. We keep seeing these articles over and over and they keep missing the point.

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.

Devilboy• 17 years ago

I don't agree with these two points:

| 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.

gnaritas7• 17 years ago

> Only ad-hoc queries that you didn't plan for will be slow

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.

mattmcknight2• 17 years ago

Except you just showed that normalization has performance tradeoffs, not that relational databases aren't a good solution. They offer you the flexibility to normalize or denormalize as the conditions demand.

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.

pavel_lishin• 17 years ago

> The database does not need round-the-clock availability (middle-of-the-night maintenance windows are no problem).

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.

lsb2• 17 years ago

SQLite version 1 was implemented on top of a gdbm backend. They ran SQL queries on top of a key-value store. I wonder if Tokyo Tyrant, which has the goal of being a modern implementation of dbm, would make a good backend.

strlen2• 17 years ago

It would be interesting to see a SQL parser (a front-end, basically) on top of a distributed, HA key-value store. It could even be a pluggable MySQL engine (there already is a Memcached engine).

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.

dasil0033• 17 years ago

If you're going to use SQL why would you use such a limited backend with a parity mismatch and a much greater set of limitations in place of many heavily-optimized-over-decades backends?

lsb2• 17 years ago

K/V stores don't provide ACID properties themselves, but you could bolt them on pretty easily. For instance:

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.

mdasen• 17 years ago

Relational database alternatives aren't as applicable as they make them sound. They've provided some examples. The first one is a social networking profile with a list of interests. "Awesome! I can just store things in a JSON-like list and not have another table or anything!" Oh, what if I want to find all the users who have 'ruby' as an interest? Oh, I can only lookup by key? And yes, there are ways around that. You could create a table of interests and each interest would be a key in that table and it would have a list of people with that interest. So, you have 'Joe' => 'interests': 'ruby', 'blokus'; 'Amanda' => 'interests': 'ruby', 'rugby'. And then you have an interests table with 'ruby': ['Joe', 'Amanda']. And when you want to search by interest, you select the document from the interests table and get the list of people and then request the documents for those people. But is that more efficient? No. It both requires more code and takes longer to execute. A RDBMS is able to optimize a join on that data in a way that you can't.

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.

patio113• 17 years ago

Oh, also, offer some consulting for clients on their database woes.

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.)

cscotta2• 17 years ago

If there are Rails developers who work with applications of any size at all and are not familiar with indexes, the problem isn't scaling - it's lack of knowledge of one's application's stack. The mentality that one's app should magically scale without any idea of what's under the hood is toxic.

Anyway, it doesn't get much simpler than: add_index :users, :account_id

nostrademons2• 17 years ago

Could be that 97% of customers have no need for indexes. I have none on my personal site; with rows numbering in the dozens and pageviews/day numbering about 20, I could probably brute-force search over the database in Python and it'd still be fast enough...

jules• 17 years ago

I have some sites that don't have indexes. At <1000 views/day it's not necessary. Caching generated HTML is as easy and more effective in making it faster.

sah• 17 years ago

"Awesome! I can just store things in a JSON-like list and not have another table or anything!" Oh, what if I want to find all the users who have 'ruby' as an interest? Oh, I can only lookup by key? And yes, there are ways around that. You could create a table of interests and each interest would be a key in that table and it would have a list of people with that interest.

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.

gaius• 17 years ago

the writes have to be done on every box while a read only has to occur on one box

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.

pj3• 17 years ago

These heroku guys are selling something. They're on a bandwagon, constantly bashing SQL databases. I'd say SQL Databases are great for 90% of information based systems.

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.

sho3• 17 years ago

While I agree 100% with your point that people need to think carefully about their actual needs and not just follow the hype of the moment, comparing "NoSQL" databases to someone writing flat files in Perl is a bit harsh! Using SQL these days, with modern ORMs, is if anything quite a lot easier than manipulating flat files. And that's the problem! SQL is the new "flat file", the hammer every programmer knows, so you know what every problem starts to look like.

pj3• 17 years ago

Yes, that was a bit harsh, but the point I'm trying to make is that the noSQL proponents aren't really talking about PRO-couchDB, they just keep hammering on how "bad" SQL is, when it's really frickin' awesome.

DanielBMarkham2• 17 years ago

These heroku guys are selling something. They're on a bandwagon, constantly bashing SQL databases.

And it's getting a little old.

No solution works for everything all the time. Can we just stipulate that?

mattmcknight2• 17 years ago

His example raises a question... 'items': [ {'product_code':'AX5718','qty':'1','price_per_unit':'5.00'}, {'product_code':'BB9388','qty':'3','price_per_unit':'2.00'}, ],

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.

OperaLover• 17 years ago

Does the OP believe we really need to have "database wars" to match the tired old "OS wars" and "browser wars"?

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 )

HoneyAndSilicon• 17 years ago

On the upside: even if the "wars" are a tad artificial... it does provide db noobs (like me) a chance to see the issues (at least when processed in the analytical way they are here at HN).

dmfdmf• 17 years ago

What does HN run on?

listic• 17 years ago

Software: Arc

http://arclanguage.org/

Hardware:

http://news.ycombinator.com/item?id=516122

burke• 17 years ago

afaik, everything is stored in memory. I might be wrong though.

silentbicycle10• 17 years ago

The dataset is stored in memory on a server written in Arc, in turn running on top of MzScheme. (The source comes bundled with Arc.) It occasionally writes changes to disk, and lazily reloads them when restarted.

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.

Tichy2• 17 years ago

While it is cool, I wonder if that really was less work than simply storing things in a db. Presumably with a good ORM mapper, caching the data in memory transparently should only be a small configuration setting.

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?

dasil0033• 17 years ago

I really can't stand the hype around these document DBs anymore. Sure SQL database are overapplied. I realize many people are incompetent with them, and that they often don't give appropriate thoughts to alternatives. The title is fair, but then it veers off very quickly into hyperbole and misrepresentation.

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.

mtts• 17 years ago

The guy quotes "Data integrity is 100% paramount, trumping all other concerns, such as performance and scalability." as somewhat of a disadvantage of SQL databases.

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.

rleisti• 17 years ago

CouchDB is ACID compliant

DanielBMarkham2• 17 years ago

"SQL databases don't scale, but that’s ok. "

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.

alphazero2• 17 years ago

"SQL Databases"? SQL is a language.

"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.

silentbicycle10• 17 years ago

When people call them "SQL Databases", it's often a sign that they don't understand the relational model. Of course, it's much harder to make a RDBMS perform well if they don't, just as code written without an understanding of algorithms is likely to be unnecessarily slow.

alphazero2• 17 years ago

Agreed. (Prepare to be downvoted for making a rational statement contrary to the prevailing fads.)

Tichy2• 17 years ago

Interesting read, but I am doubtful about some aspects. For example "SQL Blobs are the wrong choice for binary data" - wouldn't that depend entirely on how the SQL Blobs are implemented? Sure, maybe apache can serve faster from the file system than from an SQL Blob (if you read the SQL Blob in the middle layer and stream it). But that is just one possible implementation. In theory, the web server could be part of the database and stream the SQL Blobs directly. SQL Blobs could be implemented as files in the file system. And so on...

wglb• 17 years ago

I am tempted to say "real men don't use databases", but that is a little dramatic. Databases in the sense of RDBMS/SQL databases. It seems that HN doesn't, Google doesn't, Yahoo doesn't, ITA does (from what I understand). (Or at least they are using a lot of Oracle--maybe that is stuff that is much more than just the RDBMS.) Is your problem as big or as hard as the problem that these outfits solve? Isn't it an interesting matter for contemplation why these smart folk aren't?

While the article has an agenda, it seems useful to understand why a database might be needed and what the useful alternatives might be.

nettdata2• 17 years ago

A lot of databases "suck" because they have to be jacks of all trades, in that they have to be able to do a ton of very general things.

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

SQL Databases Are An Overapplied Solution

Related Stories