I agree largely with this post. I'm tired of hearing the same arguments the author has. This technology we use is hardly infallible and as programmers and system administrators of these systems the defects should be obvious! Yet somehow people still think that the "clueless users" are to blame for these "defects."
Sure, we all know the saying about assumptions. But why are we at a point where it's cool to use a data persistance layer that assumes it doesn't have to live up to any of those assumptions? It's why we have things like ACID, the LSB, etc. Contracts about how these systems should work for the end user.
It's hard to get all that stuff right off the bat, sure. But losing the entire database when a few bytes land in the wrong place? Yikes. Hardly something I'd trust real data to.
This stuff can always get better. Of course the solution as pointed out is tried and true. Glad to see that the developers aren't covering their ears...
I read once that every developer ought to use 'kill -9' as their method of stopping services they are writing, then ensure that doing so never causes a problem. Even if you can keep people from executing 'kill -9' (hint: you can't), there are always cases that you can't control that can stop a box in its tracks. By developing explicitly for this, you save you and your customer a lot of pain.
Honestly, I would never use a database that doesn't make safeguarding my data as its number one priority. Even if I'm just dumping logs in it, the time I really need to know what's going on is likely the time those safeguards are put to the test.
It was briefly mentioned in the article, but this is exactly how couchdb shuts down:
On-disk, CouchDB never overwrites committed data or associated structures, ensuring the database file is always in a consistent state. This is a “crash-only” design where the CouchDB server does not go through a shut down process, it’s simply terminated.
This is exciting. MongoDB (and the suite of libraries building up around it) has made prototyping web applications an order of magnitude easier than using SQL. Lack of single-server durability, however, is a showstopper in production, before you've grown enough to justify scaling the database beyond one machine.
In my case, that meant going back to MySQL once our schema was finalized (sadly). Looks like that won't be necessary in the not-so-near future, which means MongoDB is usable in new projects without worrying about replication before otherwise necessary.
I have an app I'm working on launching right now, it's in what you might call 'private beta': real people accessing the app is <10. It's been in this state for a little less than 2 months and there's been a replica-set and backups happening since even before that.
What I'm saying is that you don't need to grow before adding a second box. My project currently runs on ec2 micro instances and the second box will probably cost you less than your GitHub code hosting costs! Just seems to me that (in ec2 terms, as an example) going from $9/month burn to $18/month is hardly a show stopper.
Aren't micro instances $15/month? (Plus EBS backing, unless you want to lose your DB on reboot.)
Regardless, it's added complexity that you can avoid with MySQL early in your application's lifecycle. Even if it's not much more expensive, you have more hardware to administer. Compare that to a Rails+MySQL stack, which you can run on a single box. (There are added benefits of replicas, sure, but they aren't easier to set up, per se, then just installing your stack on one machine when you're getting going.)
Yea sorry $15/month (EBS is practically free, as is Elastic IP), but anyway my point still stands at $30/month. I think you'd be hard pressed to find a developer who doesn't spend that much on coffee per month (not to mention merchant accounts and support and code hosting and bug tracking and blah blah blah) :)
As to the rest, sure, it's sys-admin-wise easier to setup. But then you have to consider stuff like database dumps/backups. Where are you storing them? Have you tested your recovery plan? Then there's hassle when you decide you need a dedicated db box and move it there (friday fat-finger, anyone?). True enough that none of these are deal-breaker level issues (evidenced by the ubiquity of MySQL/etc).
With mongo you spend a tiny bit of extra money and a tiny bit of extra time setting up a replica set which handles your backup for you. And then when you need more db horsepower you just start sharding or just adding more replicas and sending reads there. Need to upgrade the db box? Cool just add a huge instance, add it to the replica set and promote it when it's done syncing.
I certainly agree that it's a trade-off... it's just one that I think is not driven by data durability. Either way I need to handle my disaster recovery with some kind of automated system and I'm just suggesting that whether this is some kind of cron-job, restore recipe, etc backup of a db-dump or a replica on a cheap extra box is a wash.
In my experience, using CouchDB for prototyping is not fun. With SQL and (from what I gather) MongoDB, you can create ad-hoc queries as you go. Creating views in CouchDB requires thought and time -- usually lots of time -- to generate the views, unless you're working with a very small amount of data. Changing your mind is painful. You end up using weird hacks to get around rebuilding your views all the time.
Maybe I was doing it wrong, but it seems like if your app has more than a little data and you plan on changing your mind about how best to present it, you will experience discomfort.
I don't have any experience with MongoDB, but I am building an application right now using CouchDB and I've found it to be pretty easy to use -- definitely easier to prototype with vs a relational database + ORM framework.
CouchDB supports ad-hoc queries in the form of 'temporary views'. The downside (and I think a difference compared with MongoDB) is that temporary views can't be used in production since they are not indexed and thus are much slower than permanent views.
I'm developing my application in Python, which has great support for CouchDB in the form of CouchDB-Python (http://packages.python.org/CouchDB/) and CouchDBKit (http://couchdbkit.org/). Using CouchDBKit, it was pretty easy to set up CouchDB artifacts (map functions, reduce functions, design documents) into a nice file/folder hierarchy and write the simple Python glue code to deploy updates to CouchDB via a single shell command.
so far we're going with couch and haven't experienced this. the best part of course is that we can add items to documents on the fly whilst keeping the existing test data! this is major for us. we can go back to testing right away. with MySQL the changes were so major that it was better just dumping and recreating from scratch (we are using django) and this wasn't exactly fast. django has a project "south" that would help with this, but we're eventually deploying to oracle and cannot rely on it or have the time to extend it to properly work with oracle.
I'm pleased Mongo is getting single server durability. I have never understood why it got so popular without this feature. I'd love to know why people choose Mongo over say, Riak, or CouchDB, as the majority of projects don't need more than one server.
Because they market infinite scalability, insane speeds, and they have a nice API. But people don't realize that, while the speed is fast, you're driving without a seatbelt, and the scalability story is more or less false. They do have a nice API though.
I use mongodb to record fine grained log for performance tests, where it works like a charm. It is fragile (in that I have lost full databases when VM were full, for example)
I don't know about riak, but we are forced to use couchdb at work, for the wrong reasons I think, and I was not impressed. First, its performances are pretty bad: document insertion is slow unless you batch them, but even though it is slower than in mysql, even though the guarantees about durability are certainly not the same (couchdb was not configured to fsync at each write). Building views is excruciatingly slow, and the view engine does not use all the CPU available (even when IO wait is low, so not a simple IO issue). There are bug reports about this issue (cannot find it at the moment). Replication is not very reliable either for large data (where large against means a few GB, so actually not that large).
I am sure we are using couchdb the wrong way at work (out of my hands), and for the wrong application, but its design choices as well as my limited experience does seem to imply couchdb is not adapted for large amount of data, especially one which are written often (there was an interview last summer with D. Katz who said that there were not yet much optimization for large data).
There are some companies which seem to use couchdb for large usage, for example bbc (http://enda.squarespace.com/tech/2010/3/4/couchdb-at-scale-4...). I don't know their infrastructure, but they claim to server 4 billions requests as of 4th march 2010 since summer 2009 on a 32 nodes (16 master, 16 backups). Assuming that summer starts in september to get an upper bound of the traffic, this means 250 rq/sec on average, which is nothing impressive for an infrastructure with 32 machines without more information about what they do. Generally, I would not say much about this kind of usecases, but since it is often advertised by couchdb proponents, the burden of the proof is theirs.
I agree we need to improve the interactive query capability of CouchDB.
As far as performance, the key with Couch is to keep the view generation speed faster than the insert rate. So if your users are generating less than about 1000 changes a second, you should be fine on a single server.
More than that and you may need to shard / partition, which can be done in a few different way. The leading option is BigCouch: https://github.com/cloudant/bigcouch
Actually, the numbers I get are much lower than that (almost two orders of magnitude), but I suspect it is at least partly due to bad usage of couchdb.
I am actually working on simple benchmarks representative of what we do to see if there is something worth submitting as bug issues.
You just have to tilt your head a bit and realize that you do need more than one server. For example - I don't know about you but I don't want to take down the site to deploy everytime I have to update some code (esp in the early days, this happens practically once a day).
With MongoDB you simply add a cheap second box and use it as at least a replica of your database. When you get tired of "we're updating the site with shiny new code" interrupting your users, you simply make it into a web-node as well.
A city could be destroyed by a nuclear bomb at any time - that doesn't mean components of a system shouldn't be reasonably durable. Often it's impractical or unnecessary to have multiple servers for a proof-of-concept, a staging/test server, etc., yet it would be nice to not have to start from scratch, or to deal with backups, just because the database is so fragile. Basically, having a non-durable system forces you to take unnecessary precautions on non-critical systems because of the higher likelihood of failure and lost time investment.
We're talking about data durability. When my SQL server crashes due to hardware failure I know that when I eventually get it back up and running the data will be consistent with at worst the last couple of transactions being rolled back.
What I'm saying is that I can go up to my SQL server and disconnect the power cord and my database will not be corrupt when I start it back up. Sure if your HDD gets taken out by a meteor then nothing will save you but that's why you have backups.
Not guaranteed. I have had more than one customer experience hardware failure, resulting in a corrupt or suspect SQL Server database, that was unrecoverable via normal means.
In each of the cases where the customer had a true standby system, implemented via replication, log shipping, or mirroring, they were able to failover with little (log shipping) or no data loss.
In the cases where they had a single, standalone server, the option was to restore the last known good backup, or sent the database files had to Microsoft for analysis and repair.
ANY system (RDBMS, NoSQL, or otherwise), should have a standby replica to prevent data loss. If you data is stored on a single machine, you are doing it wrong.
I'm not disagreeing with you, all I'm saying is that SQL Server databases are built from the ground up to resist data corruption and does exceedingly well at it. Not so with NoSQL data stores.
Yeah. There are several nosql backends which offer resistance to corruption under sudden failure. Bitcask, for example, is structured for on-disk data durability using some of the ideas for log-structured filesystems.
Personally, Mongo maps so nicely to Ruby that sometimes, I just use the regular driver rather than an ORM. A database that naturally understands hashes and arrays is lots of fun.
Also, it seems like with Mongo, I can often design most of my pages to need just one query, and therefore they're crazy fast. This might be possible with other NoSQL stores too, and I'm trying to get more experience with more of them.
In my admittedly limited experience with mongo, it felt like they were designing the ultimate database for the web. Not in a "web-scale" sort of way, more in a swiss army database for the web dev way. This came to me when I wanted to do location queries on a db. Most people using standard DBs do this with PostGIS for postgres, which is incredibly powerful and accurate. But mongoDB supports the "find shit near me with good enough accuracy" query that 90% of people want to run right now.
I was able to make a simple test page that told you the IP of the previous visitor who was closest to you. It took me thirty minutes from "hey mongo has some sort of geo support" to that (and I'm not a web dev and was using ATT 3G to read doc). PostGIS took me longer to figure out how to set a lat,long pair in the DB.
I'm not saying raw speed of development is the best measure of a database, but if your #1 priority is the ability to rapidly prototype, I'd imagine mongo is a good fit in your toolbelt.
I assure you that that anyone who has used any kind of SQL database can do the same stuff you did on Mongo in the same time or less. Development speed is much faster on the SQL database once you have some more data and you can start using joins and searches and all the other nice RDBMS features.
I've built a lot of stuff on RDBMSen, and often, those 'nice features' are actually what I call 'a giant pain in the ass.'
Sometimes, data doesn't fit a relational model well. In those cases, something like Mongo is a godsend. And yeah, you could de-normalize your data and get halfway to Mongo, but I'd rather use each tool for what it's good at.
We have exactly the same kind of data relationships in our database. Sure it's complicated and in your example sure you'll end up with a bunch of tables if you normalize it all the way to 3NF. But there's a lot of GOOD STUFF you get when you do, for example if you decide to rename a FooType or a FooSize then you only have to update one record rather than searching every document in your document DB for instances of those names so you can rename them. Hell if you have many documents and 24/7 clients this may become IMPOSSIBLE to do while still guaranteeing consistency in your data.
Your example is not that unusual or even that complicated. What is about it that doesn't work on a RDBMS? You can't just say the solution sucks because it uses more tables that you feel it should.
It's good to know that I'm not totally spewing bullshit, and that's the way it should be done.
In this case, I would have less than 200 types, and probably under 1000 Foo records total. Updates wouldn't have been a problem. The pain was much bigger than the upside. Especially designing the forms that would end up creating those rows...
It's not that it doesn't work. It's that it's not a 'this always fits' solution.
This is what views are for. If it is easier to work with the data denormalized, create a view with a few joins. Some views can even have updates directly applied to them.
If the amount of data you're storing is so small you should definitely consider de-normalizing it until it's easy enough to work with. When you eventually scale up to millions of Foos you'll probably want to get it back to 3NF.
There is nothing wrong with denormalized data in your DB and denormalizing doesn't mean you'd have been better off with some random NoSQL solution. Especially considering that the entire rest of your DB is still very much relational data so what happens to all of that when you dump Postgre for Mongo?
It's not really random, though. I made a comment elsewhere in this thread, but Mongo is _almost_ a freaking Object database as far as Ruby is concerned. I could do it either way, but doing it with an RDBMs feels like BDSM compared to the loose and fast feel of developing with Mongo.
That said, I always try to use what's right. I don't care how easy it is to store credit card numbers is... I'll just use BrainTree's Vault product instead. ;)
Just like most things, it's not that simple. For example...
Cassandra is good for 'big data' if you're ok dealing with the repercussions of eventual consistency, and questionable performance characteristics. Not to mention the mediocre community support.
Redis is a fantastic choice, if you're ok with a limited data model and no diskstore option(coming soon though!)
Picking a datastore that fits your data model, access patterns, support needs, etc. can be a daunting task, and simplifying it to this degree just leaves too much out of the equation.
I don't know if this is what happened to the reluctant hero of this story, but if the database really had hung, and wasn't responding to "sensible" methods of shutting down, what are you supposed to do besides kill -9? (By "sensible" I mean something the DB could trap in order to finish saving a consistent state to disk - shutdown command, -INT, -QUIT, etc.)
(I'm not excusing trying -9 as a first resort, but if a process is deadlocked (or pegging the CPU in a tight loop) and won't respond to -QUIT, there's not much else you can do.)
Actually that doesn't add any additional durability guarantees in the general case. The only way to use MongoDB durably with a single server is to use the new --dur flag coming in 1.8. I assure you there is no other way.
Sure, we all know the saying about assumptions. But why are we at a point where it's cool to use a data persistance layer that assumes it doesn't have to live up to any of those assumptions? It's why we have things like ACID, the LSB, etc. Contracts about how these systems should work for the end user.
It's hard to get all that stuff right off the bat, sure. But losing the entire database when a few bytes land in the wrong place? Yikes. Hardly something I'd trust real data to.
This stuff can always get better. Of course the solution as pointed out is tried and true. Glad to see that the developers aren't covering their ears...