Slashdot Mirror
Ask Slashdot: Which NoSQL Database For New Project?
DorianGre writes: "I'm working on a new independent project. It involves iPhones and Android phones talking to PHP (Symfony) or Ruby/Rails. Each incoming call will be a data element POST, and I would like to simply write that into the database for later use. I'll need to be able to pull by date or by a number of key fields, as well as do trend reporting over time on the totals of a few fields. I would like to start with a NoSQL solution for scaling, and ideally it would be dead simple if possible. I've been looking at MongoDB, Couchbase, Cassandra/Hadoop and others. What do you recommend? What problems have you run into with the ones you've tried?"
Do you need a database to do what you're trying to do? Why not just write the information to a text file (csv or tab seperated?), and use other programs to query the data?
try to make ends meet, you're a slave to money, then you die
If you need to store less than a few hundred million rows just use PostgreSQL.
It supports JSON and transactions.
You might want to consider a SQL database.
Based on your information no one can give you solid advice. It highly depends on the load you expect and on the data model you will use. for a simple twitter, you can use a log file, or any NoSQL technology. If you only have a few transactions and not billions of entries, you could use PostgreSQL or even MySQL. However, PostgreSQL scales better. If you want to make complex interpretations on graph like data you may consider Neo4J as a graph DB.
I would like to start with a NoSQL solution for scaling
And there it is, the proverbial premature optimization ...
These guys are committed, meaning mongo has a future. 2.6 that came out the other day has some nice new features and many bug fixes.
SQLite is a relational database management system contained in a C programming library. In contrast to other database management systems, SQLite is not a separate process that is accessed from the client application, but an integral part of it.
"I'll need to be able to pull by date or by a number of key fields"
So, in other words, you have already decided on key fields. If you use a database, this has things call index's, that can search billions of rows for a key field in a fraction of a second.
If you don't use something with INDEX's then you can't do this.
Where has this idea that Databases can't scale come from? - The world runs on Database for heaven sake. Do you think when you take money out of an ATM, its going to MONGODB? - And yet there are millions of ATM's and you can take money out of your VISA account in almost all of them anywhere in the world. That is called scale.
I would consider using the latest release of MariaDB.
You can use it as a standard MySQL server, but they also have Cassandra NoSQL as an engine for it now (since the release of 10)... So you would be easily able to play with things on different database types and see what suits your situation better.
It's a mistake to think that "NoSQL" is a silver bullet for scalability. You can scale just fine using MySQL (FlockDB) or Postresgl if you know what you're doing. On the other, if you don't know what you're doing, NoSQL may create problems where you didn't have them.
An important advantage of NoSQL (which has its costs) is that it's schema-free. This can allow for more rapid iteration in your development cycle. It pays off to plan document structures carefully, but if you need to make changes at some point (or just want to experiment), you can handle it at the code level. You can also support older "schemas" if you plan accordingly: for example, adding a version tag or something similar that can tell your code how to handle it. So, even ignoring the dubious potential of better scalability, NoSQL can still be beneficial for your project.
More so than SQL, NoSQL database are designed for different kinds of applications, and have different strengths:
MongoDB is a really good backend engine that gives programmers lot of control over performance and its costs: if you need faster writes, you can allow for eventual integrity, or if you need faster reads, you can allow for data not being the absolute freshest. For many massive multiuser applications, not having immediately up-to-date data is a reasonable compromise. It also offers an excellent set of atomic operations, which from my experience compensate well for the lack of transactions. Furthermore, MongoDB is by far the most feature-rich of these, supporting aggregate queries and map-reduce, which again can make up for the lack of joins. It also offers good sharding tools, so if you do need to scale, you can. Again, I'll emphasize that you need a good understanding of how MongoDB works in order to properly scale. For example, map-reduce locks the database, so you don't want to rely on it too much. The bottom line is that MongoDB can offer similar features to SQL databases (though they work very differently), so it's good for first-timers.
Couchbase is very good at dispersed synchronization. For example, if parts of your database live in your clients (mobile applications come to mind), it does a terrific job at resynching itself and handling divergences. This is also "scalable," but in a quite different meaning of the term than in MongoDB.
I would also take a look at OrientDB: it's not quite as feature rich as MongoDB (and has no atomic operations), but it can work in schema-mode, and generally offers a great set of tools that can make it easy to migrate from SQL. It's query language, for example, looks a lot like SQL.
The above are all "document-oriented" databases, where you data is not opaque: the database actually does understand how your data is structured, and can allow for deep indexing and updating of your documents. Cassandra and REDIS (and Tokyo Cabinet, and BerkeleyDB) are key-value stores: much simpler databases offering fewer querying features: your data is simply a blob as far the engine is concerned. I would be less inclined to recommend them unless your use case is very specific. Where appropriate, of course simpler is better. With these kinds of databases, there are actually very few ways in which you can create an obstacle for scalability: simply because they don't do very much, from a programming perspective.
There are also in-between databases that are sometimes called "column-oriented": Google and Amazon's hosted big data services are both of this type. Your data is structured, but the structure is flat. Generally, I would prefer full-blown "document-oriented" databases, such as MongoDB and OrientDB. However, if you're using a hosted service, you might not have a choice.
It's also entirely possible to mix different kinds of databases. For example, use MongoDB for your complex data and use REDIS for a simple data store. I've even seen sophisticated deployments that very smartly archive data from one DB to another, and migrate it back again when necessary.
I just felt I have to comment on this. So many developers start with the phrase "I need NoSQL so I can scale" and almost all of them are wrong. The chances are your project will never ever ever scale to the kind of size where the NoSQL design decision will win. Its far more likely that NoSQL design choice will cause far more problems (performance etc), than the theoretical scaling issues.
Take for example two systems I've been involved with for managing WiFi access to large scale networks (100,000+ concurrent users, 1000's of APs), one uses MongoDB the other based on PostgresSql. The MongoDB based solution has very real performance problems, its reporting takes a very long time to run taking very large amounts of system ram (24G in some cases) and that performance is only degrading as the system grows, there are also many other performance issue. These issues are not just mongo issues but simply that NoSQL is not well suited to the task. The system has been rewritten using an SQL backend and now works much better but importantly it's scaling but better. Growth in the system is no-longer degrading performance and the point where we need hardware upgrades or extra servers etc are now much more predictable so we can predict cost base growth in relation to user growth.
NoSQL does not guarantee scaling, in many cases it scales worse than an SQL based solution. Workout what your scaling problems will be for your proposed application and workout when they will become a problem and will you ever reach that scale. Being on a bandwagon can be fun, but you would be in a better place if you really think through any potential scaling issues. NoSQL might be the right choice but in many places I've seen it in use it was the wrong choice, and it was chosen base on one developers faith that NoSQL scales better rather than think through the scaling issues.
Telecommunications data is eminently suitable to schema table storage in any relational database, which with a little work, will let you index by the keys you intend to query by.
NoSQL solutions are better for unstructured data that doesn't come in predictable formats or value sets.
You need to take a step back and look at the problem before you decide on a solution. Don't be one of those idiots who tries to use a hammer to drive a screw.
I do not fail; I succeed at finding out what does not work.
Now scale that. Or just lock it properly.
If you want simple, scalable and low sysadmin overhead and all you need are key -> value lookups then Amazon's S3 can be an excellent choice. You don't need to manage it, you don't need to work out how to add servers and its well proven at extremely large scales.
However, like a lot of other posters, I'm very sceptical that NoSQL is the place to start. SQL databases can do a LOT for you, are very robust and can scale very considerably. As your requirements grow you might find yourself wanting things like indexes, transactions, referential integrity, the ability to manually inspect and edit data using SQL and the ability to store and access more complex structures. You're likely to give yourself a lot of pain if you go straight for NoSQL, and even if you DO need to scale later combining existing SQL and new NoSQL data stores can be a useful way to go.
"I'm working on a new independent project. It will soon become the new Facebook, and I'll be billionaire next quarter. The only problem is that I don't know which luxury yacht to buy with all this money. I've been looking at Lady Moura, Christina O, Pelorus, Venus and others. What do you recommend? What problems have you run into with the ones you've tried?"
Premature Optimisation.
CouchBase/CouchDB is probably the easiest and most available one out there. It's particularly well suited for app backends too, as both the backend and mobile apps can talk to the same database, in theory eliminating the need for the backend to handle data syncing.
Those are good reasons, and it's also true that CouchDB will use a lot less resource overhead than a full-bore RDBMS under load. Depending on the use case, it might also prove decidedly easier to scale.
But the place where NoSQL really shines is storing amorphous or heterogeneous data. Because you have no constraints about what goes into a given record, you can record more or less name/value pairs at your whim. As with Perl, though, freedom comes at the cost of potential disorder.
But honestly, with the tiny amount of detail provided, it seems like it's really six of one and half a dozen of the other. If it's just call data being recorded, and the same call data every time, it won't make a huge difference if you use a full-blown RDBMS or a NoSQL database. Either one has its costs (individual PUTs and POSTs in CouchDB for example, can be expensive, whereas queuing and write contention might cause headaches at extreme scales in PostGres or Oracle).
Both an RDBMS and a NoSQL database will deal with replication fairly well, though my personal inclination is to prefer the simplicity of replication in CouchDB right up until the noise level gets out of hand.
Crumb's Corollary: Never bring a knife to a bun fight.
NoSQL is a good solution for horizontal scaling, CSV and SQL DB are not.
I'd like to dispute this. Based on the OP's description of his application, two things come to mind:
Blanket statements like yours above can't really be made without reference to the intended application, as some applications scale much more easily than others, and OP's sounds like it's one of the easy kind.
First. everyone who is pointing out your premature optimization is probably right. You can get a lot of scalability out of existing databases, particularly if you optimize your data schema with indexes. Even if you store all possible 9,999,999,999 phone numbers, the log base-2 of that is 34. So you'll need a b-tree 34 levels deep. That's big, real big, but b-trees are fast. Worst case you are reading 34 blocks from disk, which is ~16kB.
Next, don't choose databases by name. Choose them by their features because you use features, not names. That said, HBase is probably what you want. It's a blend of distributable hadoop and tables. You don't need atomicity (it doesn't sound like) which is one thing you give up when leaving SQL behind.
Slashdot's rate-of-post filter: Preventing you from posting too many great ideas at once.
...so that you simply write an adapter for pushing/pulling data.
Then you don't have to worry so much about making what appears to be an extremely premature optimization.
In other words, have your backend web services (presuming you're using them and not manually POSTing from a socket yourself to your own socket server) instantiate an instance of iMyDBAdapter and use it.
Later, when you find out that you actually do need MongoDB, PostgreSQL, sharded MariaDB, whatever, you can simply write another adapter class that simply has to satisfy the iMyDBAdapter interface.
The reason this works so well is that it will force you to separate your business logic from your underlying DB implementation (which requires a lot of discipline to do otherwise, especially when you just want to get something 'done'.)
Also, as another poster pointed out, you're much more likely to suffer from other issues relating to scaling (and issues better solved elsewhere) than a modern database.
My advice, stick rigidly to the interface/adapter mechanism and implement an adapter for whichever DB you're most comfortable with right now.
Loading...
I would like to start with a NoSQL solution for scaling,
This is a solution looking for a problem. Or more precisely, you are looking for an excuse to use a piece of technology or paradigm. Don't get me wrong, your systems requirements might indeed be best served using a NoSQL solution, but what exactly has your analysis shown regarding this?
Scaling is not just a technical feature (NoSQL, SQL, Jedi mind-meld tricks). Scaling is a function of your architecture. You can NoSQL the shit out of your solution, but if your software and system architecture is not scalable, then having NoSQL will mean chicken poop as solutions go.
and ideally it would be dead simple if possible.
If you want simple, put a simple RDBMs schema (a properly normalized that) in place, and have your code use a simple, technology-agnostic persistence layer that maps your domain-level artifacts to database artifacts. If you ever had to replace the back-end, then you can do so with minimal changes to the API that domain-level artifacts use to persist themselves with the persistence layer.
Design your domain solution around domain-specific artifacts. Persistence technology is typically a low-level design/implementation detail, an important one obviously (and a critical one for some classes of systems).
But for what you are describing, the choice shouldn't even be coming into the picture without first having an architectural notion of your solution.
It means you don't have any big data requirements so you're better off sticking with MySQL or something easier to manage at a small scale.
If growth is high or you have a lot of data to analyze, you can look into importing data into Hadoop using sqoop and query it with Hive and HBase. But you most likely won't need that for at least a couple of years.
and get to know it later :-). Fast here: your prototype creation, not primary the database I/O. The general comments are right: there is no one-fits-all solution and the database might change. It looks very much like you also haven't decided on the server platform: Ruby, PHP... you could look at node.js or vert.x too - server side JavaScript is at least neat for prototyping (I'm not making a statement that is is *only* neat for prototyping - that's a completely different discussion).
We did a number of super rapid prototypes with datasets roughly in the range you describe using CouchDB (not CouchBase!). There we took advantage of CouchApps - the ability to store the application itself inside the database - works like a charm when replicating data and you need a http server (Apache, NGix) for the URL mapping (which is already kind of optional) and CouchDB. You can authenticate with OAuth or via the Webserver and it replicates - so you can have local data easily (gold for testing). Since you can specify the direction I usually replicate all data from the server into local, but not the design. So I can try new app features local against the live dataset. It also does Map-Reduce using JavaScript.
Give it a shot. If it can handle the data from CERN you also have quite a growth path. One fun project we did: run it on a Rasberry Pi to collect weather data from Arduinos all mounted in a small sail boat (the Pi in the cabin, the Arduinos on the masts). Occasionally when the Wifi or 3G shield picks up a network, it replicated back to a cloud server.
Use MongoDB, it's web-scale. They produce kick-ass benchmarks by piping all your data to /dev/null.
Rules of Conduct:
#1 - The DM is always right.
#2 - If the DM is wrong, see rule #1
If the goal really is just to amass data and then do offline reports on it (not completely clear from the question) then I can report that at my company we've been doing this at scale for over five years. Here's how:
* A bunch of web servers accept data and append it to a local disk file.
* Every hour, that "log" is pushed from each host into HDFS and a new log file started. (HDFS as in the Hadoop Distributed Filesystem)
* Querying is done later, using Hive with a custom deserializer that natively understands our on-disk format. (You could also just make sure your on disk-format is the delimited text format Hive natively understands, of course. We had some unique needs here.)
* An hourly task runs a small set of Hive aggregation queries (Hive presents a SQL-like interface to defining and running MapReduce jobs) on the raw "table" to produce some smaller datasets that can return aggregate-based results faster than the raw data, including copying some of the smaller aggregates into a MySQL database for online access via some reporting applications.
At this point our daily dataset is a few terabytes in size, when considering the sum of all of the collecting servers across all of the hours. (There are some peak hours due to the nature of our business, so the volume isn't even across the whole day.)
The only thing we've ever disliked about this system is the delay between data arriving and it being available to query. For a little while experimented with using Apache Storm to with realtime log streaming, and produced a working prototype that was shown to work for a one-tenth sample of the data, but ultimately we concluded that the need for faster data wasn't strong enough to justify the additional complexity and stuck with the above design. Therefore I can't speak to how far that solution would scale, but if real-time analysis isn't a requirement -- and scaling up in data size is -- then I can certainly recommend the above design.
... then everything looks like a (NoSQL) nail. Who says you need NoSQL? Nothing against using cool, newish stuff, but as others have pointed out, you didn't describe the scale of your project. Don't blindly pick trendy technology just because you want to sit with the cool kids at lunchtime. If this is an alpha or beta product with under 1 billion records, use a regular database and be done with that. Move onto the interesting parts of your project and fix the plumbing later if you need to.
I have used Oracle, MySQL, and Mongo in prod situations. I have looked at Cassandra for evaluating it for potential usage in prod.
I can imagine situations where I could recommend any of the above. For example, if you are large financial company with billions of rows, I would go with Oracle. If you have smarts but not money and didn't need somebody to sue if something went wrong, then maybe Postgres would do . If I were a simple web based app with simple form submits, I would go with MySQL. If I had complex unpredictable data blobs and unpredictable needs to do certain types of queries against the data, I might recommend Mongo. If I have large amounts of data on which I want to do analytics I would use Cassandra.
Cassandra wins when you have a lot of data and not a lot of complex real time queries against it. It is especially good at scaling up on cheap data storage (think 100s of terabytes). It also has an unreal "write" throughput (important for certain types of analytics which write out complex intermediate results) though that is not relevant for the case described.
The problem generally with noSql solutions is that they increase the amount of storage to store the equivalent amount of information. You are essentially redundantly storing schema design with each "record" that you store. This really matters more than some might suspect, because when you can put an entire collection into memory, the read performance is much higher. You usually need 1/5th to 1/10th as much RAM to do the job with a traditional relational database (especially since MySQL and their brethren handle getting in and out memory better than mongo). This isn't so much the case for Cassandra because of its distributed storage nature, but it really isn't usable for real time transactions.
My recommendation, use a traditional database -- if in a Microsoft shop use SQL Server, otherwise I like postgres or mysql. If however, you have complex data storage needs that a noSql solution is perfect for, then I would go with that. If you are into back end analytics, copy the data as it comes in and put into a Cassandra (or one of its similar brethren) as well.