We have ~1Tb of user profiles and need to perform two types operations on them:
random reads and writes (~20k profile updates per second)
queries on predefined dimensions (e.g. for reporting)
For example, if we encounter user in a transaction, we want to update his profile with a URL he came from. At the end of the day we want to see all users who visited particular URL. We don't need joins, aggregations, etc., only filtering by one or several fields.
We don't really care about latency, but need high throughput.
Most databases we looked at belong to one of two categories - key-value DBs with fast random access or batch DBs optimized for querying and analytics.
Key-value storages
Aerospike can store terabyte-scale data and is very well-optimized for fast key-based lookup. However, queries on secondary index are deadly slow, which makes it unsuitable for our purposes.
MongoDB is pretty flexible, but requires too much hardware to handle our load. In addition, we encountered particular issues with massive exports from it.
HBase looks attractive since we already have Hadoop cluster. Yet, it's not really clear how to create secondary index for it and what its performance will be.
Cassandra - may be an option, but we don't have experience with it (if you do, please share it)
Couchbase - may be an option, but we don't have experience with it (if you do, please share it)
Analytic storages
Relational DBMS (e.g. Oracle, PostreSQL) provide both - random access and efficient queries, but we have doubts that they can handle terabyte data.
HDFS / Hive / SparkSQL - excellent for batch processing, but doesn't support indexing. The closest thing is partitioning, but it's not applicable given many-to-many relations (e.g. many users visited many URLs). Also, to our knowledge none of HDFS-backed tools except for HBase support updates, so you can only append new data and read latest version, which is not very convenient.
Vertica has very efficient queries, but updates boil down to rewriting the whole file, so are terribly slow.
(Because of limited experience some of information above may be subjective or wrong, please feel free to comment about it)
Do any of the mentioned databases have useful options that we missed?
Is there any other database(s) optimized for your use case? If not, how would you address this task?
For large web sites (traffic wise) that has alot of incoming reads and updates that end up being database I/Os, what're the best ways to mitigate the performance impact? one solution that I can think of is - for write, to cache and then do delayed write (using separate job); for read, use memcached concept. any other better solutions?
Here are the most common solutions to database performance:
Caching (Memcache, etc)
Add memory to your database
More database servers (master/slave or sharding)
Use a different database type (NoSQL, Redis, etc)
Indexes to speed up read perf. (careful, too many will affect write performance)
SSDs (fast SSDs will help a lot)
RAID
Optimize/tune SQL queries
Don't forget to optimize your queries. Most of the times it is not the disk I/O, but poorly written queries which turn out to be the bottleneck.
You can also cache query results and also entire web pages if the content isn't going to change too often.
It very much depends on the usage pattern and data type. There are really different things to do depending on whether transaction are going to be supported, whether you are interested in full consistency or "eventual consistency", how big the data is (will it all fit in huge memory?), how complex the data and queries are, the list might go on and on.... Lots of variables and only after listing all the constraints/requirements you will be able to make a proper decision. Two general advices though:
Use SSDs
Use distributed architecture with distributed "NoSQL" (key/value) approach (only if you do not have to use complex relations and transactions)
10 years ago, the standard answer - besides optimizing your particular database - was scale-out using MySQL in two ways.
Reads can be scaled out in two ways. The first is through caching, which introduces possible inconsistancies and creates a separate cache layer. Reads can also be scaled in MySQL by creating "read replicas", where any database can be queried. Any write must be applied to all servers, so replication doesn't help write throughput.
Writes are scaled through sharding. For example, imagine all users with the last name 'a' are assigned to a certain server. Now imagine a more complicated shard algorithm, where a particular row's primary ID is hashed using a hash function, and distributed to one of a pool of servers.
Facebook is one of the most advanced proponents of a sharded MySQL architecture. You can have individual tables "joined" but you have to write custom code, because you might have to hop from server to server - imagine you want to get your friend's timeline posts, you can't simply join it, you have to write some application code.
Once you shard your database, you can't do joins and range lookups become difficult. This subset is sometimes called CRUD operations, and thus MySQL is overkill. Many Chinese social networks realized this, and use sharded Redis (which is much quicker than MySQL), and have written their own shard layer and application logic layers.
Imagine the next problem in sharding - you want to add a new server, and start assigning some users to that new server.
Another approach is to use a distributed database, which generally comes under the names NoSQL or NewSQL, and have a variety of approaches. Some, like MongoDB, have a sharding system to manage this mapping, but require manual steps to add servers. Cassandra has a more flexible clustering scheme, called a chorded architecture. Systems like CouchBase and Aerospike use a random distribution mechanism that remove the need for a shard layer. Some of these databases can exceed 100,000 to 200,000 requests per second per server, with the lateral scale to add new servers - enough for very large operations. With this style of clustering, you can often get a higher level of redundancy and reliability.
Other distributed approaches represent data in a more efficient way, like a graph database. If you have a problem that is better represented as a graph, then a clustered graph database may be more appropriate.
Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 2 years ago.
Improve this question
There has been a lot of talk related to Cassandra lately.
Twitter, Digg, Facebook, etc all use it.
When does it make sense to:
use Cassandra,
not use Cassandra, and
use a RDMS instead of Cassandra.
There is nothing like a silver bullet, everything is built to solve specific problems and has its own pros and cons. It is up to you, what problem statement you have and what is the best fitting solution for that problem.
I will try to answer your questions one by one in the same order you asked them. Since Cassandra is based on the NoSQL family of databases, it's important you understand why use a NoSQL database before I answer your questions.
Why use NoSQL
In the case of RDBMS, making a choice is quite easy because all the databases like MySQL, Oracle, MS SQL, PostgreSQL in this category offer almost the same kind of solutions oriented toward ACID properties. When it comes to NoSQL, the decision becomes difficult because every NoSQL database offers different solutions and you have to understand which one is best suited for your app/system requirements. For example, MongoDB is fit for use cases where your system demands a schema-less document store. HBase might be fit for search engines, analyzing log data, or any place where scanning huge, two-dimensional join-less tables is a requirement. Redis is built to provide In-Memory search for varieties of data structures like trees, queues, linked lists, etc and can be a good fit for making real-time leaderboards, pub-sub kind of system. Similarly there are other databases in this category (Including Cassandra) which are fit for different problem statements. Now lets move to the original questions, and answer them one by one.
When to use Cassandra
Being a part of the NoSQL family, Cassandra offers a solution for problems where one of your requirements is to have a very heavy write system and you want to have a quite responsive reporting system on top of that stored data. Consider the use case of Web analytics where log data is stored for each request and you want to built an analytical platform around it to count hits per hour, by browser, by IP, etc in a real time manner. You can refer to this blog post to understand more about the use cases where Cassandra fits in.
When to Use a RDMS instead of Cassandra
Cassandra is based on a NoSQL database and does not provide ACID and relational data properties. If you have a strong requirement for ACID properties (for example Financial data), Cassandra would not be a fit in that case. Obviously, you can make a workaround for that, however you will end up writing lots of application code to simulate ACID properties and will lose on time to market badly. Also managing that kind of system with Cassandra would be complex and tedious for you.
When not to use Cassandra
I don't think it needs to be answered if the above explanation makes sense.
When evaluating distributed data systems, you have to consider the CAP theorem - you can pick two of the following: consistency, availability, and partition tolerance.
Cassandra is an available, partition-tolerant system that supports eventual consistency. For more information see this blog post I wrote: Visual Guide to NoSQL Systems.
Cassandra is the answer to a particular problem: What do you do when you have so much data that it does not fit on one server ? How do you store all your data on many servers and do not break your bank account and not make your developers insane ? Facebook gets 4 Terabyte of new compressed data EVERY DAY. And this number most likely will grow more than twice within a year.
If you do not have this much data or if you have millions to pay for Enterprise Oracle/DB2 cluster installation and specialists required to set it up and maintain it, then you are fine with SQL database.
However Facebook no longer uses cassandra and now uses MySQL almost exclusively moving the partitioning up in the application stack for faster performance and better control.
The general idea of NoSQL is that you should use whichever data store is the best fit for your application. If you have a table of financial data, use SQL. If you have objects that would require complex/slow queries to map to a relational schema, use an object or key/value store.
Of course just about any real world problem you run into is somewhere in between those two extremes and neither solution will be perfect. You need to consider the capabilities of each store and the consequences of using one over the other, which will be very much specific to the problem you are trying to solve.
Besides the answers given above about when to use and when not to use Cassandra, if you do decide to use Cassandra you may want to consider not using Cassandra itself, but one of the its many cousins out there.
Some answers above already pointed to various "NoSQL" systems which share many properties with Cassandra, with some small or large differences, and may be better than Cassandra itself for your specific needs.
Additionally, recently (several years after this question was originally asked), a Cassandra clone called Scylla (see https://en.wikipedia.org/wiki/Scylla_(database)) was released. Scylla is an open-source re-implementation of Cassandra in C++, which claims to have significantly higher throughput and lower latencies than the original Java Cassandra, while being mostly compatible with it (in features, APIs, and file formats). So if you're already considering Cassandra, you may want to consider Scylla as well.
I will focus here on some of the important aspects which can help you to decide if you really need Cassandra. The list is not exhaustive, just some of the points which I have at top of my mind-
Don't consider Cassandra as the first choice when you have a strict requirement on the relationship (across your dataset).
Cassandra by default is AP system (of CAP). But, it supports tunable consistency which means it can be configured to support as CP as well. So don't ignore it just because you read somewhere that it's AP and you are looking for CP systems. Cassandra is more accurately termed “tuneably consistent,” which means it allows you to easily decide the level of consistency you require, in balance with the level of availability.
Don't use Cassandra if your scale is not much or if you can deal with a non-distributed DB.
Think harder if your team thinks that all your problems will be solved if you use distributed DBs like Cassandra. To start with these DBs is very simple as it comes with many defaults but optimizing and mastering it for solving a specific problem would require a good (if not a lot) amount of engineering effort.
Cassandra is column-oriented but at the same time each row also has a unique key. So, it might be helpful to think of it as an indexed, row-oriented store. You can even use it as a document store.
Cassandra doesn't force you to define the fields beforehand. So, if you are in a startup mode or your features are evolving (as in agile) - Cassandra embraces it. So better, first think about queries and then think about data to answer them.
Cassandra is optimized for really high throughput on writes. If your use case is read-heavy (like cache) then Cassandra might not be an ideal choice.
Right. It makes sense to use Cassandra when you have a huge amount of data, a huge number of queries but very little variety of queries. Cassandra basically works by partitioning and replicating. If all your queries will be based on the same partition key, Cassandra is your best bet. If you get a query on an attribute that is not the partition key, Cassandra allows you to replicate the whole data with a new partition key. So now you have 2 replicas of the same data with 2 different partition keys.
Which brings me to your next question. When not to use Cassandra. As I mentioned, Cassandra scales by replicating the complete database for every new partitioning key. But you can't keep making new copies again and again. So when you have a high variety in queries i.e. each query has a different column in the where clause, Cassandra is not a good option.
Now for the third question. The whole point of using RDBMS is when you want the ACID properties. If you are building something like a payment service and want each transaction to be isolated, each transaction to either complete or not happen at all, changes to be persistent despite system failure, and the money to be consistent across bank accounts before and after the transaction completes, an RDBMS is the only option that will help you achieve this.
This article actually explains the whole thing, especially when to use Cassandra or not (as opposed to some other NoSQL option) part of the question -> Choosing the best Database. Do check it out.
EDIT: To answer the question in the comments by proximab, when we think of banking systems we immidiately think "ACID is the best solution". But even banking systems are made up of several subsystems that might not even be dealing with any transaction related data like account holder's personal information, account statements, credit card details, credit histories, etc.
All of this information needs to be stored in some database or the another. Now if you store the account related information like account balance, that is something that needs to be consistent at all times. For example, if you try to send money from account A to account B, then the money that disappears from account A should instantaneousy show up in account B, and it cannot be present in both accounts at the same time. This system cannot be inconsistant at any point. This is where ACID is of utmost importance.
On the other hand if you are saving credit card details or credit histories, that should not get into the wrong hands, then you need something that allows access only to authorised users. That I believe is supported by Cassandra. That said, data like credit history and credit card transactions, I think that is an ever increasing data. Also there is only so much yo can query on this data i.e. it has a very finite number of queries. These two conditions make Cassandra a perfect solution.
Talking with someone in the midst of deploying Cassandra, it doesn't handle the many-to-many well. They are doing a hack job to do their initial testing. I spoke with a Cassandra consultant about this and he said he wouldn't recommend it if you had this problem set.
You should ask your self the following questions:
(Volume, Velocity) Will you be writing and reading TONS of information , so much information that no one computer could handle the writes.
(Global) Will you need this writing and reading capability around the world so that the writes in one part of the world are accessible in another part of the world?
(Reliability) Do you need this database to be up and running all the time and never go down regardless of which Cloud, which country, whether it's VM , Container, or Bare metal?
(Scale-ability) Do you need this database to be able to continue to grow easily and scale linearly
(Consistency) Do you need TUNABLE consistency where some writes can happen asynchronously where as others need to be certified?
(Skill) Are you willing to do what it takes to learn this technology and the data modeling that goes with creating a globally distributed database that can be fast for everyone, everywhere?
If for any of these questions you thought "maybe" or "no," you should use something else. If you had "hell yes" as an answer to all of them, then you should use Cassandra.
Use RDBMS when you can do everything on one box. It's probably easier than most and anyone can work with it.
Heavy single query vs. gazillion light query load is another point to consider, in addition to other answers here. It's inherently harder to automatically optimize a single query in a NoSql-style DB. I've used MongoDB and ran into performance issues when trying to calculate a complex query. I haven't used Cassandra but I expect it to have the same issue.
On the other hand, if your load is expected to be that of very many small queries, and you want to be able to easily scale out, you could take advantage of eventual consistency that is offered by most NoSql DBs. Note that eventual consistency is not really a feature of a non-relational data model, but it is much easier to implement and to set up in a NoSql-based system.
For a single, very heavy query, any modern RDBMS engine can do a decent job parallelizing parts of the query and take advantage of as much CPU and memory you throw at it (on a single machine). NoSql databases don't have enough information about the structure of the data to be able to make assumptions that will allow truly intelligent parallelization of a big query. They do allow you to easily scale out more servers (or cores) but once the query hits a complexity level you are basically forced to split it apart manually to parts that the NoSql engine knows how to deal with intelligently.
In my experience with MongoDB, in the end because of the complexity of the query there wasn't much Mongo could do to optimize it and run parts of it on multiple data. Mongo parallelizes multiple queries but isn't so good at optimizing a single one.
Let's read some real world cases:
http://planetcassandra.org/apache-cassandra-use-cases/
In this article: http://planetcassandra.org/blog/post/agentis-energy-stores-over-15-billion-records-of-time-series-usage-data-in-apache-cassandra
They elaborated the reason why they didn't choose MySql is because db synchronization is too slow.
(Also due to 2-phrase commit, FK, PK)
Cassandra is based on Amazon Dynamo paper
Features:
Stability
High availability
Backup performs well
Read and Write is better than HBase, (BigTable clone in java).
wiki http://en.wikipedia.org/wiki/Apache_Cassandra
Their Conclusion is:
We looked at HBase, Dynamo, Mongo and Cassandra.
Cassandra was simply the best storage solution for the majority of our data.
As of 2018,
I would recommend using ScyllaDB to replace classic cassandra, if you need back support.
Postgres kv plugin is also quick than cassandra. How ever won't have multi-instance scalability.
another situation that makes the choice easier is when you want to use aggregate function like sum, min, max, etcetera and complex queries (like in the financial system mentioned above) then a relational database is probably more convenient then a nosql database since both are not possible on a nosql databse unless you use really a lot of Inverted indexes. When you do use nosql you would have to do the aggregate functions in code or store them seperatly in its own columnfamily but this makes it all quite complex and reduces the performance that you gained by using nosql.
Cassandra is a good choice if:
You don't require the ACID properties from your DB.
There would be massive and huge number of writes on the DB.
There is a requirement to integrate with Big Data, Hadoop, Hive and Spark.
There is a need of real time data analytics and report generations.
There is a requirement of impressive fault tolerant mechanism.
There is a requirement of homogenous system.
There is a requirement of lots of customisation for tuning.
If you need a fully consistent database with SQL semantics, Cassandra is NOT the solution for you. Cassandra supports key-value lookups. It does not support SQL queries. Data in Cassandra is "eventually consistent". Concurrent lookups of data may be inconsistent, but eventually lookups are consistent.
If you need strict semantics and need support for SQL queries, choose another solution such as MySQL, PostGres, or combine use of Cassandra with Solr.
Apache cassandra is a distributed database for managing large amounts of structured data across many commodity servers, while providing highly available service and no single point of failure.
The archichecture is purely based on the cap theorem, which is availability , and partition tolerance, and interestingly eventual consistently.
Dont Use it, if your not storing volumes of data across racks of clusters,
Dont use if you are not storing Time series data,
Dont Use if you not patitioning your servers,
Dont use if you require strong Consistency.
Mongodb has very powerful aggregate functions and an expressive aggregate framework. It has many of the features developers are accustomed to using from the relational database world. It's document data/storage structure allows for more complex data models than Cassandra, for example.
All this comes with trade-offs of course. So when you select your database (NoSQL, NewSQL, or RDBMS) look at what problem you are trying to solve and at your scalability needs. No one database does it all.
According to DataStax, Cassandra is not the best use case when there is a need for
1- High end hardware devices.
2- ACID compliant with no roll back (bank transaction)
It does not support complete transaction management across the
tables.
Secondary Index not supported.
Have to rely on Elastic search /Solr for Secondary index and the custom sync component has to be written.
Not ACID compliant system.
Query support is limited.
I've heard a lot about couchdb lately, and am confused about what it offers.
It's hard to explain all the differences in strict advantage/disadvantage form.
I would suggest playing with CouchDB a little yourself. The first thing you'll notice is that the learning curve during initial usage is totally inverted from RDBMS.
With RDBMS you spend a lot of up front time modeling your real world data to get it in to the Database. Once you've dealt with the modeling you can do all kinds of queries.
With CouchDB you just get all your data in JSON and stored in the DB in, literally, minutes. You don't need to do any normalization or anything like that, and the transport is HTTP so you have plenty of client options.
Then you'll notice a big learning curve when writing map functions and learning how the key collation works and the queries against the views you write. Once you learn them, you'll start to see how views allow you to normalize the indexes while leaving the data un-normalized and "natural".
CouchDB is a document-oriented database.
Wikipedia:
As opposed to Relational Databases, document-based databases do not store data in tables with uniform sized fields for each record. Instead, each record is stored as a document that has certain characteristics. Any number of fields of any length can be added to a document. Fields can also contain multiple pieces of data.
Advantages:
You don't waste space by leaving empty fields in documents (because they're not necessarily needed)
By providing a simple frontend for editing it is possible to quickly set up an application for maintaining data.
Fast and agile schema updates/changes
Map Reduce queries in a turing complete language of your choice. (no more sql)
Flexible Schema designs
Freeform Object Storage
Really really easy replication
Really Really easy Load-Balancing (soon)
Take a look here.
I think what better answers you is:
Just as CouchDB is not always the
right tool for the job, RDBMS's are
also not always the right answer.
CouchDB is a disk hog because it doesn't update documents -- it creates a new revision each time you update so the not-wasting-space-part because you don't have empty fields is trumped by the revisions.
New school datastore paradigms like Google BigTable and Amazon SimpleDB are specifically designed for scalability, among other things. Basically, disallowing joins and denormalization are the ways this is being accomplished.
In this topic, however, the consensus seems to be that joins on large tables don't necessarilly have to be too expensive and denormalization is "overrated" to some extent
Why, then, do these aforementioned systems disallow joins and force everything together in a single table to achieve scalability? Is it the sheer volumes of data that needs to be stored in these systems (many terabytes)?
Do the general rules for databases simply not apply to these scales?
Is it because these database types are tailored specifically towards storing many similar objects?
Or am I missing some bigger picture?
Distributed databases aren't quite as naive as Orion implies; there has been quite a bit of work done on optimizing fully relational queries over distributed datasets. You may want to look at what companies like Teradata, Netezza, Greenplum, Vertica, AsterData, etc are doing. (Oracle got in the game, finally, as well, with their recent announcement; Microsoft bought their solition in the name of the company that used to be called DataAllegro).
That being said, when the data scales up into terabytes, these issues become very non-trivial. If you don't need the strict transactionality and consistency guarantees you can get from RDBMs, it is often far easier to denormalize and not do joins. Especially if you don't need to cross-reference much. Especially if you are not doing ad-hoc analysis, but require programmatic access with arbitrary transformations.
Denormalization is overrated. Just because that's what happens when you are dealing with a 100 Tera, doesn't mean this fact should be used by every developer who never bothered to learn about databases and has trouble querying a million or two rows due to poor schema planning and query optimization.
But if you are in the 100 Tera range, by all means...
Oh, the other reason these technologies are getting the buzz -- folks are discovering that some things never belonged in the database in the first place, and are realizing that they aren't dealing with relations in their particular fields, but with basic key-value pairs. For things that shouldn't have been in a DB, it's entirely possible that the Map-Reduce framework, or some persistent, eventually-consistent storage system, is just the thing.
On a less global scale, I highly recommend BerkeleyDB for those sorts of problems.
I'm not too familiar with them (I've only read the same blog/news/examples as everyone else) but my take on it is that they chose to sacrifice a lot of the normal relational DB features in the name of scalability - I'll try explain.
Imagine you have 200 rows in your data-table.
In google's datacenter, 50 of these rows are stored on server A, 50 on B, and 100 on server C. Additionally server D contains redundant copies of data from server A and B, and server E contains redundant copies of data on server C.
(In real life I have no idea how many servers would be used, but it's set up to deal with many millions of rows, so I imagine quite a few).
To "select * where name = 'orion'", the infrastructure can fire that query to all the servers, and aggregate the results that come back. This allows them to scale pretty much linearly across as many servers as they like (FYI this is pretty much what mapreduce is)
This however means you need some tradeoffs.
If you needed to do a relational join on some data, where it was spread across say 5 servers, each of those servers would need to pull data from eachother for each row. Try do that when you have 2 million rows spread across 10 servers.
This leads to tradeoff #1 - No joins.
Also, depending on network latency, server load, etc, some of your data may get saved instantly, but some may take a second or 2. Again, when you have dozens of servers, this gets longer and longer, and the normal approach of 'everyone just waits until the slowest guy has finished' no longer becomes acceptable.
This leads to tradeoff #2 - Your data may not always be immediately visible after it's written.
I'm not sure what other tradeoffs there are, but off the top of my head those are the main 2.
So what I'm getting is that the whole "denormalize, no joins" philosophy exists, not because joins themselves don't scale in large systems, but because they're practically impossible to implement in distributed databases.
This seems pretty reasonable when you're storing largely invariant data of a single type (Like Google does). Am I on the right track here?
If you are talking about data that is virtually read-only, the rules change. Denormalisation is hardest in situations where data changes because the work required is increased and there are more problems with locking. If the data barely changes then denormalisation is not so much of a problem.
Novaday You need to find more interoperational environment for databases. More frequently You don't need only an relational DBs, like MySQL or MS SQL but also Big Data farms as Hadoop or non-relational DBs like MongoDB. In some cases all those DBs will be used in one solution so their performance must be as equal as possible in macro scale. It means, that You will not be able to use let say Azure SQL as relational DB and one VM with 2 cores and 3GB of RAM for MongoDB. You must scale-up Your solution and use DB as a Service when it is possible (if it is not possible, then build Your own cluster in a cloud).