Dominic Williams

Occasionally useful posts about RIAs, Web scale computing & miscellanea

HBase vs Cassandra: why we moved

with 94 comments

My team is currently working on a brand new product – the forthcoming MMO This has given us the luxury of building against a NOSQL database, which means we can put the horrors of MySQL sharding and expensive scalability behind us. Recently a few people have been asking why we seem to have changed our preference from HBase to Cassandra. I can confirm the change is true and that we have in fact almost completed porting our code to Cassandra, and here I will seek to provide an explanation.

For those that are new to NOSQL, in a following post I will write about why I think we will see a seismic shift from SQL to NOSQL over the coming years, which will be just as important as the move to cloud computing. That post will also seek to explain why I think NOSQL might be the right choice for your company. But for now I will simply relay the reasons why we have chosen Cassandra as our NOSQL solution.

Caveat Emptor – if you’re looking for a shortcut to engaging your neurons be aware this isn’t an exhaustive critical comparison, it just summarizes the logic of just another startup in a hurry with limited time and resources!!

Did Cassandra’s bloodline foretell the future?

One of my favourite tuppences for engineers struggling to find a bug is “breadth first not depth first”. This can be annoying for someone working through complex technical details, because it implies that the solution is actually much simpler if they only looked (advice: only use this saying with established colleagues who will forgive you). I coined this saying because in software matters I find that if we force ourselves to examine the top level considerations first, before tunnelling down into the detail of a particular line of enquiry, we can save enormous time.

So before getting technical, I’ll mention I might have heeded my motto better when we were making our initial choice between HBase and Cassandra. The technical conclusions behind our eventual switch might have been predicted: HBase and Cassandra have dramatically different bloodlines and genes, and I think this influenced their applicability within our business.

Loosely speaking, HBase and its required supporting systems are derived from what is known of the original Google BigTable and Google File System designs (as known from the Google File System paper Google published in 2003, and the BigTable paper published in 2006). Cassandra on the other hand is a recent open source fork of a standalone database system initially coded by Facebook, which while implementing the BigTable data model, uses a system inspired by Amazon’s Dynamo for storing data (in fact much of the initial development work on Cassandra was performed by two Dynamo engineers recruited to Facebook from Amazon).

In my opinion, these differing histories have resulted in HBase being more suitable for data warehousing, and large scale data processing and analysis (for example, such as that involved when indexing the Web) and Cassandra being more suitable for real time transaction processing and the serving of interactive data. Writing a proper study of that hypothesis is well beyond this post, but I believe you will be able to detect this theme recurring when considering the databases.

NOTE: if you are looking for lightweight validation you’ll find the current makeup of the key committers interesting: the primary committers to HBase work for Bing (M$ bought their search company last year, and gave them permission to continue submitting open source code after a couple of months). By contrast the primary committers on Cassandra work for Rackspace, which supports the idea of an advanced general purpose NOSQL solution being freely available to counter the threat of companies becoming locked in to the proprietary NOSQL solutions offered by the likes of Google, Yahoo and Amazon EC2.

Malcolm Gladwell would say my unconscious brain would have known immediately that my business would eventually prefer Cassandra based upon these differing backgrounds. It is horses for courses. But of course, justifying a business decision made in the blink of an eye is difficult…

Which NOSQL database has the most momentum?

Another consideration that has persuaded us to move to Cassandra is a belief that it is now has the most general momentum in our community. As you know, in the business of software platforms the bigger you get the bigger you get – where platforms are perceived as similar, people tend to aggregate around the platform that is going to offer the best supporting ecosystem in the long term (i.e. where the most supporting software is available from the community, and where the most developers are available for hire). This effect is self-reinforcing.

When starting with HBase, my impression then was that it had the greatest community momentum behind it, but I now believe that Cassandra is coming through much stronger. The original impression was partly created by two very persuasive and excellently delivered presentations given by the CTOs of StumpleUpon and Streamy, two big players in the Web industry who committed to HBase some time before Cassandra was really an option, and also from a quick reading of an article entitled “HBase vs Cassandra: NoSQL Battle!” (much of which has now been widely debunked).

Proving momentum comprehensively is difficult to do, and you will have to poke about for yourself, but one simple pointer I offer you is the developer activity on IRC. If you connect to and compare the #hbase and #cassandra developer channels, you will find Cassandra typically has twice the number of developers online at any time.

If you consider Cassandra has been around for half as long as HBase, you can see why this is quite a clear indication of the accelerating momentum behind Cassandra. You might also take note of the big names coming on board, such as Twitter, where they plan broad usage (see here).

Note: Cassandra’s supporting website looks much lovelier than HBase’s, but seriously, this could be a trend driven by more than the marketing. Read on!

Deep down and technical: CAP and the myth of CA vs AP

There is a very powerful theorem that applies to the development of distributed systems (and here we are talking about distributed databases, as I’m sure you’ve noticed). This is known as the CAP Theorem, and was developed by Professor Eric Brewer, Co-founder and Chief Scientist of Inktomi.

The theorem states, that a distributed (or “shared data”) system design, can offer at most two out of three desirable properties – Consistency, Availability and tolerance to network Partitions. Very basically, “consistency” means that if someone writes a value to a database, thereafter other users will immediately be able to read the same value back, “availability” means that if some number of nodes fail in your cluster the distributed system can remain operational, and “tolerance to partitions” means that if the nodes in your cluster are divided into two groups that can no longer communicate by a network failure, again the system remains operational.

Professor Brewer is an eminent man and many developers, including many in the HBase community, have taken it to heart that their systems can only support two of these properties and have accordingly worked to this design principle. Indeed, if you search online posts related to HBase and Cassandra comparisons, you will regularly find the HBase community explaining that they have chosen CP, while Cassandra has chosen AP – no doubt mindful of the fact that most developers need consistency (the C) at some level.

However I need to draw to your attention to the fact that these claims are based on a complete non sequitur. The CAP theorem only applies to a single distributed algorithm (and here I hope Professor Brewer would agree). But there is no reason why you cannot design a single system where for any given operation, the underlying algorithm and thus the trade-off achieved is selectable. Thus while it is true that a system may only offer two of these properties per operation, what has been widely missed is that a system can be designed that allows a caller to choose which properties they want when any given operation is performed. Not only that, reality is not nearly so black and white, and it is possible to offer differing degrees of balance between consistency, availability and tolerance to partition. This is Cassandra.

This is such an important point I will reiterate: the beauty of Cassandra is that you can choose the trade-offs you want on a case by case basis such that they best match the requirements of the particular operation you are performing. Cassandra proves you can go beyond the popular interpretation of the CAP Theorem and the world keeps on spinning!

For example, let’s look at two different extremes. Let us say that I must read a value from the database with very high consistency – that is, where I will be 100% sure to receive the last copy of that data which was previously written. In this case, I can read the value from Cassandra specifying consistency level “ALL”, which requires that all the nodes that hold replicated copies of that data agree on its value. In this case, I have zero tolerance to either node failure, or network partition. At the other extreme, if I do not care about consistency particularly, and simply want the maximum possible performance, I can read the value from Cassandra using consistency level “ONE”. In this case, a copy is simply taken from a random node amongst those holding the replicas – and in this case, if the data is replicated three times, it does not matter if either of the two other nodes holding copies have failed or been partitioned from us, although now of course it is also possible that such conditions may mean the data I read is stale.

And better still, you are not forced to live in a black and white world. For example, in our particular application important read/write operations typically use consistency level “QUORUM”, which basically means – and I simplify so please research before writing your Cassandra app – that a majority of nodes in the replication factor agree. From our perspective, this provides both a reasonable degree of resilience to node failure and network partition, while still delivering an extremely high level of consistency. In the general case, we typically use the aforementioned consistency level of “ONE”, which provides maximum performance. Nice!

For us this is a very big plus for Cassandra. Not only can we now easily tune our system, we can also design it so that, for example, when a certain number of nodes fail, or the network connecting those nodes falters, our service continues operating in many respects, and only those aspects that require data consistency fail. HBase is not nearly so flexible, and the pursuit of a single approach within the system (CP) reminds me of the wall that exists between SQL developers and the query optimizer – something it is good to get beyond!

In our project then, Cassandra has proven by far the most flexible system, although you may find your brain at first loses consistency when considering your QUORUMs.

When is monolithic better than modular?

An important distinction between Cassandra and HBase, is that while Cassandra comes as a single Java process to be run per node, a complete HBase solution is really comprised of several parts: you have the database process itself, which may run in several modes, a properly configured and operational hadoop HDFS distributed file system setup, and a Zookeeper system to coordinate the different HBase processes. Does this mean then that this is a modularity win for HBase?

Although it is true that such a setup might promise to leverage the collective benefits of different development teams, in terms of systems administration the modularity of HBase cannot be considered a plus. In fact, especially for a smaller startup company, the modularity of HBase might be a big negative. Let me explain…

The underpinnings of HBase are pretty complex, and anyone in doubt of this should read the original Google File System and BigTable papers. Even setting up HBase in pseudo distributed mode on a single server is difficult – so difficult in fact that I did my best to write a guide that takes you past all the various gotchas in the minimum time (see if you wish to try it). As you will see from that guide, getting HBase up and running in this mode actually involves setting up two different system systems manually: first hadoop HDFS, then HBase itself.

Now to the point: the HBase configuration files are monsters, and your setup is vulnerable to the quirks in default network configurations (in which I include both the default networking setups on Ubuntu boxes, and the subtleties of Elastic IPs and internally assigned domain names on EC2). When things go wrong, you will be presented with reams of output in the log file. All the information you need to fix things is in there, and if you are a skilled admin you are going to get through it.

But what happens if it does wrong in production and you need to fix it in a hurry? And what happens if like us, you have a small team of developers with big ambitions and can’t afford a team of crack admins to be on standby 247?

Look seriously, if you’re an advanced db admin wanting to learn a NOSQL system, choose HBase. It’s so damn complex that safe pairs of hands are going to get paid well.

But if you’re a small team just trying to get to the end of the tunnel like us, wait ’til you hear the Gossip…

It’s Gossip talk dude, Gossip!

Cassandra is a completely symmetric system. That is to say, there are no master nodes or region servers like in HBase – every node plays a completely equal role in the system. Rather than any particular node or entity taking on a coordination role, the nodes in your cluster coordinate their activities using a pure P2P communication protocol called “Gossip”.

A description of Gossip and the model using it is beyond this post, but the application of P2P communication within Cassandra has been mathematically modelled to show that, for example, the time taken for the detection of node failure to be propagated across the system, or for a client request to be routed to the node(s) holding the data, occur deterministically within well bounded timeframes that are surprisingly small. Personally I believe that Cassandra represents one of the most exciting uses of P2P technology to date, but of course this idea is not relevant to choosing your NOSQL database!

What is relevant are the real benefits that the Gossip-based architecture gives to Cassandra’s users. Firstly, continuing with the theme of systems administration, life becomes much simpler. For example, adding a new node to the system becomes as simple as bootstrapping its Cassandra process and pointing it at a seed node (an existing node within your cluster). When you think of the underlying complexity of a distributed database running across, potentially, hundreds of nodes, the ability to add new nodes to scale up with such ease is incredible. Furthermore, when things go wrong you no longer have to consider what kind of nodes you are dealing with – everything is the same, which can make debugging a more progressive and repeatable process.

Secondly I have come to the conclusion that Cassandra’s P2P architecture provides it with performance and availability advantages. Load can be very evenly balanced across system nodes thus maximizing the potential for parallelism, the ability to continue seamlessly in the face of network partitions or node failures is greatly increased, and the symmetry between nodes prevents the temporary instabilities in performance that have been reported with HBase when nodes are added and removed (Cassandra boots quickly, and its performance scales smoothly as new nodes are added).

If you are looking for more evidence, you will be interested to read a report from a team with a vested interest in hadoop (i.e. which should favor HBase)…

A report is worth a thousand words. I mean graph right?

The first comprehensive benchmarking of NOSQL systems performed by Yahoo! Research now seems to bear out the general performance advantage that Cassandra enjoys, and on the face of it the figures do currently look very good for Cassandra.

At the time of writing these papers are in draft form and you can check them out here:

NOTE: in this report HBase performs better than Cassandra only respect of range scans over records. Although the Cassandra team believes they will quickly approach the HBase times, it is also worth pointing out that in a common configuration of Cassandra range scans aren’t even possible. I recommend this to you as being of no matter, because actually in practice you should implement your indexes on top of Cassandra, rather than seek to use range scans. If you are interested in issues relating to range scans and storing indexes in Cassandra, see my post here

FINAL POINT OF INTEREST: the Yahoo! Research team behind this paper are trying to get their benchmarking application past their legal department and make it available to the community. If they succeed, and I hope they do, we will be treated to an ongoing speed competition galore, and both HBase and Cassandra will doubtless be improving their times further.

A word on locking, and useful modularity

You may no doubt hear from the HBase camp that their more complex architecture is able to give you things that Cassandra’s P2P architecture can’t. An example that may be raised is the fact that HBase provides the developer with row locking facilities whereas Cassandra cannot (in HBase row locking can be controlled by a region server since data replication occurs within the hadoop layer below, whereas in Cassandra’s P2P architecture all nodes are equal, and therefore none can act as a gateway that takes responsibility for locking replicated data).

However, I would reflect this back as an argument about modularity, which actually favours Cassandra. Cassandra implements the BigTable data model but uses a design where data storage is distributed over symmetric nodes. It does that, and that’s all, but in the most flexible and performant manner possible. But if you need locking, transactions or any other functionality then that can be added to your system in a modular manner – for example we have found scalable locking quite simple to add to our application using Zookeeper and its associated recipes (and other systems such as Hazelcast might also exist for these purposes, although we have not explored them).

By minimizing its function to a narrower purpose, it seems to me that Cassandra manages to implement a design that executes that purpose better – as indicated for example by its selectable CAP tradeoffs. This modularity means you can build a system as you need it – want locking, grab yourself Zookeeper, want to store a full text index, grab yourself Lucandra, and so on. For developers like us, this means we don’t have to take on board more complexity than we actually need, and ultimately provides us with a more flexible route to building the application we want.

MapReduce, don’t mention MapReduce!

One thing Cassandra can’t do well yet is MapReduce! For those not versed in this technology, it is a system for the parallel processing of vast amounts of data, such as the extraction of statistics from millions of pages that have been downloaded from the Web. MapReduce and related systems such as Pig and Hive work well with HBase because it uses hadoop HDFS to store its data,  which is the platform these systems were primarily designed to work with. If you need to do that kind of data crunching and analysis, HBase may currently be your best option.

Remember, it’s horses for courses!

Therefore as I finish off my impassioned extolation of Cassandra’s relative virtues, I should point out HBase and Cassandra should not necessarily be viewed as out and out competitors. While it is true that they may often be used for the same purpose, in much the same way as MySQL and Postgres, what I believe will likely emerge is that they will become preferred solutions for different applications. For example, as I understand StumbleUpon has been using HBase with the associated hadoop MapReduce technologies to crunch the vast amounts of data added to its service. Twitter is now using Cassandra for real time interactive community posts. Our needs fit better with the interactive serving and processing of data and so we are using Cassandra, and probably to some degree there you have it.

As a controversial parting shot though the gloves are off for the next point!

NOTE: before I continue I should point out Cassandra has hadoop support in 0.6, so its MapReduce integration may be about to get a whole load better.

O boy, I can’t afford to lose that data…

Perhaps as a result of the early CAP Theorem debates, an impression has grown that data is somehow safer in HBase than Cassandra. This is a final myth that I wish to debunk: in Cassandra, when you write new data it is actually immediately written to the commit log on one of the nodes in the quorum that will hold the replicas, as well as being replicated across the memory of the nodes. This means that if you have a complete power failure across your cluster, you will likely lose little data. Furthermore once in the system, data entropy is prevented using Merkle trees, which further add to the security of your data 🙂

In truth I am not clear exactly what the situation with HBase is – and I will endeavour to update this post as soon as possible with details – but my current understanding is that because hadoop does not yet support append, HBase cannot efficiently regularly flush its modified blocks of data to HDFS (whereupon the new mutations to data will be replicated and persisted). This means that there is a much larger window where your latest changes are vulnerable (if I am wrong, as I may be, please tell me and I will update the post).

So while the Cassandra of Greek mythology had a rather terrible time, the data inside your Cassandra shouldn’t.

NOTE: Wade Arnold points out below that (at the time of writing this) hadoop .21 is about to be released, which will solve this problem with HBase.


Written by dominicwilliams

February 24, 2010 at 7:27 pm

94 Responses

Subscribe to comments with RSS.

  1. Cassandra 0.6 (in beta now) supports Hadoop map/reduce, btw.

    Jonathan Ellis

    February 24, 2010 at 8:54 pm

  2. Your HDFS append comment is correct. This will require hadoop .21 which is currently being voted on with an imminent release as all tests are passing.

    Wade Arnold

    February 24, 2010 at 9:25 pm

  3. I think it’s so cool when people abbreviate Microsoft with a dollar sign (M$) . They seem so uber-1337 and take such a stand.


    February 24, 2010 at 10:24 pm

    • You’re probably right its a bit pathetic. The trouble is they’ve had so much of my money and I’m still getting over it.


      February 24, 2010 at 10:47 pm

  4. […] February 24, 2010 at 6:02 pm · Filed under Cloud computing ·Tagged Databases [From HBase vs Cassandra: why we moved « Bits and Bytes.] […]

  5. interesting read — the very first 1-2 paragraphs almost turned me off however….. anyways I kept reading and it definitely kept my interest — as stated I think comparing cassandra to hbase and friends is like comparing apples and oranges…. we’ve been looking at big data for a while for one of our projects — nice to see the steam building up around the web


    February 25, 2010 at 2:04 am

  6. CouchDB! It offers a robust map/reduce implementation, replication, and it talks http and json natively. It’s incredible.


    February 25, 2010 at 5:26 am

  7. As Wade says, the Hadoop team works very closely with the HBase team to get the last issue with HDFS append resolved for the next release. Here is my take on HBase’s write-ahed-log:

    One thing that sort of was never answered is scalability, while both systems are designed to do so it is known from the Dynamo paper that the Gossip protocol may eventually saturate the network and is something that needs to be improved on. Is that “fixed” in Cassandra?

    With BigTable you get a virtually boundless scalability with linear scan times and logarithmic read performance, not matter the size of the cluster. So with HBase the use-case I see is to be able to create a huge storage system and be able to efficiently process the data. Cassandra seems to be easier for rapid prototyping and flexibility along the way to determine what is needed for an otherwise unknown use-case. HBase has indeed a more strict model compared to that but is also on the other hand not promising the world. Bottom line is as you say, “horses for courses” and that this is no and should never be a shootout between NoSQL systems.

    As far as community is concerned, this is to be expected. A system backed by trendy sites like Facebook and Twitter ought to attract more developers while the rather more involved stup for a basic HBase cluster plus the fact that “M$” now uses it is just less spiffy. To me this is no indicator for success but for hype. And hype must not necessarily be a gauge for what is “better or worse”.


    February 25, 2010 at 8:31 am

    • FB, Twitter, etc. are using Cassandra because it is a better technical choice, not to be “trendy.” Ryan King listed some of their criteria for choosing Cassandra over at .

      Jonathan Ellis

      February 25, 2010 at 7:17 pm

      • Of course, I was just referring to the fact that there are more people on IRC for Cassandra as opposed to HBase’s channel. Not why they chose it.


        February 25, 2010 at 7:36 pm

      • Ah. Context is important there. Remember, in Bradford’s anti-Cassandra FUD piece, he claims that one of the reasons you should use HBase is that it’s more popular. So whether it is a useful metric or not, it’s worth correcting the record. 🙂

        Jonathan Ellis

        February 25, 2010 at 7:40 pm

    • I can’t believe you think FaceBook and Twitter technical architecture is “Trendy” as a point for your argument!….I almost spilled my drink.

      Jeryl Cook

      September 2, 2010 at 10:57 pm

  8. This was a great reading indeed. Thanks for taking the time 🙂


    February 25, 2010 at 9:24 am

  9. Interesting insights. Thanks for the in-depth post.


    February 25, 2010 at 5:57 pm

  10. […] HBase vs Cassandra: why we moved My team is currently working on a brand new product – the forthcoming MMO This has given us […] […]

    Top Posts —

    February 26, 2010 at 12:36 am

  11. Thanks Dominic for sharing your thoughts.

    I’m at a small startup and we’ve been happily using HBase in production on Amazon EC2 for 7 months. We have only 2 engineers who are primarily developers rather than db admins, and I can relate to your experience with the complexity of getting started. You have to learn how to work with HDFS, Zookeeper as well as HBase and there are so many configuration options.

    We also have the fear you talk about: what will we do if we run into a serious low-level production problem. There don’t seem to be any companies offering HBase support contracts. However, the community support has been great. We’ve always received quick responses on the HBase Users mailing list and IRC channel. The HBase committers patiently guided us through the one or two production problems we did encounter.

    Anyway, your post and the recent news from Twitter has definitely made me more curious about Cassandra. So much to learn in this space…

    Ken Weiner

    February 26, 2010 at 7:27 pm

    • Hi Ken,

      If you do decide you require HBase support, I’m happy to discuss the support levels we provide at Cloudera. Let me know.


      Kylie Clement

      March 4, 2010 at 1:38 am

    • IBM adopted Hadoop project and gives support to HFDS and HBASE.

      Tomer Paz

      June 14, 2011 at 7:43 pm

  12. Great comparison thanks Dom. In the spirit of cloud it’ll be interesting to see how services can remove even more of the admin work. Take SimpleDB and their recent “consistent read” announcement or the comment above re: hbase on EC2….

    Very cool.

    Chad Arimura

    February 27, 2010 at 6:39 pm

  13. nice post!

    just one thing that I’d like to mention about the “C” in Cassandra’s CAP. even when consistency level is set to ALL this doesn’t guarantee does it?

    consider the case when 2 clients perform a WRITE operation (WRITE1 & WRITE2), at the same time, to the same data, with consistency level ALL. WRITE1 may arrive first at half of the replicas while WRITE2 may arrive first at the other half.
    if these writes have no logic order there will be no way for the system to know which occurred “first”, and the decision will be left to the clients on future reads.
    I realize this is in the nature of eventually consistent systems, my point is only that people should be made aware of Cassandra’s/Riak’s/Dynamo’s consistency limitations

    On the other hand… I may be wrong too. But if you agree, then I think its a valuable piece of information to include in your post



    February 28, 2010 at 6:28 pm

    • Cassandra uses client-supplied timestamps to resolve conflicts like the one you describe.

      Jonathan Ellis

      March 1, 2010 at 2:10 pm

      • Yeah, that’s what I was getting at.

        From my understanding, even with consistency level set to ALL you are not guaranteed that all replicas will have a consistent view of the world.

        Do I have that right?

        My point (if the above is not completely wrong) is that Cassandra NEVER promises to be strongly consistency, regardless of settings.


        March 1, 2010 at 2:40 pm

      • In the improbable case where the client-supplied time stamps are of:
        (1) “long” type and are identical
        (2) “vector clock” type and have no causal order

        Does a client resolve the inconsistency at some point in the future when they perform a read?
        Or is the inconsistency resolved deterministically at server side?


        March 1, 2010 at 4:26 pm

      • 1) on the first read the server will pick one version and write it back to any replicas that have another version w/ the same ts during read repair

        2) cassandra does not yet support vector clocks, but one of the points of adding them is so you can push this kind of conflict back to the client and say “you know more about your domain than I do, you tell me what you want to do here”

        Jonathan Ellis

        March 1, 2010 at 4:28 pm

    • Hi Alex, the point is, that if you use consistency level ALL then Cassandra ***isn’t*** an “eventually consistent” system i.e. Cassandra does not behave like Dynamo in that case.

      If you use ALL for reads and writes on a value, then once you have written that value, subsequent reads will always return the same value (or fail, if a node in the replication factor is unavailable). The reason is that the operation contacts all nodes in the replication factor to ensure consistency.

      But note in practice you will get similar results using QUORUM, in which a simple majority of nodes in the replication factor are contacted. This provides better performance, and allows some nodes in the replication factor to be unavailable.

      In your Cassandra application, you will vary your consistency level depending upon the operation you are performing. Where consistency is not important, and you are emphasising performance instead, you will use the lowest consistency level of 1.


      March 1, 2010 at 2:13 pm

      • Hi Dominic, it seems like either we’re talking about different things or one of us is incorrect… very possibly me!

        As far as I understand, the scenario I described (2 writers, consistency level = ALL, half receive write1 1st, other half receive write2 1st) would NOT result in all replicas having consistent values, even in the case when there are zero failures.

        Just to clarify… I’m not inferring that Cassandra is flawed, I’m just trying to point out what seems to be a common misconception


        March 1, 2010 at 3:13 pm

    • Hi Alex, hopefully this is a better answer. Read it too quickly last time..

      Firsly, in NOSQL systems reads and writes are performed in separate operations. Thus, even if you always read back the value that you have written (i.e. because you are using ALL or QUORUM) without further measures a lost update can occur. For example two clients might read a number simultaneously, add differing amounts to it, then write it back. When the clients write the value back, the update of at least one client will be lost i.e. the number will only have been incremented by the amount added by the “winning” client.

      By contrast, with an SQL system you can read and write in a single statement, and thus the database systems themselves can serialize these client operations preventing lost updates happening (although in practice, as applications become complex, it is often not possible to encapsulate each operation into a single SQL statement and thus these problems can still occur).

      In order to avoid lost updates and related problems with NOSQL systems (and sometimes as mentioned with SQL systems, and certainly when you are sharding SQL systems) you have to use some kind of global synchronization mechanism. We are using ZooKeeper for this, which comes with Java recipes you can can use to create scalable global write locks and other things. Hazelcast is another system Java developers can look at.

      When I talk about a system being “eventually consistent”, I mean that it has the characteristic that you can write a value for some data, and then read it back with another value. This is the case for example with Amazon EC2’s SimpleDb (or at least it was until a few days ago, since they have just released new features allowing caller to specify desired consistency level).

      If you have this kind of “eventual consistency”, you cannot use global synchronization to be sure you are performing operations like debiting one bank account and crediting another correctly. However so long as you can always read back the value you just wrote, you can use global synchronization to do this (note that you can also do transactions with ZooKeeper, but since in our game the accounts we deal with don’t contain real money, we haven’t gone that far :))

      Personally, I think vector clocks will be a useful addition for high performance or high availability scenarios where you don’t want to enforce global serialization (or maybe even use strong consistency) but still need to be able to resolve conflicts in data in the rare case where they occur. However I reckon for most people global serialization will be the preferred option if conflicts must be dealt with simply because writing client-side conflict resolution code is always going to drag on productivity.


      March 1, 2010 at 5:19 pm

      • Unfortunately, this blog post and the comments misunderstand one fundamental element of Cassandra.

        In Cassandra, the write and read server-count requirements are only executed in the context of that specific operation. The ‘repair’ propagation does not use these values, but simply propagates any value with a newer timestamp.

        As a result, there is no way to achieve consistency. Regardless of the requirements of your write, if it makes it to any servers (even if it returns a failure to meet you server count requirement) it will be propagated by repair. Regardless of the read/write settings you use, this means that after a write, that write may or may not be visible immediately, and it may or may not show up later.

        Perhaps some future version will allow configuring repair to require a quorum to propagate a value. However, if that occurs, it won’t be something you can configure per call. At best it could be configured per keyrange.

        For a discussion of this and other nuances, see the Cassandra list.


        November 26, 2010 at 5:36 am


        Dave, I believe the important point you raise is *not* an issue as concerns implementation of systems that require consistent relationships amongst data.

        Yes it is true that if even if you write some data with CL QUORUM, and this write operation fails, that if some or all of the write reaches just one node in your cluster then read repair may propagate that write throughout the cluster, in actual fact this does *not* matter.

        The reasons is, in order to implement even complex systems that require consistency across data, it is only necessary that you can know when you *have* successfully written data with CL QUORUM (i.e. to a number of nodes greater than half your replication factor) so long as you are using a global synchronization mechanism.

        Very basically, where code requires that it sees data in a consistent state, access to that data must be mediated by for example a distributed lock system. When you wish to make updates to such data, or read that data in a consistent way, it is necessary that you do so when holding the appropriate locks. Furthermore, it is necessary that operations writing data do not release the locks until they have succeeded making the writes with CL QUORUM (more later on what happens if there is a crash) e.g.

        ZkMultiLock lock = new ZkMultiLock();
        try {
        while (!success)
        // try to write data with ConsistencyLevel.QUORUM
        finally {

        It does not matter if you try and write some data with CL QUORUM and it fails. Anyone that needs to read these values in a consistent state will be waiting for the lock. They will only attempt to read the data with CL QUORUM after it has been *successfully* written with CL QUORUM and the necessary locks have changed ownership.

        We use Cages/ZooKeeper for distributed locking. This makes our system possible, which supports huge numbers of transaction-style operations on highly contended data (buy/sell etc). We use Pelops to read and write data, and it automatically retries operations that fail because, for example, a cluster node is unavailable.

        What Cages cannot do for the moment – although this functionality is planned soon – is offer transaction mechanisms that deal with a situation where there is a system outage while waiting for all necessary writes to succeed at CL QUORUM. But please see the post on Cages that explains how these transactions will integrate with locks to make this possible.

        The only reason this functionality hasn’t been added to Cages already is that the transactions on our site ( involve virtual money and objects, so what we have already with Cages and CL QUORUM is enough for us in business terms. But the key thing is, Cassandra’s model is no barrier in principle to the implementation of systems that require data to have consistent data relationships.

        Personally I believe that Cassandra will eventually be widely used for systems where maintenance of consistent data relationships is required, just like in ours, which means it can easily move out from just being a niche product.

        Please checkout the post on Cages for discussion.


        December 1, 2010 at 12:43 am

  14. Zeitgeist at work: Amazon CTO Werner Vogels on the same wavelength?

    Just 3.5 hours after this post was made (after timezone differences), Werner Vogels made a short post covering similar ground and announcing the introduction of configurable consistency features to SimpleDb.

    This post was widely tweeted within the NOSQL community within minutes of being posted, and included two key subjects:

    1. The advantages of the configurable consistency model (which at the time of writing, only Cassandra provided).

    2. A discussion of Eric Brewer’s CAP Theorem in the context of NOSQL (although this post also focuses on how this theorem has been widely misunderstood in relation to NOSQL and other shared data systems – Eric Brewer sometimes describes the CAP Theorem as applying to “shared data systems” in general, e.g. see, which mislead a lot of people. In fact, the CAP Theorem can only possibly be applied to a specific set of algorithms that are used to minimally implement a shared data system, and there is no reason why you cannot design a system that makes the algorithms and thus tradeoffs configurable on an operation by operation basis, as does Cassandra).

    No doubt EC2 was planning to add configurable consistency to SimpleDb anyway, but maybe this post and the buzz it was getting bumped Werner Vogels into announcing it?

    Who knows, but it’s a nice coincidence!

    Check out his post at


    March 1, 2010 at 6:54 pm

  15. […] Getting Real about NoSQL and the SQL-Isn’t-Scalable Lie. В продолжение темы: HBase vs Cassandra: why we moved и Brewer’s CAP […]

  16. Thank you for this, Dominic. It covers a lot of the issues that most people seem to miss when they’re making these same sorts of decisions.

    Jeff Darcy

    March 19, 2010 at 4:10 pm

  17. If you need point queries (value by key), short scan queries (lookup all the versions of the same row/record ) and efficient large scan implementation for MapReduce jobs in one package than you choice is obvious – HBase. Good write up anyway.

    Vladimir Rodionov

    March 19, 2010 at 9:16 pm

  18. […] 原作者:Dominic Williams 原文发布日期:February 24, 2010 at 7:27 pm […]

  19. […] FightMyMonster team switched from HBase to Cassandra after concluding that "HBase is more suitable for data warehousing, and large scale data […]

  20. Hadoop .21 is a farce with Yahoo proposing to scrapping it and back porting some of its fixes/features to .20 and to begin .22. HDFS Append is still in limbo.


    March 28, 2010 at 6:06 am

  21. The descriptions are almostly right, but following tow:
    1. The reference of Yahoo’s benchmark report is not good.
    2. Pithy is better, and save writer and reader’s time.

    Thanks for you good post.


    March 31, 2010 at 11:15 am

  22. […] April 3, 2010 Sam Baskinger Leave a comment Go to comments Dominic Williams has an excellent post at Bits and Bytes on why he (his company, really) moved to Cassandra. This post is a simple […]

  23. […] HBase vs Cassandra: why we moved […]

  24. Dominic, I’d love to hear more about your experiences since you made the switch… especially around the day-to-day admin of the cluster nodes and the amount of attention they’ve needed? Have you found that your original design decisions were on the money or do you wish you’d structured things differently?



    April 7, 2010 at 11:25 am

    • Hi, is not on general release yet, so I won’t vouch for its scalability in production – but we’ve now been working with Cassandra for a while, have ported all our existing code base for FMM over and have also integrated related systems like Lucandra.

      What I can report is that in terms of general ease of configuration, setting up clusters and maintaining them, developing against its model etc we are very happy. As compared with our experiences with SQL sharding, HBase etc we are in a much better place, at least for us given our requirements. Based on our experiences, and having convinced ourselves of the NoSQL and in particular Cassandra models, we are planning to base future projects on Cassandra and indeed port some other unrelated ones across.

      Bearing in mind the above, below are the only real difficulties we have experienced so far. These are not so much criticisms as a reality check because Cassandra is cutting edge software under constant development:-

      1. We access Cassandra from Java. We started off using which wrapped the thrift interface (we thought quite nicely). This wasn’t tracking Cassandra releases so we migrated to a new client called Hector, which is the one currently most widely in use. The API of this client has changed several times, so there has been quite a bit of refactoring going on. Hector doesn’t yet surface some very useful features from the latest Thrift interface like batch mutate – we are considering writing basic connection pool / utility system for working with raw Thrift so keeping up to date easier.

      2. Recently had a strange problem where if we deleted a row from database, then created another row with same id, although the operation succeeded we could not see the row. The reason we have concluded was that the rows in question contained super columns and had been originally created with an earlier version of Cassandra that had a bug documented in JIRA 703. This problem does not exist with super column rows created / deleted / recreated since 0.6.

      3. If your app does some complex and/or critical multi-row data manipulation like us, you are going to need some kind of centralized locking system for case where race conditions could screw things up. We are using ZooKeeper for this purpose, and I plan to open source our locking and synchronization primitives library for general use asap. HBase offers single row locking, and multi-row locking is planned, but for a whole load of reasons I believe this is much better handled by ZooKeeper where/if this is needed. Once our library is out, this won’t be something people have to worry about because managing a ZooKeeper cluster is also very straightforward.


      April 7, 2010 at 6:32 pm

      • Thanks for the details. When is FightMyMonster going live? Also… where in London you guys based?


        April 7, 2010 at 11:20 pm

      • We are doing development work in Queens Park area, but FMM is actually a Swiss operation. Not clear yet when we will be ready to go live, but product is beginning to look v.good so hopefully soon.


        April 8, 2010 at 10:48 pm

  25. […] HBase vs Cassandra In my opinion, these differing histories have resulted in HBase being more suitable for data warehousing, and large scale data processing and analysis (for example, such as that involved when indexing the Web) and Cassandra being more suitable for real time transaction processing and the serving of interactive data. Writing a proper study of that hypothesis is well beyond this post, but I believe you will be able to detect this theme recurring when considering the databases. via […]

  26. […] HBase vs Cassandra: why we moved: interesting background on a cassandra choice View full post on Jeremy Zawodny’s linkblog […]

  27. Did you pick a persistence engine such as DataNucleus or OCM? Or did you roll your own? We’ve moved to Cassandra as well, but there don’t seem to be any mature object persistence frameworks written for it.


    April 23, 2010 at 1:25 am

    • Hi, we’re not using an object persistence engine. If we want to store a whole object into a Cassandra column value, we just serialize/deserialize it using XStream. But I guess you are looking for something more heavyweight.


      April 23, 2010 at 11:19 am

  28. […] to comments So if you are looking for a good NoSQL read of HBase vs. Cassandra you can check out  In short HBase is good for reads and Cassandra for writes.  Cassandra does a great job on reads […]

  29. Your words are persuasive, but our system have to use the map/reduce, so hbase is still our preference!

    Wang Xiaojun

    May 6, 2010 at 3:01 am

  30. jaybae의 생각…

    아니 카산드라가 그렇게 좋아?…

    jaybae's me2DAY

    May 6, 2010 at 9:01 am

  31.’s done it once more. Amazing post.

    Luke Zimmerman

    May 28, 2010 at 9:19 am

  32. Incredibly awesome post. Really!

    Dionne Rizzo

    May 29, 2010 at 11:26 pm

  33. Thanks for writing this! Very valuable!

    Robert McTupper

    June 11, 2010 at 8:30 pm

  34. […] 英文原文地址: […]

  35. […] HBase -> Cassandra migration experience and some CAP context, excellent read […]

  36. Thanks for the experience report and pointer to the ycsb paper.

    Architecturally, HBase is uniquely situated to provide both:

    1. Low Latency random access
    2. High Throughput batch streaming with Hadoop

    These two performance extremes are difficult to achieve in a distributed system that allows record updating and hardware expansion without downtime. Both of these extremes are important when the goal is to have the same data used for both an interactive service and batch analysis.

    If both extremes cannot be achieved, the alternative is to snapshot the data of the interactive system into another system for streaming processing. The snapshots are reminiscent of database replication that can get far behind when something goes wrong. Facebook probably uses Scribe to copy changes to their Hive/Hadoop cluster for their analytics.

    It is true that setting up HBase is not easy when compared to a P2P style architecture such as Cassandra, but if you already have a Hadoop cluster, then you are halfway there.

    Perhaps one day Hadoop, HBase, and Zookeeper could use the wonderful gossip algorithm for self-organizing.

    Paul Baclace

    November 13, 2010 at 5:07 am

  37. […] HBase vs Cassandra: why we moved (Chinese Translation) […]

  38. Are you still happy with your move to Cassandra? You site Twitter and Facebook as big users of Cassandra but they have both moved to HBase now. The momentum seems to favor HBase at this point. I would gather most large distributed systems will want to run analytics on the data. With HBase you can do this using Hadoop map-reduce natively. There is no need to transfer data in and out. This was the deal breaker for me with Cassandra and similar systems.


    December 6, 2010 at 4:00 am

    • Hi, we are still happy with our move to Cassandra, but then our requirements are different. We have a relatively small budget and team, but big scalability requirements. We simply don’t have the technical bandwidth to manage HBase at the moment, and nobody should be under any illusions, maintaining an HBase/Hadoop cluster does require *much* more work and expertise. That said this is a big win for HBase, at least as concerns absolutely huge systems where it is possible to field a well resourced dedicated technical team, but the full picture about why this choice was made is not available yet. We require locking and consistency in much of our system and are doing this over Cassandra using ZooKeeper/Cages and writing/reading where necessary with QUORUM (and without this things quickly go wrong, as we experienced at the beginning before we fixed some bugs in Cages). So far, neither Cassandra nor ZooKeeper have proven the bottleneck in our system (ZooKeeper/Cages in particular was predicted to cause trouble and hasn’t), it is actually the application layer where most data processing takes place that we are now having to examine ways of optimizing. It will be interesting as we scale out properly next year to see how Cassandra fares, but while we are aiming for several M users we are never going to have the load that Facebook does in its messaging system.


      December 14, 2010 at 11:40 am

  39. When would you use Riak, HyperTable or MongoDB?…

    If you are looking for general insights, this blog post was a good read on Cassandra, HBase and NoSQL in general…


    December 22, 2010 at 8:05 pm

  40. […] Холивар Почему я выбрал Cassandra вместо HBase Почему я выбрал HBase вместо Cassandra […]

  41. by the way.. facebook’s realtime messaging system is now running on hbase: not Cassandra the article is an interesting read and offers another perspective on this topic.

    Arvind K

    February 25, 2011 at 6:15 am

  42. […] My team is currently working on a brand new product – the forthcoming MMO This has given us the luxury of building against a NOSQL database, which means we can put the horrors of MySQL sharding and expensive scalability behind us. Recently a few people have been asking why we seem to have changed our preference from HBase to Cassandra. I can confirm the change is true and that we have in fact almost completed porting our c … Read More […]

  43. Great article, however I perceived it to be a little biased in favor of Cassandra, rather than a balanced comparison.

    Thank you for the great information tho.

    Asad Hasan

    March 20, 2011 at 7:35 am

  44. The simplicity of adding a new node (aka “Elastic Speedup” in the YCSB paper) to a Cassandra system is definitely a strength. However, the YCSB paper noted that after adding a new node “… In fact, under load, it takes many hours for Cassandra to stabilize.”, which is evident in Figure 7(a) of the paper. Have you experienced the same behavior in your implementation?


    April 1, 2011 at 6:10 am

    • Hi this isn’t a specific thing we’ve had issues with but it is true that hinted handoffs, ops that move data between nodes etc are CPU intensive. For example, just try turning off one of your nodes for while and then turning it back on – you’ll seen the load ramp up as it receives the data it missed while down (btw we did have a real problem with this, albeit one of our own making, described in a recent post). The problem with reports like the YCSB is that the software itself is evolving rapidly. After upgrading from 0.7.0 to 0.7.4 we’ve noticed it seems to be using memory quite a bit better. Furthermore, things like the replication factor you use will also make a big difference (for the record, we use RF 3 so everything continues as before when a node goes down). What I can report is that the performance we’re getting with Cassandra is really good and have no complaints. Personally I haven’t had a chance to properly critique the node splitting algorithms used by Cassandra, but I will say that a system we use internally called Starburst uses consistent hashing to distribute addressable network resources (such as connected users). That’s a really nice way of redistributing resource in a way that all nodes play an equal part and if I had to nitpick, albet from a position of ignorance, I would like to see a system more like that.


      April 1, 2011 at 9:12 am

  45. […] for us, juggling with Consistency, Availability and Partition tolerance is not as important as making sure that data is 100% […]

  46. Thanks for taking the time to write this. Very helpful.


    May 13, 2011 at 8:58 pm

  47. Hi,

    If you have the chance to make this decision today, do you consider mongodb ? Can you please give some details why mongo is not ideal for this project ?


    Serdarou have the chance to

    June 3, 2011 at 10:12 pm

    • Hi I am not so familiar with MongoDb, but the big advantage of Cassandra is that all nodes have equal status. That makes it much easier to scale a cluster, because you can just add nodes without worrying about giving them special roles, adjusting the roles of existing nodes etc.


      June 6, 2011 at 4:32 pm

    • (1) mongodb has no roll-forward-log reliability… if a node goes down uncleanly for any reason you can’t trust the data on it anymore and must rebuild the node from a copy. This is a really tenuous situation to be in. A single sitewide power-outage can leave all data in an untrustworthy state. Better have good backups.

      (2) mongodb is a heap + b-tree, and thus is not write-optimized like hbase and cassandra. As working-set exceeds memory-size, random write performance of mongodb will be closer to innodb/mysql than hbase/cassandra.

      (3) mongodb’s heap + b-tree index means that a sorted scan retrieval by any index requires a seek-per-record. If the dataset is larger than memory size, this can be much more expensive than mysql-innodb primary-key ordered data, or hbase key-ordered data. (the state of cassandra’s ordered-scan has been so questionable I don’t include it)

      David Jeske

      June 7, 2011 at 8:30 pm

      • thank you very much, great information.

        Serdar Irmak

        June 10, 2011 at 7:03 am

  48. […] HBase vs Cassandra […]

  49. Thanks for this information.
    Two issues concern me in this regards:
    1) Well, we are now in the mid of 2011, and the “trend” is no longer “Facebook == Cassandra” as Facebook decided to “divorce” from Cassandra and switch to… you guessed – HBASE. It’s almost like a slap in the face to the Trend claim.

    I would love to get your comment on that – how do you perceive their brave (?) decision, what was wrong with Cassandra for them etc’.
    I read their blogs and FB Engineering pages but seems they willing to say more what they liked in HBase that suits them than discover what was missing or went wrong in Cassandra for their type of usage (of-course).

    2) Regarding the flexibility of Cassandra to play with CAP, if I choose full consistency(“ALL”), I lose all Cassandra benefits, so it is not a proper claim in my opinion to say that it is not black and white and one can choose. If you chose for some of your data to be fully consistent, you do a terrible decision to put it in Cassandra database. Am I wrong?

    Tomer Paz

    June 14, 2011 at 7:59 pm

    • I’ll disclaim that I have not used Cassandra, so someone who has actual experience may be more informed. However, from my investigation, these may be two Cassandra vs hbase differences that are worth considering.

      1) If you replicate 3 ways, and you wish your data to be kept in memory, Cassandra quorum consistency will require it to take memory on 3 machines. Hbase uses a different model where there is a ‘single master’ for the memory copy which is always consistent but is not fully available. This means hbase consumes 1/3 as much memory for an in-memory dataset, with the downside that the data may briefly become unavailable. There was a Cassandra plan to do something to have a ‘memory-master’, but i don’t know the current state of it. As a another note…. waiting for an in-memory response from two of three servers will be slower on average than an in-memory response from a single server.

      2) range queries seem more robust on hbase. There may be Cassandra issues with slice/range queries and quorum consistency. Rather than try to explain it, you can read the bug…

      There used to be an issue with read-repair not honoring quorum consistency, but reading up on the bugs it looks like it may be fixed now.

      David Jeske

      June 15, 2011 at 6:27 am

      • The relative performance implications of the Cassandra vs HBase architectures is a complex subject that really needs a proper paper – if there’s a distributed computing Phd student out there this is a call to action! It’s very difficult to draw conclusions from individual points. For example, while its true that, for example, a RF3 Cassandra cluster will block until 2 nodes have written with a QUORUM consistency level, these writes are very fast because the commit log (write ahead logs) are stored on the actual nodes themselves. This contrasts with HBase, where, as I understand, the region servers which are solely responsible for ranges of rows have to store their commit logs on the DFS for durability. The few comparative speed tests out there seem to reflect that in fact Cassandra is much faster for writes. On the other hand, although I don’t think the test results actually reflect this, Cassandra has been heavily optimized for writes rather than reads. Not only will a greater number of nodes hold the hot data set in memory (at least in say a RF3 cluster like we use) but the process pursued for retrieval of data is more complex than in HBase and in the common case of say QUORUM reads the data sets returned by different nodes must be resolved before being returned to the client. Therefore when an organization compares the relative suitability of HBase and Cassandra, they need to consider whether their application is write or read heavy. Unfortunately there are no simple answers.


        June 15, 2011 at 8:58 am

      • I can shed a little light here.

        1) Your reasoning is correct, although Cassandra will optimistically optimize down to just the minimum reads if read repair is off (two for quorum-of-three). The majority of use cases are fine with ConsistencyLevel.ONE, though, and it’s worth noting that Cassandra is roughly 3x faster at in-memory reads than HBase (according to the YCSB benchmark on current releases of each as of a couple months ago) at ConsistencyLevel.ALL — even faster if you drop down to QUORUM or ONE.

        2) Range queries are well-tested in Cassandra at this point and widely used by Hadoop/Pig/Hive queries as well as manually. (The linked issue has to do with “slicing” to page through a single very large row, not range queries across multiple rows.)

        Jonathan Ellis

        July 6, 2011 at 7:24 pm

    • Hi the reality is that both Cassandra and HBase now have loyal followers and numerous production clusters running. Detailed technical information about the specific applications of NoSQL at Facebook and Twitter aren’t really available atm so it is too early to draw inferences about the relative systems, although it is worth noting that both those organizations have huge resources and therefore certain advantages offered by Cassandra – namely simpler cluster setup and operation – would prove less of a draw.
      I think it is a fair point that using Cassandra introduces issues related to consistency that many engineers will be unfamiliar with and may not fully appreciate until they have worked with the database for some time. There are ways of dealing with these and in fact we rely on heavily serialized consistent operations for which we use distributed locking and QUORUM writes (we have RF3) based upon ZooKeeper where needed. Neither system has reached a major version number and I think a clearer picture will emerge when they do and the supporting open source application architectures enter wider user (for example, we rely on the open source systems at some of which have yet to be released).


      June 15, 2011 at 8:31 am

  50. […] View FightMyMonster’s Hbase vs Cassandra: why we moved Blog Post […]

  51. Dear Dominic Williams
    I have read your post and I liked a lot, sorry for my english, because I’m still learning.
    I’m searching and writing an article for graduate about NoSQL and I’d like to know if you have or if you know a free book to send to me, because here in Brazil doesn’t have anyone and I need one immediately.
    My email is:

    Hugs and thanks man!

    • Hi Bruno checkout the documentation from the website, and also the presentations from the last Cassandra conference in San Francisco.


      August 17, 2011 at 9:01 am

  52. […] HBase vs Cassandra: why we moved « Dominic Williams (tags: cassandra hbase nosql bigdata) […]

  53. Brilliant article! Clearly differentiates the differences between Hbase and Cassandra….by the way at Ooyala we use a combination of Hadoop and Cassandra for dynamic realtime user analytics

    Joseph Oliver

    October 7, 2011 at 10:01 pm

  54. thanx for the post ! although i’m all new to all the distributed processing and NoSQL thing, i find your comparison very informative and close to a new-in-the-field developer.

    Mohammad Habbab

    December 26, 2011 at 6:57 pm

  55. one of the best article on HBase and Cassandra.. (any clue why Facebook using HBase over Cassandra..)

    viji (@vijiconnect)

    April 4, 2012 at 4:38 am

    • Many of these choices come down to idiosyncrasies of individual teams, situations and engineers at the time decisions were made, but here are some thoughts: (i) at the time the decision was made Cassandra was a much less mature database that HBase, a situation that has since changed, and (ii) Facebook has hundreds of engineers, and the additional administrative overhead of running a large HBase + HDFS installation won’t have looked so bad, especially considering they were already expending resources to manage their sharded MySQL setup, which is the ultimate design/coding and admin time eater.


      April 4, 2012 at 10:28 am

  56. I have read in this article that it is difficult to install HBase. I think it is no longer true. There is an open source installer with Hadoop, HBase, Zookeeper and Flume at

    Ladislav Urban

    May 10, 2012 at 5:20 pm

  57. I am revisiting HBase vs. Cassandra for a client. HBase is having it’s first summit on May 22, 2012, so expect some announcements there.

    Here are some notes that update discussion as of May 12 2012:
    * Composite keys and ordering of records is now standard (just like HBase)
    ** this enables efficient parallel processing at the Mapper input stage for map-reduce. Although it would not be as fast as map-reduce directly over hdfs, it would be similar to a Mapper stage reading from HBase RegionServers
    * super columns are deprecated (good thing)
    * atomic counters were added
    * Both have TTL, time to live metadata, which means self-cleaning (yeah).
    * When FB switched from Cassandra to HBase, they said the change was driven by Cassandra team members leaving (perhaps going to DataStax) and also because eventual consistency was a problem in Cassandra.
    * So far, Cassandra is easier to administer since they have a funded support company (DataStax) that gives out plenty of free advice.
    ** HBase might catch up eventually; big-table technology is proven to work fine for Google.
    * Michael Stack has often said HBase is good for TB sized data and larger.
    ** For me, setting up HBase for 100GB on 4 nodes certainly seems harder than necessary.
    ** A P2P installation mechanism for HBase is not yet available (I continue to nudge).
    * Any distributed store can have performance killing hotspots (all activity on one or 2 nodes) if keys are not chosen wisely.
    * Under a heavy write and insert load ANY storage mechanism (distributed or not) can get behind on compacting/splitting/re-org.
    ** This is a matter of admin discipline, there must be headroom for periodic cleanup/reorg (when Scotty says the warp drives cannot go faster without later down time, he means it.)
    ** The safer approach is to separate batch and interactive operations.
    * Both have commit logs and can detect/repair disk errors (over-write, wrong sector written, etc.)
    * Cassandra deployments split across data centers has actually been attempted by FB, but I have not seen any experience reports on how well that works.
    * Cassandra only supports the Thrift interface which does not have a way to stream large objects.

    Paul Baclace

    May 13, 2012 at 1:30 am

    • That’s a pretty good list. I’d add a few:

      * Cassandra supports 1000s of ColumnFamilies in a cluster; HBase supports a handful (HBase book says “two or three” but let’s be generous)
      * Cassandra’s p2p design gives you “read slaves” for free; HBase reads and writes all go through a single regionserver per region
      * Cassandra offers built-in caching; HBase installations tend to rely heavily on memcached
      * Cassandra is far, far ahead of HBase in dealing with compaction; this includes parallel (multiple threads per compaction) and concurrent (multiple simultaneous compactions per server), throttling to avoid latency spikes, “leveled” compaction to optimize for reads, and the ability to remove deleted or expired data during minor compactions as well as major
      * Cassandra’s p2p design provides full availability, as well as the multi-datacenter capabilities you mentioned. HBase relies on the HDFS namenode SPOF as well as regionserver mini-SPOFs
      * Multi-DC can also solve the batch/interactive conflict you mention; best practice for Cassandra is to create a “virtual DC” for your Hadoop jobs and let Cassandra replicate between that and your interactive “DC,” bidirectionally
      * Cassandra manages its own local storage per node, which allows you to (for instance) dedicate a spindle to your commitlog, or pin certain columnfamilies to SSDs


      May 16, 2012 at 7:38 pm

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: