HBase in Production at Facebook – Jonathan Gray at Hadoop World 2010

Interesting presentation from Facebook’s Jonathan Gray at Hadoop World 2010 on Facebook’s current and future plans for using HBase in their data platform. Here’s the video:

A couple of slides that position the role of HBase within Facebook. First, Facebook’s core application/data platform – a LAMP stack with Memcache and Hadoop/Hive:

And then a slide that hints at the impact HBase has on various elements of the stack:

Note that HBase does not actually replace any of these element of the stack, but rather plays an interesting intermediate role between online transactional and offline batch data processing.

In his presentation, Gray speaks to the advantages of HBase. Here’s a few snippets:

And then on the Data Analysis side, HBase doesn’t actually do data analysis. And it doesn’t actually store data. But HDFS stores data, and Hive does the analysis. But with HBase in the middle you can do random access and you can do incremental updating.

You also have fast Index writes. HBase is a sorted store. So every single table is sorted by row, every single row is sorted by the column. And then columns are sorted by versions. That’s a really powerful thing that you can build inverted search indexes on, you can build secondary indexes on. So you can do a lot more that just what you can do with a key-value store. So it has a very powerful data structure.

And lastly, there’s real tight integration with Hadoop. And my favorite thing: It’s an interesting product that kind of bridges this gap between the online world and the offline world – the serving, transactional world and the offline, batch-processing world.

HBase Use Case #1: Near real-time Incremental updates to the Data Warehouse

Says Gray:

Right now (at Facebook), we’re doing night updates of UDBs into the data warehouse. And the reason we’re doing that is because HDFS doesn’t have incremental operations. I can only append to something. I can’t edit something, I can’t delete something. So merging in the changes of transactional data, you basically have to rewrite the entire thing.

But with HBase, what we’re able to do is, all of our MySQL data is already being replicated. So we already have existing replication streams. So we can actually hook directly into those replication streams, and then write them into HBase. And then HBase then allows us to expose Hive tables, so we can actually have completely up-to-date UBD data in the datawarehouse. So now we can have UDB data into our data warehouse in minutes, rather than in hours or a day.

HBase Use Case #2: High Frequency Counters and Real-time Analytics

Again quoting Gray:

The second use case is around high-frequency counters, and then real-time analytics of those counters. This is something I think a lot of people have used HBase for for a long time. …

It’s a really interesting use case. Counters aren’t writes, they’re read-modify-writes. So they’re actually a real expensive operation. In a relational database, I’d actually have to do a read lookup, and then write that data back. So it’s a real expensive thing. And if you’re talking about billions of counter updates per hour – or I think right now on one of our clusters it’s about 100,000 updates per second. So doing a 100,000 increments a second on a SQL machine, it’s a cluster of machines now – a lot of machines.

And then the other part is, well now that I’m taking all this increment data, I want to be able to do analysis on it. If I’m taking click-stream data, I want to say What’s the most popular link today? And this past hour? And of all time? So I want to be able to do all that stuff, and if I have all my data sitting in MySQL or HDFS, it’s not necessarily very efficient to compute these things.

So the way we do it now is Scribe Logs. So everytime you click on something, for example, that’s going into Scribe as a log line saying this user clicked this link. That’s being fed into HDFS. And then periodically we’re saying, OK once an hour, or once a day or whatever, let’s take all of our click data and do some analysis on it. Let’s do sums and max-mins and averages and group-by, and different kinds of queries like that. And then once we have our computations, let’s feed it back into the UDBs so people can read it.

So looking at this flow here, we have Scribe going downline into HDFS. And then once we’re in HDFS, we’re writing things as Hive tables so we can get the dimensions that we need. And then we’re doing these huge MapReduce joins to join everything by URL. So it takes a long time to do that job. It’s really, really inefficient. It uses lots and lots of I/O. And it’s not real-time. If this job takes an hour to run, we’ll always have at least an hour of stale data.

But with HBase what we’re doing is we’re going from Scribe directly into HBase. Which means that as soon as that edit comes into HBase, it’s available. You can read it. You can randomly read it. You can do MapReduce on it. You can do whatever you want. And like I was talking about before, you can do real-time reads of it. And I could say “How many clicks have there been for newyorktimes.com today, and I can just grab that data out of HBase.

Or, if I want to do things like “What are the top 10 links across all domains?” Well, the way we do that is kind of like through a trigger-based system. Because the increments are so efficient, I can actually increment 10 things for each increment. So I can say “Increment this domain. But then increment the link. Increment it for today. Increment it for these demographics.” Just do a whole bunch of increments because the increments are so efficient.

So you can actually pre-compute a lot of this stuff. And then when you want to do big aggregations, you can do a MapReduce directly on HBase. And then when you’re done and you have your results, rather than having to feed them back to the UDBs, you just put them in HBase and they’re there.

So it’s really, really cool – storage, serving, analysis as one system. And we’re able to basically keep up with huge, huge numbers of increments and operations, and at the same time do analytics on it.

HBase Use Case #3: User-facing Database for Write-intensive workloads

And the third scenario:

The last case I want to talk about … is using HBase as a user-facing database – almost as a transactional database, and specifically for Write workloads. When you have lots and lots of Writes and very few Reads, or you have a huge amount of data liked I talked about before. If I’m storing 500K, I don’t necessarily want to put that in my UDB. …

I’m not going to elaborate further on this use case. Please listen to the presentation for a full discussion.

HBase and Hive Integration

On production development of HBase at Facebook, Gray has this to say:

But the first thing we did was the Hive integration. … This unlocks a whole new potential, and not just the way I was describing it earlier that we can now randomly write into our data warehouse. You can also randomly read into the data warehouse. So for certain kinds of joins, for example, rather than having to stream the joins we can actually do point lookups into HBase tables. So it unlocks a whole new bunch of ways that we can potentially optimize Hive queries.

But the base of Hive integration is really HBase tables become Hive tables. So you can map Hive tables into HBase. You can use that then as an ETL data target, meaning that we can write our data into it. It can also be a Query data source so we can read data from it. And like I was saying, the Hive integration supports different read and write patterns.

So on the Write side, it supports API random writing like we would do with UDBs. It also supports this bulk load facility through something called HFile output format. So HFile is the ondisk format that HBase uses. And it just looks like a sequence file or a Map file or anything else, but it has some special facilities for HBase.

And we extensively are doing this, which is taking data, writing it out as HFiles, which basically means you’re writing into HBase at the same speed you write to HDFS. And then you just kind of hit a button, and HBase loads all those files in. And now you have really efficient random access to all that data. We’re using that extensively.

Also, on the Read side. You can randomly read into stuff. Or you can do full table scans. Or you can do range scans. All that kind of stuff through Hive.

In Summary

Another great presentation from Hadoop World.

glenn

Facebook’s Architectural Stack – designing for Big Data

This is my fourth of a series of posts exploring the topic of Big Data. The previous posts in this series are:

• Designing Large-Scale Retrieval Systems at Google – Jeff Dean from 2009
• Introduction to Hadoop – understanding Big Data
• Keen Insight on Big Data from Cloudera CEO Mike Olson

This post provides two videos in which Facebook’s David Recordon discusses Facebook’s architectural stack as a platform that must scale to massive amounts of data and traffic. The first video is a short video where Recordon discusses Facebook use of the LAMP stack at OSCON 2010:

On Database Technology and NoSQL Databases at Facebook

In the first video, Recordon first addresses how Facebook implements database technology generally, and the topic of NoSQL databases. Says Recordon:

The primary way that we store data – all of our user data that you’re going and accessing when we’re working on the sight, with the exception of some services like newsfeed – is actually stored in MySQL.

So we run thousands of nodes of a MySQL cluster – but we largely don’t care that MySQL is a relational database. We generally don’t use it for joins. We’re not going and running complex queries that are pulling multiple tables together inside a database using views or anything like that.

But the fundamental idea of a relational database from the ’70s hasn’t gone away. You still need those different components.

Recordon says that there are really three different layers Facebook thinks about when working with data, illustrated in the following visual:

Continues Recordon:

You have the database, which is your primary data store. We use MySQL because it’s extremely reliable. [Then you have] Memcache and our web servers.

So we’re going and getting the data from our database. We’re actually using our web server to combine the data and do joins. And this is some of where HipHop becomes so important, because our web server code is fairly CPU-intensive because we’re going and doing all these different sorts of things with data.

And then we use Memcache as our distributed secondary index.

These are all the components that you would traditionally use a relational database for:

Recordon continues:

[These are the same layers that were] talked about 30-40 years ago in terms of database technology, but they’re just happening in different places.

And so whether you’re going and using MySQL, or whether you’re using a NoSQL database, you’re not getting away from the fact that you have to go and combine data together, that you’re needing to have a way to look it up quickly, or any of those things that you would traditionally use a database for.

On the topic of NoSQL databases, Recordon says:

And then when you dig into the NoSQL technology stack, there are a number of different families of NoSQL databases which you can go and use. You have document stores, you have column family stores, you have graph databases, you have key-value pair databases.

And so the first question that you really have is what problem am I trying to solve, and what family of SQL database do I want to go and use.

And then even when you dig into one of these categories – if we just go and look at Cassandra and HBase – there are a number of differences inside of this one category of database. Cassandra and HBase make a number of different tradeoffs from a consistency perspective, from a relationship perspective. And so overall you really go and think about what problem am I trying to solve; how can I pick the best database to do that, and use it.

While we store the majority of our user data inside of SQL, we have about 150 terabytes of data inside Cassandra, which we use for Inbox search on the site. And over 36 petabytes of uncompressed data in Hadoop overall.

On the topic of Big Data

Recordon:

So that leads me into Big Data. We run a Hadoop cluster with a little over 2,200 servers, about 23,000 CPU cores inside of it. And we’ve seen the amount of data which we go and store and process growing rapidly – it’s increased about 70 times over the past 2 years. And by the end of the year, we expect to be storing over 50 petabytes of uncompressed information – which is more than all the works of mankind combined together.

And I think this is really both the combination of the increase in terms of user activity on Facebook … But also just in terms of how important data analysis has become to running large, successful websites.

The diagram below shows Facebook’s Big Data infrastructure:

Says Recordon:

So this is the infrastructure which we use. I’ll take a minute to walk through it.

With all our web servers we use an open source technology we created called Scribe to go and take the data from tens of thousands of web servers, and funnel them into HDFS and into our Hadoop warehouses. The problem that we originally ran into was too many web servers going and trying to send data to one place. And so Scribe really tries to break it out into a series of funnels collecting this data over time.

This data is pushed into our Platinum Hadoop Cluster about every 5-to-15 minutes. And then we’re also going and pulling in data from our MySQL clusters on about a daily basis. Our Platinum Hadoop Cluster is really what is vital to the business. It is the cluster where if it goes down, it directly affects the business. It’s highly maintained, it’s highly monitored. Every query that’s being run across it, a lot of thought has gone into it.

We also then go and replicate this data to a second cluster which we call the Silver Cluster – which is where people can go and run ad-hoc queries. We have about 300 to 400 people which are going running Hadoop and Hive jobs every single month, many of them outside of engineering. We’ve tried to make this sort of data analysis to help people throughout the company make better product decisions really accessible.

And so that’s one of the other technologies which we use, Apache Hive, which gives you an SQL interface on top of Hadoop to go and do data analysis. And all of these components are open source.

So when Facebook thinks about how there stack has evolved over the past few years, it looks something like this:

Where the major new component is the Hadoop technology stack and its related components to manage massive amounts of data, and do data analysis over top of that data.

A deeper look at Scaling challenges at Facebook

The second video is a presentation delivered by David Recordon and Scott MacVicar – both Facebook software engineers – at FOSDEM in February 2010 provides a deeper look into Facebook’s use of open source technology to provide a massively scalable infrastructure:

The question that I am interested in, and isn’t answered in these videos, is how Facebook implements its Open Graph data model in its infrastructure. That would be very interesting to learn. For more specifically about Facebook’s Open Graph technology, please see Facebook’s Open Graph and the Semantic Web – from Facebook F8.

Very interesting stuff.

glenn