After GFS and MapReduce, Google solve again big data problems designing BigTable: a compressed, high performance, and proprietary data storage system that forms the basis for most of its projects. HBase and Cassandra are inspired from it.


Title: Bigtable: A Distributed Storage System for Structured Data (PDF), November 2006
Authors: Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E. Gruber

Bigtable is a distributed storage system for managing structured data that is designed to scale to a very large size: petabytes of data across thousands of commodity servers.

Many projects at Google store data in Bigtable, including web indexing, Google Earth, and Google Finance. These applications place very different demands on Bigtable, both in terms of data size (from URLs to web pages to satellite imagery) and latency requirements (from backend bulk processing to real-time data serving). Despite these varied demands, Bigtable has successfully provided a flexible, high-performance solution for all of these Google products.

In this paper we describe the simple data model provided by Bigtable, which gives clients dynamic control over data layout and format, and we describe the design and implementation of Bigtable.

Check out the list of interesting papers and projects (Github).

The big-data environment at the moment is really “collaborative”. Each project is ready to run on almost every available platform and this is good. However, recently two factions are forming: people who use the Hadoop 2.0 Stack and people who use the BDAS.

Hadoop 2.0 Stack


The most important difference between Hadoop 2.0 and previous versions is YARN, the new cluster resource manager and next generation MapReduce. It can run almost every kind of big-data project:

  • Traditional Map Reduce (new version is backward compatible) with Hive or Pig or Cascading for query.
  • Interactive near real-time Map Reduce (using Tez)
  • HBase and Accumulo
  • Storm and S4 for stream processing
  • Giraph for graph processing
  • OpenMPI for message passing
  • Spark as In-memory Map Reduce
  • HDFS as distributed filesystem
  • more…

Most interesting companies here are IntelCloudera, MapR and Hortonworks.

BDAS (Berkely Data Analytics Stack)


On the BDAS everything is built around Mesos: the cluster resource manager. It’a relative new project is already widely used. Traditional HDFS is accelerated by Tachyon (in-memory file system). The main integration is around Spark which is the base for:

Mesos can also run traditional Hadoop environment and other projects (such as Storm and OpenMPI). You can also run traditional applications (also Rails apps) using Marathon.

The most interesting companies here are Databricks and Mesosphere.

Who will win? 😀


I always underestimated the contribute of Google to the evolution of big-data processing. I used to think that Google only manages and shows some search results.  Not so much data. Not so much as Facebook or Twitter at least…

Obviously I was wrong. Google has to manage a HUGE amount of data and big-data processing was already a problem on 2002! Its contribution to the current processing technologies such as Hadoop and its filesystem HDFS and HBase was fundamental.

We can split contribution into two periods. The first of these (from 2003 to 2008) influenced technologies we are using today. The second (from 2009 since today) is influencing product we are going to use is the near future.

The first period gave us

  • GFS Google FileSystem (PDF paper), a scalable distributed file system for large distributed data-intensive applications which later inspire HDFS
  • BigTable (PDF paper), a columnar oriented database designed to store petabyte of data across large clusters which later inspire HBase
  • the concept of MapReduce (PDF paper), a programming model to process large datasets distributed across large cluster. Hadoop implements this programming model over the HDFS or similar filesystems.

This series of paper revolutionized the strategies behind data warehouse and now all the largest companies uses products, inspired by these papers, we all knows.

The second period is less popular at the moment. Google faced many limits in its previous infrastructure and tried to fix them and move ahead. This behavior gave as many other technologies, some of these not yet completely public:

  • Caffeine, a new search infrastructure who use GFS2, next-generation MapReduce and next-generation BigTable.
  • Colossus, formerly known as Google FileSystem 2 the next generation GFS.
  • Spanner (PDF paper), a scalable, multi-version, globally-distributed, and synchronously-replicated database, the NewSQL evolution of BigTable
  • Dremel (PDF paper), a scalable, near-real-time ad-hoc query system for analysis of read-only nested data, and its implementation for the GAEBigQuery.
  • Percolator (PDF paper), a platform for incremental processing which continually update the search index.
  • Pregel (PDF paper), a system for large-scale graph processing similar to MapReduce for columnar data.

Now market is different than 2002. Many companies such Cloudera and MapR are working hard for big-data and Apache Foundation as well. Anyway Google has 10 years of advantages and its technologies are still stunning.

Probably many of these papers are going to influence the next 10 year. First results are already here. Apache Drill and Cloudera Impala implement the Dremel paper specification, Apache Giraph implements the Pregel one and HBase Coprocessor the Percolator one.

And they are just some examples, a Google search can show you more 😉


In previous post I analyzed Facebook main infrastructure. Now I’m going deeper into services.

Facebook Images

Facebook is the biggest photo sharing service in the world and grows by several millions of images every week. Pre-2009 infrastructure uses three NFS tier. Also with some optimization this solution can’t easily scale over a few billions of images.

So in 2009 Facebook develop Haystack, an HTTP based photo server. It is composed by 5 layers: HTTP server, Photo Store, Haystack Object Store, Filesystem and Storage.

Storage is made on storage blades using a RAID-6 configuration who provides adequate redundancy and excellent read performance. The poor write performance is partially mitigated by the RAID controller NVRAM write-back cache. Filesystem used is XFS and manage only storage-blade-local files, no NFS is used.

Haystack Object Store is a simple log structured (append-only) object store containing needles representing the stored objects. A Haystack consists of two files:

the actual haystack store file containing the needles


plus an index file


Photo Store server is responsible for accepting HTTP requests and translating them to the corresponding Haystack store operations. It keeps an in-memory index of all photo offsets in the haystack store file. The HTTP framework we use is the simple evhttp server provided with the open source libevent library.

Insights and Sources

Facebook Messages and Chat

Facebook messaging system is powered by a system called Cell. The entire messaging system (email, SMS, Facebook Chat, and the Facebook Inbox) is divided into cells, and each cell contains only a subset of users. Every Cell is composed by a cluster of application server (where business logic is defined) monitored by different ZooKeeper instances.

Application servers use a data acces layer to communicate with metadata storage, an HBase based system (old messaging infrastructure relied on Cassandra) which contains all the informations related to messages and users.

Cells are the “core” of the system. To connect them to the frontend there are different “entry points”. An MTA proxy parses mail and redirect data to the correct application. Emails are stored in the same structure than photos: Haystack. There are also discovering service to map user-to-Cell (based on hashing) and service-to-Cell (based on ZooKeeper notifications) and everything expose an API.

There is a “dirty” cache based on Memcached to serve messages (from a local cache of datacenter) and social information about the users (like social indexes).


The search engine for messages is built using an inverted index stored in HBase.

Chat is based on an Epoll server developed in Erlang and accessed using Thrift and there is also a subsystem for logging chat messages (in C++). Both subsystems are clustered and partitioned for reliability and efficient failover.

Real-time presence notification is the most resource-intensive operation performed (not sending messages): keeping each online user aware of the online-idle-offline states of their friends. Real-time messaging is done using a variation of Comet, specifically XHR long polling, and/or BOSH.

Insights and Sources

Facebook Search

Original Facebook search engine simply searches into cached users informations: friends list, like lists and so on. Typeahead search (the search box on the top of Facebook frontend) came on 2009. It try to suggest you most interesting results. Performances are really important and results must be available within 100ms. It has to be fast and scalable and the structure of the system is built as follow:


First attempts are still on browser cache where are stored informations about user (friends, like, pages). If cache misses starts and AJAX request.

Many Leaf services search for results inside theirs indexes (stored into an inverted index called Unicorn). When results references are fetched from the indexes, they are merged and loaded from global datastore. An aggregator provide a single channel to send data to client. Obviously query are cached.

On 2012 Facebook starts from the core part of typeahead search to build a new search tool. Unicorn is the core part of the new Graph Search. Formally is an in-memory inverted index which maps Facebook contents as a graph database would do and you can query it as graph traversing tool. To be used for Graph SearchUnicorn was updated to be more than a traversing tool. Now supports nested queries, scoring and support for different kind of resources. Results are aggregated on different levels.


Query lifecycle is usually made by 2 steps: the Suggestion Phase and the Search Phase.

The Suggestion Phase works like “autocomplete” and Is powered by a Natural Language Processing (NLP) module attempts to parse text based on a grammar. It identifies parts of the query as potential entities and passes these parts down to Unicorn to search for them.

The Search Phase begins when the searcher has made a selection from the suggestions. The parse tree, along with the fbids of the matched entities, is sent back to the Top Aggregator. A user readable version of this query is displayed as part of the URL.

Currently Graph Search is still in beta.

Insights and Sources

Resources and insights

This post and the previous one are based and inspired by the answer of Michaël Figuière to the following question on Quora: What is Facebook’s architecture?

Additional stuff:

Facebook has one of the biggest hardware infrastructure in the world with more than 60.000 servers. But servers are worthless if you can’t efficiently use them. So how does Facebook works?

Facebook born in 2004 on a LAMP stack. A PHP website with MySQL datastore. Now is powered by a complex set of interconnected systems grown over the original stack. Main components are:

  1. Frontend
  2. Main persistence layer
  3. Data warehouse layer
  4. Facebook Images
  5. Facebook Messages and Chat
  6. Facebook Search

1. Frontend

hiphopThe frontend is written in PHP and compiled using HipHop (open sourced on github) and g++. HipHop converts PHP in C++ you can compile. Facebook frontend is a single 1,5 GB binary file deployed using BitTorrent. It provide logic and template system. Some parts are external and use HipHop Interpreter (to develop and debug) and HipHop Virtual Machine to run HipHop “ByteCode”.

Is not clear which web server they are using. Apache was used in the beginning and probably is still used for main fronted. In 2009 they acquired (and open sourced) Tornado from FriendFeed and are using it for Real-Time updates feature of the frontend. Static contents (such as fronted images) are served through Akamai.

Everything not in the main binary (or in the ByteCode source) is served using using Thrift protocol. Java services are served without Tomcat or Jetty. Overhead required by these application server is worthless for Facebook architecture. Varnish is used to accelerate responses to HTTP requests.

To speed up frontend rendering they use BigPipe, a sort of mod_pagespeed. After a HTTP request servers fetch data and build HTML structure. HTML is sent before data retrieval and is rendered by the browser. After render, Javascript retrieve and display data (already available) in asynchronous way.

memcachedThe Memcached infrastructure is based on more than 800 servers with 28TB of memory. They use a modded version of Memcached which reimplement the connection pool buffer (and rely on some mods on the network layer of Linux) to better manage hundred of thousand of connections.

Insights and Sources:

2. Main Persistence layer

Facebook rely on MySQL as main datastore. I know, I can’t believe it too…

mysql_clusterThey have one of the largest MySQL Cluster deploy in the world and use the standard InnoDB engine. As many other people they designd the system to easìly handle MySQL failure, they simply enforce backup. Recently Facebook Engineers published a post about their 3-level backup stack:

  • Stage 1: each node (all the replica set) of cluster has 2 rack backup. One of these backup binary log (to speed up slave promotion) and one with mysqldump snapshot taken every night (to handle node failure).
  • Stage 2: each binlog and backup are copied in a larger Hadoop cluster mapped using HDFS. This layer provide a fast and distributed source of data useful to restore nodes also in differenti locations.
  • Stage 3: provide a long-term storage copying data from Hadoop cluster to a discrete storage in a separate region.

Facebook collect several statistics about failure, traffic and backups. These statistics contribute to build a “score” of a node. If a node fails or has a score too low an automatic system provide to restore it from Stage 1 or Stage 2. They don’t try to avoid problem, they simply handle them as fast as possible.

Insights and Sources:

3. Data warehouse layer

hadoopMySQL data is also moved to a “cold” storage to be analyzed. This storage is based on Hadoop HDFS (which leverage on HBase) and is queried using Hive. Data such as logging, clicks and feeds transit using Scribe and are aggregating and stored in Hadoop HDFS using Scribe-HDFS.

Standard Hadoop HDFS implementation has some limitations. Facebook adds a different kind of NameNodes called AvatarNodes which use Apache ZooKeeper to handle node failure. They also adds RaidNode that use XOR-parity files to decrease storage usage.


Analysis are made using MapReduce. Anyway analyze several petabytes of data is hard and standard Hadoop MapReduce framework became a limit on 2011. So they developed Corona, a new scheduling framework that separates cluster resource management from job coordination. Results are stored into Oracle RAC (Real Application Cluster).

Insights and Sources:

In the next post I’m going to analyze the other Facebook services: Images, Message and Search.

Yesterday @lastknight was looking for something to store and query a huge graph dataset. He found Titan developed by Aurelius and released last August. It is a distributed graph database who can rely on Apache Cassandra, Apache HBase or Oracle BerkeyDB for storage. It promises to be fully distributed and horizontally scalable, it’s really ambitious and presentation seem really interesting 🙂

Aurelius also develops Faunus, a Apache Hadoop-based graph analytics engine for analyzing massive-scale graphs.

More about Titan and Faunus:

[UPDATE 2013-03-10] Subscription page describe details about software (developed and planned) by Aurelius. The Fulgora processor seems really interesting.