Bigtable: A Distributed Storage System for Structured Data (2006): GFS(Google File System)和Bigtable是Google数据基础架构的两个核心组件。 BigTable是基于GFS构建的高性能分布式数据存储,并且BigTable解释了如何在GFS之上构建低延迟数据存储。 其中一些可能已经被谷歌内部新的专有技术所取代,但这些想法仍然存在。



1. Bigtable: A Distributed Storage System for Structured Data Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach Mike Burrows, Tushar Chandra, Andrew Fikes, Robert E. Gruber {fay,jeff,sanjay,wilsonh,kerr,m3b,tushar,fikes,gruber} Google, Inc. Abstract achieved scalability and high performance, but Bigtable Bigtable is a distributed storage system for managing provides a different interface than such systems. Bigtable structured data that is designed to scale to a very large does not support a full relational data model; instead, it size: petabytes of data across thousands of commodity provides clients with a simple data model that supports servers. Many projects at Google store data in Bigtable, dynamic control over data layout and format, and al- including web indexing, Google Earth, and Google Fi- lows clients to reason about the locality properties of the nance. These applications place very different demands data represented in the underlying storage. Data is in- on Bigtable, both in terms of data size (from URLs to dexed using row and column names that can be arbitrary web pages to satellite imagery) and latency requirements strings. Bigtable also treats data as uninterpreted strings, (from backend bulk processing to real-time data serving). although clients often serialize various forms of struc- Despite these varied demands, Bigtable has successfully tured and semi-structured data into these strings. Clients provided a flexible, high-performance solution for all of can control the locality of their data through careful these Google products. In this paper we describe the sim- choices in their schemas. Finally, Bigtable schema pa- ple data model provided by Bigtable, which gives clients rameters let clients dynamically control whether to serve dynamic control over data layout and format, and we de- data out of memory or from disk. scribe the design and implementation of Bigtable. Section 2 describes the data model in more detail, and Section 3 provides an overview of the client API. Sec- tion 4 briefly describes the underlying Google infrastruc- 1 Introduction ture on which Bigtable depends. Section 5 describes the fundamentals of the Bigtable implementation, and Sec- Over the last two and a half years we have designed, tion 6 describes some of the refinements that we made implemented, and deployed a distributed storage system to improve Bigtable’s performance. Section 7 provides for managing structured data at Google called Bigtable. measurements of Bigtable’s performance. We describe Bigtable is designed to reliably scale to petabytes of several examples of how Bigtable is used at Google data and thousands of machines. Bigtable has achieved in Section 8, and discuss some lessons we learned in several goals: wide applicability, scalability, high per- designing and supporting Bigtable in Section 9. Fi- formance, and high availability. Bigtable is used by nally, Section 10 describes related work, and Section 11 more than sixty Google products and projects, includ- presents our conclusions. ing Google Analytics, Google Finance, Orkut, Person- alized Search, Writely, and Google Earth. These prod- ucts use Bigtable for a variety of demanding workloads, which range from throughput-oriented batch-processing 2 Data Model jobs to latency-sensitive serving of data to end users. The Bigtable clusters used by these products span a wide A Bigtable is a sparse, distributed, persistent multi- range of configurations, from a handful to thousands of dimensional sorted map. The map is indexed by a row servers, and store up to several hundred terabytes of data. key, column key, and a timestamp; each value in the map In many ways, Bigtable resembles a database: it shares is an uninterpreted array of bytes. many implementation strategies with databases. Paral- lel databases [14] and main-memory databases [13] have (row:string, column:string, time:int64) → string To appear in OSDI 2006 1

2. "contents:" "" "" "<html>..." t3 "com.cnn.www" "<html>..." t5 "CNN" t9 "" t8 "<html>..." t6 Figure 1: A slice of an example table that stores Web pages. The row name is a reversed URL. The contents column family con- tains the page contents, and the anchor column family contains the text of any anchors that reference the page. CNN’s home page is referenced by both the Sports Illustrated and the MY-look home pages, so the row contains columns named and Each anchor cell has one version; the contents column has three versions, at timestamps t3 , t5 , and t6 . We settled on this data model after examining a variety Column Families of potential uses of a Bigtable-like system. As one con- crete example that drove some of our design decisions, Column keys are grouped into sets called column fami- suppose we want to keep a copy of a large collection of lies, which form the basic unit of access control. All data web pages and related information that could be used by stored in a column family is usually of the same type (we many different projects; let us call this particular table compress data in the same column family together). A the Webtable. In Webtable, we would use URLs as row column family must be created before data can be stored keys, various aspects of web pages as column names, and under any column key in that family; after a family has store the contents of the web pages in the contents: col- been created, any column key within the family can be umn under the timestamps when they were fetched, as used. It is our intent that the number of distinct column illustrated in Figure 1. families in a table be small (in the hundreds at most), and that families rarely change during operation. In contrast, a table may have an unbounded number of columns. A column key is named using the following syntax: family:qualifier. Column family names must be print- Rows able, but qualifiers may be arbitrary strings. An exam- ple column family for the Webtable is language, which The row keys in a table are arbitrary strings (currently up stores the language in which a web page was written. We to 64KB in size, although 10-100 bytes is a typical size use only one column key in the language family, and it for most of our users). Every read or write of data under stores each web page’s language ID. Another useful col- a single row key is atomic (regardless of the number of umn family for this table is anchor; each column key in different columns being read or written in the row), a this family represents a single anchor, as shown in Fig- design decision that makes it easier for clients to reason ure 1. The qualifier is the name of the referring site; the about the system’s behavior in the presence of concurrent cell contents is the link text. updates to the same row. Access control and both disk and memory account- Bigtable maintains data in lexicographic order by row ing are performed at the column-family level. In our key. The row range for a table is dynamically partitioned. Webtable example, these controls allow us to manage Each row range is called a tablet, which is the unit of dis- several different types of applications: some that add new tribution and load balancing. As a result, reads of short base data, some that read the base data and create derived row ranges are efficient and typically require communi- column families, and some that are only allowed to view cation with only a small number of machines. Clients existing data (and possibly not even to view all of the can exploit this property by selecting their row keys so existing families for privacy reasons). that they get good locality for their data accesses. For example, in Webtable, pages in the same domain are Timestamps grouped together into contiguous rows by reversing the hostname components of the URLs. For example, we Each cell in a Bigtable can contain multiple versions of store data for under the the same data; these versions are indexed by timestamp. key Storing pages from Bigtable timestamps are 64-bit integers. They can be as- the same domain near each other makes some host and signed by Bigtable, in which case they represent “real domain analyses more efficient. time” in microseconds, or be explicitly assigned by client To appear in OSDI 2006 2

3.// Open the table Scanner scanner(T); Table *T = OpenOrDie("/bigtable/web/webtable"); ScanStream *stream; stream = scanner.FetchColumnFamily("anchor"); // Write a new anchor and delete an old anchor stream->SetReturnAllVersions(); RowMutation r1(T, "com.cnn.www"); scanner.Lookup("com.cnn.www"); r1.Set("", "CNN"); for (; !stream->Done(); stream->Next()) { r1.Delete(""); printf("%s %s %lld %s\n", Operation op; scanner.RowName(), Apply(&op, &r1); stream->ColumnName(), stream->MicroTimestamp(), stream->Value()); Figure 2: Writing to Bigtable. } Figure 3: Reading from Bigtable. applications. Applications that need to avoid collisions must generate unique timestamps themselves. Different versions of a cell are stored in decreasing timestamp or- Bigtable supports several other features that allow the der, so that the most recent versions can be read first. user to manipulate data in more complex ways. First, To make the management of versioned data less oner- Bigtable supports single-row transactions, which can be ous, we support two per-column-family settings that tell used to perform atomic read-modify-write sequences on Bigtable to garbage-collect cell versions automatically. data stored under a single row key. Bigtable does not cur- The client can specify either that only the last n versions rently support general transactions across row keys, al- of a cell be kept, or that only new-enough versions be though it provides an interface for batching writes across kept (e.g., only keep values that were written in the last row keys at the clients. Second, Bigtable allows cells seven days). to be used as integer counters. Finally, Bigtable sup- In our Webtable example, we set the timestamps of ports the execution of client-supplied scripts in the ad- the crawled pages stored in the contents: column to dress spaces of the servers. The scripts are written in a the times at which these page versions were actually language developed at Google for processing data called crawled. The garbage-collection mechanism described Sawzall [28]. At the moment, our Sawzall-based API above lets us keep only the most recent three versions of does not allow client scripts to write back into Bigtable, every page. but it does allow various forms of data transformation, filtering based on arbitrary expressions, and summariza- tion via a variety of operators. 3 API Bigtable can be used with MapReduce [12], a frame- work for running large-scale parallel computations de- The Bigtable API provides functions for creating and veloped at Google. We have written a set of wrappers deleting tables and column families. It also provides that allow a Bigtable to be used both as an input source functions for changing cluster, table, and column family and as an output target for MapReduce jobs. metadata, such as access control rights. Client applications can write or delete values in 4 Building Blocks Bigtable, look up values from individual rows, or iter- ate over a subset of the data in a table. Figure 2 shows Bigtable is built on several other pieces of Google in- C++ code that uses a RowMutation abstraction to per- frastructure. Bigtable uses the distributed Google File form a series of updates. (Irrelevant details were elided System (GFS) [17] to store log and data files. A Bigtable to keep the example short.) The call to Apply performs cluster typically operates in a shared pool of machines an atomic mutation to the Webtable: it adds one anchor that run a wide variety of other distributed applications, to and deletes a different anchor. and Bigtable processes often share the same machines Figure 3 shows C++ code that uses a Scanner ab- with processes from other applications. Bigtable de- straction to iterate over all anchors in a particular row. pends on a cluster management system for scheduling Clients can iterate over multiple column families, and jobs, managing resources on shared machines, dealing there are several mechanisms for limiting the rows, with machine failures, and monitoring machine status. columns, and timestamps produced by a scan. For ex- The Google SSTable file format is used internally to ample, we could restrict the scan above to only produce store Bigtable data. An SSTable provides a persistent, anchors whose columns match the regular expression ordered immutable map from keys to values, where both anchor:*, or to only produce anchors whose keys and values are arbitrary byte strings. Operations are timestamps fall within ten days of the current time. provided to look up the value associated with a specified To appear in OSDI 2006 3

4.key, and to iterate over all key/value pairs in a specified dynamically added (or removed) from a cluster to acco- key range. Internally, each SSTable contains a sequence modate changes in workloads. of blocks (typically each block is 64KB in size, but this The master is responsible for assigning tablets to tablet is configurable). A block index (stored at the end of the servers, detecting the addition and expiration of tablet SSTable) is used to locate blocks; the index is loaded servers, balancing tablet-server load, and garbage col- into memory when the SSTable is opened. A lookup lection of files in GFS. In addition, it handles schema can be performed with a single disk seek: we first find changes such as table and column family creations. the appropriate block by performing a binary search in Each tablet server manages a set of tablets (typically the in-memory index, and then reading the appropriate we have somewhere between ten to a thousand tablets per block from disk. Optionally, an SSTable can be com- tablet server). The tablet server handles read and write pletely mapped into memory, which allows us to perform requests to the tablets that it has loaded, and also splits lookups and scans without touching disk. tablets that have grown too large. Bigtable relies on a highly-available and persistent As with many single-master distributed storage sys- distributed lock service called Chubby [8]. A Chubby tems [17, 21], client data does not move through the mas- service consists of five active replicas, one of which is ter: clients communicate directly with tablet servers for elected to be the master and actively serve requests. The reads and writes. Because Bigtable clients do not rely on service is live when a majority of the replicas are running the master for tablet location information, most clients and can communicate with each other. Chubby uses the never communicate with the master. As a result, the mas- Paxos algorithm [9, 23] to keep its replicas consistent in ter is lightly loaded in practice. the face of failure. Chubby provides a namespace that A Bigtable cluster stores a number of tables. Each ta- consists of directories and small files. Each directory or ble consists of a set of tablets, and each tablet contains file can be used as a lock, and reads and writes to a file all data associated with a row range. Initially, each table are atomic. The Chubby client library provides consis- consists of just one tablet. As a table grows, it is auto- tent caching of Chubby files. Each Chubby client main- matically split into multiple tablets, each approximately tains a session with a Chubby service. A client’s session 100-200 MB in size by default. expires if it is unable to renew its session lease within the lease expiration time. When a client’s session expires, it 5.1 Tablet Location loses any locks and open handles. Chubby clients can also register callbacks on Chubby files and directories We use a three-level hierarchy analogous to that of a B+ - for notification of changes or session expiration. tree [10] to store tablet location information (Figure 4). Bigtable uses Chubby for a variety of tasks: to ensure UserTable1 that there is at most one active master at any time; to Other ... store the bootstrap location of Bigtable data (see Sec- METADATA tablets ... tion 5.1); to discover tablet servers and finalize tablet ... .. server deaths (see Section 5.2); to store Bigtable schema Root tablet . ... information (the column family information for each ta- Chubby file (1st METADATA tablet) ... UserTableN ble); and to store access control lists. If Chubby becomes ... .. . ... unavailable for an extended period of time, Bigtable be- .. comes unavailable. We recently measured this effect ... . in 14 Bigtable clusters spanning 11 Chubby instances. ... The average percentage of Bigtable server hours during which some data stored in Bigtable was not available due Figure 4: Tablet location hierarchy. to Chubby unavailability (caused by either Chubby out- ages or network issues) was 0.0047%. The percentage The first level is a file stored in Chubby that contains for the single cluster that was most affected by Chubby the location of the root tablet. The root tablet contains unavailability was 0.0326%. the location of all tablets in a special METADATA table. Each METADATA tablet contains the location of a set of user tablets. The root tablet is just the first tablet in the 5 Implementation METADATA table, but is treated specially—it is never split—to ensure that the tablet location hierarchy has no The Bigtable implementation has three major compo- more than three levels. nents: a library that is linked into every client, one mas- The METADATA table stores the location of a tablet ter server, and many tablet servers. Tablet servers can be under a row key that is an encoding of the tablet’s table To appear in OSDI 2006 4

5.identifier and its end row. Each METADATA row stores The master is responsible for detecting when a tablet approximately 1KB of data in memory. With a modest server is no longer serving its tablets, and for reassign- limit of 128 MB METADATA tablets, our three-level lo- ing those tablets as soon as possible. To detect when a cation scheme is sufficient to address 234 tablets (or 261 tablet server is no longer serving its tablets, the master bytes in 128 MB tablets). periodically asks each tablet server for the status of its The client library caches tablet locations. If the client lock. If a tablet server reports that it has lost its lock, does not know the location of a tablet, or if it discov- or if the master was unable to reach a server during its ers that cached location information is incorrect, then last several attempts, the master attempts to acquire an it recursively moves up the tablet location hierarchy. exclusive lock on the server’s file. If the master is able to If the client’s cache is empty, the location algorithm acquire the lock, then Chubby is live and the tablet server requires three network round-trips, including one read is either dead or having trouble reaching Chubby, so the from Chubby. If the client’s cache is stale, the location master ensures that the tablet server can never serve again algorithm could take up to six round-trips, because stale by deleting its server file. Once a server’s file has been cache entries are only discovered upon misses (assuming deleted, the master can move all the tablets that were pre- that METADATA tablets do not move very frequently). viously assigned to that server into the set of unassigned Although tablet locations are stored in memory, so no tablets. To ensure that a Bigtable cluster is not vulnera- GFS accesses are required, we further reduce this cost ble to networking issues between the master and Chubby, in the common case by having the client library prefetch the master kills itself if its Chubby session expires. How- tablet locations: it reads the metadata for more than one ever, as described above, master failures do not change tablet whenever it reads the METADATA table. the assignment of tablets to tablet servers. We also store secondary information in the When a master is started by the cluster management METADATA table, including a log of all events per- system, it needs to discover the current tablet assign- taining to each tablet (such as when a server begins ments before it can change them. The master executes serving it). This information is helpful for debugging the following steps at startup. (1) The master grabs and performance analysis. a unique master lock in Chubby, which prevents con- current master instantiations. (2) The master scans the servers directory in Chubby to find the live servers. 5.2 Tablet Assignment (3) The master communicates with every live tablet server to discover what tablets are already assigned to Each tablet is assigned to one tablet server at a time. The each server. (4) The master scans the METADATA table master keeps track of the set of live tablet servers, and to learn the set of tablets. Whenever this scan encounters the current assignment of tablets to tablet servers, in- a tablet that is not already assigned, the master adds the cluding which tablets are unassigned. When a tablet is tablet to the set of unassigned tablets, which makes the unassigned, and a tablet server with sufficient room for tablet eligible for tablet assignment. the tablet is available, the master assigns the tablet by sending a tablet load request to the tablet server. One complication is that the scan of the METADATA Bigtable uses Chubby to keep track of tablet servers. table cannot happen until the METADATA tablets have When a tablet server starts, it creates, and acquires an been assigned. Therefore, before starting this scan (step exclusive lock on, a uniquely-named file in a specific 4), the master adds the root tablet to the set of unassigned Chubby directory. The master monitors this directory tablets if an assignment for the root tablet was not dis- (the servers directory) to discover tablet servers. A tablet covered during step 3. This addition ensures that the root server stops serving its tablets if it loses its exclusive tablet will be assigned. Because the root tablet contains lock: e.g., due to a network partition that caused the the names of all METADATA tablets, the master knows server to lose its Chubby session. (Chubby provides an about all of them after it has scanned the root tablet. efficient mechanism that allows a tablet server to check The set of existing tablets only changes when a ta- whether it still holds its lock without incurring network ble is created or deleted, two existing tablets are merged traffic.) A tablet server will attempt to reacquire an ex- to form one larger tablet, or an existing tablet is split clusive lock on its file as long as the file still exists. If the into two smaller tablets. The master is able to keep file no longer exists, then the tablet server will never be track of these changes because it initiates all but the last. able to serve again, so it kills itself. Whenever a tablet Tablet splits are treated specially since they are initi- server terminates (e.g., because the cluster management ated by a tablet server. The tablet server commits the system is removing the tablet server’s machine from the split by recording information for the new tablet in the cluster), it attempts to release its lock so that the master METADATA table. When the split has committed, it noti- will reassign its tablets more quickly. fies the master. In case the split notification is lost (either To appear in OSDI 2006 5

6.because the tablet server or the master died), the master 5.4 Compactions detects the new tablet when it asks a tablet server to load the tablet that has now split. The tablet server will notify As write operations execute, the size of the memtable in- the master of the split, because the tablet entry it finds in creases. When the memtable size reaches a threshold, the the METADATA table will specify only a portion of the memtable is frozen, a new memtable is created, and the tablet that the master asked it to load. frozen memtable is converted to an SSTable and written to GFS. This minor compaction process has two goals: it shrinks the memory usage of the tablet server, and it 5.3 Tablet Serving reduces the amount of data that has to be read from the The persistent state of a tablet is stored in GFS, as illus- commit log during recovery if this server dies. Incom- trated in Figure 5. Updates are committed to a commit ing read and write operations can continue while com- log that stores redo records. Of these updates, the re- pactions occur. cently committed ones are stored in memory in a sorted Every minor compaction creates a new SSTable. If this buffer called a memtable; the older updates are stored in a behavior continued unchecked, read operations might sequence of SSTables. To recover a tablet, a tablet server need to merge updates from an arbitrary number of SSTables. Instead, we bound the number of such files by periodically executing a merging compaction in the memtable Read Op background. A merging compaction reads the contents of a few SSTables and the memtable, and writes out a new SSTable. The input SSTables and memtable can be Memory discarded as soon as the compaction has finished. GFS A merging compaction that rewrites all SSTables tablet log into exactly one SSTable is called a major compaction. SSTables produced by non-major compactions can con- Write Op tain special deletion entries that suppress deleted data in SSTable Files older SSTables that are still live. A major compaction, on the other hand, produces an SSTable that contains Figure 5: Tablet Representation no deletion information or deleted data. Bigtable cy- cles through all of its tablets and regularly applies major reads its metadata from the METADATA table. This meta- compactions to them. These major compactions allow data contains the list of SSTables that comprise a tablet Bigtable to reclaim resources used by deleted data, and and a set of a redo points, which are pointers into any also allow it to ensure that deleted data disappears from commit logs that may contain data for the tablet. The the system in a timely fashion, which is important for server reads the indices of the SSTables into memory and services that store sensitive data. reconstructs the memtable by applying all of the updates that have committed since the redo points. When a write operation arrives at a tablet server, the 6 Refinements server checks that it is well-formed, and that the sender is authorized to perform the mutation. Authorization is The implementation described in the previous section performed by reading the list of permitted writers from a required a number of refinements to achieve the high Chubby file (which is almost always a hit in the Chubby performance, availability, and reliability required by our client cache). A valid mutation is written to the commit users. This section describes portions of the implementa- log. Group commit is used to improve the throughput of tion in more detail in order to highlight these refinements. lots of small mutations [13, 16]. After the write has been committed, its contents are inserted into the memtable. Locality groups When a read operation arrives at a tablet server, it is similarly checked for well-formedness and proper autho- Clients can group multiple column families together into rization. A valid read operation is executed on a merged a locality group. A separate SSTable is generated for view of the sequence of SSTables and the memtable. each locality group in each tablet. Segregating column Since the SSTables and the memtable are lexicograph- families that are not typically accessed together into sep- ically sorted data structures, the merged view can be arate locality groups enables more efficient reads. For formed efficiently. example, page metadata in Webtable (such as language Incoming read and write operations can continue and checksums) can be in one locality group, and the while tablets are split and merged. contents of the page can be in a different group: an ap- To appear in OSDI 2006 6

7.plication that wants to read the metadata does not need Caching for read performance to read through all of the page contents. To improve read performance, tablet servers use two lev- In addition, some useful tuning parameters can be els of caching. The Scan Cache is a higher-level cache specified on a per-locality group basis. For example, a lo- that caches the key-value pairs returned by the SSTable cality group can be declared to be in-memory. SSTables interface to the tablet server code. The Block Cache is a for in-memory locality groups are loaded lazily into the lower-level cache that caches SSTables blocks that were memory of the tablet server. Once loaded, column fam- read from GFS. The Scan Cache is most useful for appli- ilies that belong to such locality groups can be read cations that tend to read the same data repeatedly. The without accessing the disk. This feature is useful for Block Cache is useful for applications that tend to read small pieces of data that are accessed frequently: we data that is close to the data they recently read (e.g., se- use it internally for the location column family in the quential reads, or random reads of different columns in METADATA table. the same locality group within a hot row). Bloom filters Compression As described in Section 5.3, a read operation has to read from all SSTables that make up the state of a tablet. Clients can control whether or not the SSTables for a If these SSTables are not in memory, we may end up locality group are compressed, and if so, which com- doing many disk accesses. We reduce the number of pression format is used. The user-specified compres- accesses by allowing clients to specify that Bloom fil- sion format is applied to each SSTable block (whose size ters [7] should be created for SSTables in a particu- is controllable via a locality group specific tuning pa- lar locality group. A Bloom filter allows us to ask rameter). Although we lose some space by compress- whether an SSTable might contain any data for a spec- ing each block separately, we benefit in that small por- ified row/column pair. For certain applications, a small tions of an SSTable can be read without decompress- amount of tablet server memory used for storing Bloom ing the entire file. Many clients use a two-pass custom filters drastically reduces the number of disk seeks re- compression scheme. The first pass uses Bentley and quired for read operations. Our use of Bloom filters McIlroy’s scheme [6], which compresses long common also implies that most lookups for non-existent rows or strings across a large window. The second pass uses a columns do not need to touch disk. fast compression algorithm that looks for repetitions in a small 16 KB window of the data. Both compression passes are very fast—they encode at 100–200 MB/s, and Commit-log implementation decode at 400–1000 MB/s on modern machines. If we kept the commit log for each tablet in a separate Even though we emphasized speed instead of space re- log file, a very large number of files would be written duction when choosing our compression algorithms, this concurrently in GFS. Depending on the underlying file two-pass compression scheme does surprisingly well. system implementation on each GFS server, these writes For example, in Webtable, we use this compression could cause a large number of disk seeks to write to the scheme to store Web page contents. In one experiment, different physical log files. In addition, having separate we stored a large number of documents in a compressed log files per tablet also reduces the effectiveness of the locality group. For the purposes of the experiment, we group commit optimization, since groups would tend to limited ourselves to one version of each document in- be smaller. To fix these issues, we append mutations stead of storing all versions available to us. The scheme to a single commit log per tablet server, co-mingling achieved a 10-to-1 reduction in space. This is much mutations for different tablets in the same physical log better than typical Gzip reductions of 3-to-1 or 4-to-1 file [18, 20]. on HTML pages because of the way Webtable rows are Using one log provides significant performance ben- laid out: all pages from a single host are stored close efits during normal operation, but it complicates recov- to each other. This allows the Bentley-McIlroy algo- ery. When a tablet server dies, the tablets that it served rithm to identify large amounts of shared boilerplate in will be moved to a large number of other tablet servers: pages from the same host. Many applications, not just each server typically loads a small number of the orig- Webtable, choose their row names so that similar data inal server’s tablets. To recover the state for a tablet, ends up clustered, and therefore achieve very good com- the new tablet server needs to reapply the mutations for pression ratios. Compression ratios get even better when that tablet from the commit log written by the original we store multiple versions of the same value in Bigtable. tablet server. However, the mutations for these tablets To appear in OSDI 2006 7

8.were co-mingled in the same physical log file. One ap- of the SSTables that we generate are immutable. For ex- proach would be for each new tablet server to read this ample, we do not need any synchronization of accesses full commit log file and apply just the entries needed for to the file system when reading from SSTables. As a re- the tablets it needs to recover. However, under such a sult, concurrency control over rows can be implemented scheme, if 100 machines were each assigned a single very efficiently. The only mutable data structure that is tablet from a failed tablet server, then the log file would accessed by both reads and writes is the memtable. To re- be read 100 times (once by each server). duce contention during reads of the memtable, we make We avoid duplicating log reads by first sort- each memtable row copy-on-write and allow reads and ing the commit log entries in order of the keys writes to proceed in parallel. table, row name, log sequence number . In the Since SSTables are immutable, the problem of perma- sorted output, all mutations for a particular tablet are nently removing deleted data is transformed to garbage contiguous and can therefore be read efficiently with one collecting obsolete SSTables. Each tablet’s SSTables are disk seek followed by a sequential read. To parallelize registered in the METADATA table. The master removes the sorting, we partition the log file into 64 MB seg- obsolete SSTables as a mark-and-sweep garbage collec- ments, and sort each segment in parallel on different tion [25] over the set of SSTables, where the METADATA tablet servers. This sorting process is coordinated by the table contains the set of roots. master and is initiated when a tablet server indicates that Finally, the immutability of SSTables enables us to it needs to recover mutations from some commit log file. split tablets quickly. Instead of generating a new set of Writing commit logs to GFS sometimes causes perfor- SSTables for each child tablet, we let the child tablets mance hiccups for a variety of reasons (e.g., a GFS server share the SSTables of the parent tablet. machine involved in the write crashes, or the network paths traversed to reach the particular set of three GFS servers is suffering network congestion, or is heavily 7 Performance Evaluation loaded). To protect mutations from GFS latency spikes, each tablet server actually has two log writing threads, We set up a Bigtable cluster with N tablet servers to each writing to its own log file; only one of these two measure the performance and scalability of Bigtable as threads is actively in use at a time. If writes to the ac- N is varied. The tablet servers were configured to use 1 tive log file are performing poorly, the log file writing is GB of memory and to write to a GFS cell consisting of switched to the other thread, and mutations that are in 1786 machines with two 400 GB IDE hard drives each. the commit log queue are written by the newly active log N client machines generated the Bigtable load used for writing thread. Log entries contain sequence numbers these tests. (We used the same number of clients as tablet to allow the recovery process to elide duplicated entries servers to ensure that clients were never a bottleneck.) resulting from this log switching process. Each machine had two dual-core Opteron 2 GHz chips, enough physical memory to hold the working set of all running processes, and a single gigabit Ethernet link. Speeding up tablet recovery The machines were arranged in a two-level tree-shaped If the master moves a tablet from one tablet server to switched network with approximately 100-200 Gbps of another, the source tablet server first does a minor com- aggregate bandwidth available at the root. All of the ma- paction on that tablet. This compaction reduces recov- chines were in the same hosting facility and therefore the ery time by reducing the amount of uncompacted state in round-trip time between any pair of machines was less the tablet server’s commit log. After finishing this com- than a millisecond. paction, the tablet server stops serving the tablet. Before The tablet servers and master, test clients, and GFS it actually unloads the tablet, the tablet server does an- servers all ran on the same set of machines. Every ma- other (usually very fast) minor compaction to eliminate chine ran a GFS server. Some of the machines also ran any remaining uncompacted state in the tablet server’s either a tablet server, or a client process, or processes log that arrived while the first minor compaction was from other jobs that were using the pool at the same time being performed. After this second minor compaction as these experiments. is complete, the tablet can be loaded on another tablet R is the distinct number of Bigtable row keys involved server without requiring any recovery of log entries. in the test. R was chosen so that each benchmark read or wrote approximately 1 GB of data per tablet server. Exploiting immutability The sequential write benchmark used row keys with names 0 to R − 1. This space of row keys was parti- Besides the SSTable caches, various other parts of the tioned into 10N equal-sized ranges. These ranges were Bigtable system have been simplified by the fact that all assigned to the N clients by a central scheduler that as- To appear in OSDI 2006 8

9. Values read/written per second 4M scans # of Tablet Servers random reads (mem) 3M random writes Experiment 1 50 250 500 sequential reads random reads 1212 593 479 241 sequential writes 2M random reads random reads (mem) 10811 8511 8000 6250 random writes 8850 3745 3425 2000 1M sequential reads 4425 2463 2625 2469 sequential writes 8547 3623 2451 1905 scans 15385 10526 9524 7843 100 200 300 400 500 Number of tablet servers Figure 6: Number of 1000-byte values read/written per second. The table shows the rate per tablet server; the graph shows the aggregate rate. signed the next available range to a client as soon as the Single tablet-server performance client finished processing the previous range assigned to it. This dynamic assignment helped mitigate the effects Let us first consider performance with just one tablet of performance variations caused by other processes run- server. Random reads are slower than all other operations ning on the client machines. We wrote a single string un- by an order of magnitude or more. Each random read in- der each row key. Each string was generated randomly volves the transfer of a 64 KB SSTable block over the and was therefore uncompressible. In addition, strings network from GFS to a tablet server, out of which only a under different row key were distinct, so no cross-row single 1000-byte value is used. The tablet server executes compression was possible. The random write benchmark approximately 1200 reads per second, which translates was similar except that the row key was hashed modulo into approximately 75 MB/s of data read from GFS. This R immediately before writing so that the write load was bandwidth is enough to saturate the tablet server CPUs spread roughly uniformly across the entire row space for because of overheads in our networking stack, SSTable the entire duration of the benchmark. parsing, and Bigtable code, and is also almost enough to saturate the network links used in our system. Most The sequential read benchmark generated row keys in Bigtable applications with this type of an access pattern exactly the same way as the sequential write benchmark, reduce the block size to a smaller value, typically 8KB. but instead of writing under the row key, it read the string Random reads from memory are much faster since stored under the row key (which was written by an earlier each 1000-byte read is satisfied from the tablet server’s invocation of the sequential write benchmark). Similarly, local memory without fetching a large 64 KB block from the random read benchmark shadowed the operation of GFS. the random write benchmark. Random and sequential writes perform better than ran- The scan benchmark is similar to the sequential read dom reads since each tablet server appends all incoming benchmark, but uses support provided by the Bigtable writes to a single commit log and uses group commit to API for scanning over all values in a row range. Us- stream these writes efficiently to GFS. There is no sig- ing a scan reduces the number of RPCs executed by the nificant difference between the performance of random benchmark since a single RPC fetches a large sequence writes and sequential writes; in both cases, all writes to of values from a tablet server. the tablet server are recorded in the same commit log. Sequential reads perform better than random reads The random reads (mem) benchmark is similar to the since every 64 KB SSTable block that is fetched from random read benchmark, but the locality group that con- GFS is stored into our block cache, where it is used to tains the benchmark data is marked as in-memory, and serve the next 64 read requests. therefore the reads are satisfied from the tablet server’s Scans are even faster since the tablet server can return memory instead of requiring a GFS read. For just this a large number of values in response to a single client benchmark, we reduced the amount of data per tablet RPC, and therefore RPC overhead is amortized over a server from 1 GB to 100 MB so that it would fit com- large number of values. fortably in the memory available to the tablet server. Figure 6 shows two views on the performance of our Scaling benchmarks when reading and writing 1000-byte values to Bigtable. The table shows the number of operations Aggregate throughput increases dramatically, by over a per second per tablet server; the graph shows the aggre- factor of a hundred, as we increase the number of tablet gate number of operations per second. servers in the system from 1 to 500. For example, the To appear in OSDI 2006 9

10. # of tablet servers # of clusters percentage of data served from memory, and complexity 0 .. 19 259 of the table schema. In the rest of this section, we briefly 20 .. 49 47 describe how three product teams use Bigtable. 50 .. 99 20 100 .. 499 50 > 500 12 8.1 Google Analytics Google Analytics ( is a service Table 1: Distribution of number of tablet servers in Bigtable that helps webmasters analyze traffic patterns at their clusters. web sites. It provides aggregate statistics, such as the number of unique visitors per day and the page views per URL per day, as well as site-tracking reports, such as performance of random reads from memory increases by the percentage of users that made a purchase, given that almost a factor of 300 as the number of tablet server in- they earlier viewed a specific page. creases by a factor of 500. This behavior occurs because To enable the service, webmasters embed a small the bottleneck on performance for this benchmark is the JavaScript program in their web pages. This program individual tablet server CPU. is invoked whenever a page is visited. It records various However, performance does not increase linearly. For information about the request in Google Analytics, such most benchmarks, there is a significant drop in per-server as a user identifier and information about the page be- throughput when going from 1 to 50 tablet servers. This ing fetched. Google Analytics summarizes this data and drop is caused by imbalance in load in multiple server makes it available to webmasters. configurations, often due to other processes contending We briefly describe two of the tables used by Google for CPU and network. Our load balancing algorithm at- Analytics. The raw click table (˜200 TB) maintains a tempts to deal with this imbalance, but cannot do a per- row for each end-user session. The row name is a tuple fect job for two main reasons: rebalancing is throttled to containing the website’s name and the time at which the reduce the number of tablet movements (a tablet is un- session was created. This schema ensures that sessions available for a short time, typically less than one second, that visit the same web site are contiguous, and that they when it is moved), and the load generated by our bench- are sorted chronologically. This table compresses to 14% marks shifts around as the benchmark progresses. of its original size. The random read benchmark shows the worst scaling The summary table (˜20 TB) contains various prede- (an increase in aggregate throughput by only a factor of fined summaries for each website. This table is gener- 100 for a 500-fold increase in number of servers). This ated from the raw click table by periodically scheduled behavior occurs because (as explained above) we transfer MapReduce jobs. Each MapReduce job extracts recent one large 64KB block over the network for every 1000- session data from the raw click table. The overall sys- byte read. This transfer saturates various shared 1 Gi- tem’s throughput is limited by the throughput of GFS. gabit links in our network and as a result, the per-server This table compresses to 29% of its original size. throughput drops significantly as we increase the number of machines. 8.2 Google Earth 8 Real Applications Google operates a collection of services that provide users with access to high-resolution satellite imagery of As of August 2006, there are 388 non-test Bigtable clus- the world’s surface, both through the web-based Google ters running in various Google machine clusters, with a Maps interface ( and through the combined total of about 24,500 tablet servers. Table 1 Google Earth ( custom client soft- shows a rough distribution of tablet servers per cluster. ware. These products allow users to navigate across the Many of these clusters are used for development pur- world’s surface: they can pan, view, and annotate satel- poses and therefore are idle for significant periods. One lite imagery at many different levels of resolution. This group of 14 busy clusters with 8069 total tablet servers system uses one table to preprocess data, and a different saw an aggregate volume of more than 1.2 million re- set of tables for serving client data. quests per second, with incoming RPC traffic of about The preprocessing pipeline uses one table to store raw 741 MB/s and outgoing RPC traffic of about 16 GB/s. imagery. During preprocessing, the imagery is cleaned Table 2 provides some data about a few of the tables and consolidated into final serving data. This table con- currently in use. Some tables store data that is served tains approximately 70 terabytes of data and therefore is to users, whereas others store data for batch processing; served from disk. The images are efficiently compressed the tables range widely in total size, average cell size, already, so Bigtable compression is disabled. To appear in OSDI 2006 10

11. Project Table size Compression # Cells # Column # Locality % in Latency- name (TB) ratio (billions) Families Groups memory sensitive? Crawl 800 11% 1000 16 8 0% No Crawl 50 33% 200 2 2 0% No Google Analytics 20 29% 10 1 1 0% Yes Google Analytics 200 14% 80 1 1 0% Yes Google Base 2 31% 10 29 3 15% Yes Google Earth 0.5 64% 8 7 2 33% Yes Google Earth 70 – 9 8 3 0% No Orkut 9 – 0.9 8 5 1% Yes Personalized Search 4 47% 6 93 11 5% Yes Table 2: Characteristics of a few tables in production use. Table size (measured before compression) and # Cells indicate approxi- mate sizes. Compression ratio is not given for tables that have compression disabled. Each row in the imagery table corresponds to a sin- The Personalized Search data is replicated across sev- gle geographic segment. Rows are named to ensure that eral Bigtable clusters to increase availability and to re- adjacent geographic segments are stored near each other. duce latency due to distance from clients. The Personal- The table contains a column family to keep track of the ized Search team originally built a client-side replication sources of data for each segment. This column family mechanism on top of Bigtable that ensured eventual con- has a large number of columns: essentially one for each sistency of all replicas. The current system now uses a raw data image. Since each segment is only built from a replication subsystem that is built into the servers. few images, this column family is very sparse. The design of the Personalized Search storage system The preprocessing pipeline relies heavily on MapRe- allows other groups to add new per-user information in duce over Bigtable to transform data. The overall system their own columns, and the system is now used by many processes over 1 MB/sec of data per tablet server during other Google properties that need to store per-user con- some of these MapReduce jobs. figuration options and settings. Sharing a table amongst The serving system uses one table to index data stored many groups resulted in an unusually large number of in GFS. This table is relatively small (˜500 GB), but it column families. To help support sharing, we added a must serve tens of thousands of queries per second per simple quota mechanism to Bigtable to limit the stor- datacenter with low latency. As a result, this table is age consumption by any particular client in shared ta- hosted across hundreds of tablet servers and contains in- bles; this mechanism provides some isolation between memory column families. the various product groups using this system for per-user information storage. 8.3 Personalized Search 9 Lessons Personalized Search ( is an opt-in service that records user queries and clicks across In the process of designing, implementing, maintaining, a variety of Google properties such as web search, im- and supporting Bigtable, we gained useful experience ages, and news. Users can browse their search histories and learned several interesting lessons. to revisit their old queries and clicks, and they can ask One lesson we learned is that large distributed sys- for personalized search results based on their historical tems are vulnerable to many types of failures, not just Google usage patterns. the standard network partitions and fail-stop failures as- Personalized Search stores each user’s data in sumed in many distributed protocols. For example, we Bigtable. Each user has a unique userid and is assigned have seen problems due to all of the following causes: a row named by that userid. All user actions are stored memory and network corruption, large clock skew, hung in a table. A separate column family is reserved for each machines, extended and asymmetric network partitions, type of action (for example, there is a column family that bugs in other systems that we are using (Chubby for ex- stores all web queries). Each data element uses as its ample), overflow of GFS quotas, and planned and un- Bigtable timestamp the time at which the corresponding planned hardware maintenance. As we have gained more user action occurred. Personalized Search generates user experience with these problems, we have addressed them profiles using a MapReduce over Bigtable. These user by changing various protocols. For example, we added profiles are used to personalize live search results. checksumming to our RPC mechanism. We also handled To appear in OSDI 2006 11

12.some problems by removing assumptions made by one the behavior of Chubby features that were seldom exer- part of the system about another part. For example, we cised by other applications. We discovered that we were stopped assuming a given Chubby operation could return spending an inordinate amount of time debugging ob- only one of a fixed set of errors. scure corner cases, not only in Bigtable code, but also in Another lesson we learned is that it is important to Chubby code. Eventually, we scrapped this protocol and delay adding new features until it is clear how the new moved to a newer simpler protocol that depends solely features will be used. For example, we initially planned on widely-used Chubby features. to support general-purpose transactions in our API. Be- cause we did not have an immediate use for them, how- 10 Related Work ever, we did not implement them. Now that we have many real applications running on Bigtable, we have The Boxwood project [24] has components that overlap been able to examine their actual needs, and have discov- in some ways with Chubby, GFS, and Bigtable, since it ered that most applications require only single-row trans- provides for distributed agreement, locking, distributed actions. Where people have requested distributed trans- chunk storage, and distributed B-tree storage. In each actions, the most important use is for maintaining sec- case where there is overlap, it appears that the Box- ondary indices, and we plan to add a specialized mech- wood’s component is targeted at a somewhat lower level anism to satisfy this need. The new mechanism will than the corresponding Google service. The Boxwood be less general than distributed transactions, but will be project’s goal is to provide infrastructure for building more efficient (especially for updates that span hundreds higher-level services such as file systems or databases, of rows or more) and will also interact better with our while the goal of Bigtable is to directly support client scheme for optimistic cross-data-center replication. applications that wish to store data. A practical lesson that we learned from supporting Many recent projects have tackled the problem of pro- Bigtable is the importance of proper system-level mon- viding distributed storage or higher-level services over itoring (i.e., monitoring both Bigtable itself, as well as wide area networks, often at “Internet scale.” This in- the client processes using Bigtable). For example, we ex- cludes work on distributed hash tables that began with tended our RPC system so that for a sample of the RPCs, projects such as CAN [29], Chord [32], Tapestry [37], it keeps a detailed trace of the important actions done on and Pastry [30]. These systems address concerns that do behalf of that RPC. This feature has allowed us to de- not arise for Bigtable, such as highly variable bandwidth, tect and fix many problems such as lock contention on untrusted participants, or frequent reconfiguration; de- tablet data structures, slow writes to GFS while com- centralized control and Byzantine fault tolerance are not mitting Bigtable mutations, and stuck accesses to the Bigtable goals. METADATA table when METADATA tablets are unavail- In terms of the distributed data storage model that one able. Another example of useful monitoring is that ev- might provide to application developers, we believe the ery Bigtable cluster is registered in Chubby. This allows key-value pair model provided by distributed B-trees or us to track down all clusters, discover how big they are, distributed hash tables is too limiting. Key-value pairs see which versions of our software they are running, how are a useful building block, but they should not be the much traffic they are receiving, and whether or not there only building block one provides to developers. The are any problems such as unexpectedly large latencies. model we chose is richer than simple key-value pairs, The most important lesson we learned is the value and supports sparse semi-structured data. Nonetheless, of simple designs. Given both the size of our system it is still simple enough that it lends itself to a very effi- (about 100,000 lines of non-test code), as well as the cient flat-file representation, and it is transparent enough fact that code evolves over time in unexpected ways, we (via locality groups) to allow our users to tune important have found that code and design clarity are of immense behaviors of the system. help in code maintenance and debugging. One exam- Several database vendors have developed parallel ple of this is our tablet-server membership protocol. Our databases that can store large volumes of data. Oracle’s first protocol was simple: the master periodically issued Real Application Cluster database [27] uses shared disks leases to tablet servers, and tablet servers killed them- to store data (Bigtable uses GFS) and a distributed lock selves if their lease expired. Unfortunately, this proto- manager (Bigtable uses Chubby). IBM’s DB2 Parallel col reduced availability significantly in the presence of Edition [4] is based on a shared-nothing [33] architecture network problems, and was also sensitive to master re- similar to Bigtable. Each DB2 server is responsible for covery time. We redesigned the protocol several times a subset of the rows in a table which it stores in a local until we had a protocol that performed well. However, relational database. Both products provide a complete the resulting protocol was too complex and depended on relational model with transactions. To appear in OSDI 2006 12

13. Bigtable locality groups realize similar compression Given the unusual interface to Bigtable, an interest- and disk read performance benefits observed for other ing question is how difficult it has been for our users to systems that organize data on disk using column-based adapt to using it. New users are sometimes uncertain of rather than row-based storage, including C-Store [1, 34] how to best use the Bigtable interface, particularly if they and commercial products such as Sybase IQ [15, 36], are accustomed to using relational databases that support SenSage [31], KDB+ [22], and the ColumnBM storage general-purpose transactions. Nevertheless, the fact that layer in MonetDB/X100 [38]. Another system that does many Google products successfully use Bigtable demon- vertical and horizontal data partioning into flat files and strates that our design works well in practice. achieves good data compression ratios is AT&T’s Day- We are in the process of implementing several addi- tona database [19]. Locality groups do not support CPU- tional Bigtable features, such as support for secondary cache-level optimizations, such as those described by indices and infrastructure for building cross-data-center Ailamaki [2]. replicated Bigtables with multiple master replicas. We The manner in which Bigtable uses memtables and have also begun deploying Bigtable as a service to prod- SSTables to store updates to tablets is analogous to the uct groups, so that individual groups do not need to main- way that the Log-Structured Merge Tree [26] stores up- tain their own clusters. As our service clusters scale, dates to index data. In both systems, sorted data is we will need to deal with more resource-sharing issues buffered in memory before being written to disk, and within Bigtable itself [3, 5]. reads must merge data from memory and disk. Finally, we have found that there are significant ad- C-Store and Bigtable share many characteristics: both vantages to building our own storage solution at Google. systems use a shared-nothing architecture and have two We have gotten a substantial amount of flexibility from different data structures, one for recent writes, and one designing our own data model for Bigtable. In addi- for storing long-lived data, with a mechanism for mov- tion, our control over Bigtable’s implementation, and ing data from one form to the other. The systems dif- the other Google infrastructure upon which Bigtable de- fer significantly in their API: C-Store behaves like a pends, means that we can remove bottlenecks and ineffi- relational database, whereas Bigtable provides a lower ciencies as they arise. level read and write interface and is designed to support many thousands of such operations per second per server. Acknowledgements C-Store is also a “read-optimized relational DBMS”, whereas Bigtable provides good performance on both We thank the anonymous reviewers, David Nagle, and read-intensive and write-intensive applications. our shepherd Brad Calder, for their feedback on this pa- Bigtable’s load balancer has to solve some of the same per. The Bigtable system has benefited greatly from the kinds of load and memory balancing problems faced by feedback of our many users within Google. In addition, shared-nothing databases (e.g., [11, 35]). Our problem is we thank the following people for their contributions to somewhat simpler: (1) we do not consider the possibility Bigtable: Dan Aguayo, Sameer Ajmani, Zhifeng Chen, of multiple copies of the same data, possibly in alternate Bill Coughran, Mike Epstein, Healfdene Goguen, Robert forms due to views or indices; (2) we let the user tell us Griesemer, Jeremy Hylton, Josh Hyman, Alex Khesin, what data belongs in memory and what data should stay Joanna Kulik, Alberto Lerner, Sherry Listgarten, Mike on disk, rather than trying to determine this dynamically; Maloney, Eduardo Pinheiro, Kathy Polizzi, Frank Yellin, (3) we have no complex queries to execute or optimize. and Arthur Zwiegincew. References 11 Conclusions [1] A BADI , D. J., M ADDEN , S. R., AND F ERREIRA , M. C. Integrating compression and execution in column- We have described Bigtable, a distributed system for oriented database systems. Proc. of SIGMOD (2006). storing structured data at Google. Bigtable clusters have [2] A ILAMAKI , A., D E W ITT, D. J., H ILL , M. D., AND S K - been in production use since April 2005, and we spent OUNAKIS , M. Weaving relations for cache performance. roughly seven person-years on design and implementa- In The VLDB Journal (2001), pp. 169–180. tion before that date. As of August 2006, more than sixty [3] BANGA , G., D RUSCHEL , P., AND M OGUL , J. C. Re- projects are using Bigtable. Our users like the perfor- source containers: A new facility for resource manage- mance and high availability provided by the Bigtable im- ment in server systems. In Proc. of the 3rd OSDI (Feb. plementation, and that they can scale the capacity of their 1999), pp. 45–58. clusters by simply adding more machines to the system [4] BARU , C. K., F ECTEAU , G., G OYAL , A., H SIAO , as their resource demands change over time. H., J HINGRAN , A., PADMANABHAN , S., C OPELAND , To appear in OSDI 2006 13

14. G. P., AND W ILSON , W. G. DB2 parallel edition. IBM [22] KX . COM . Product page. Systems Journal 34, 2 (1995), 292–322. [23] L AMPORT, L. The part-time parliament. ACM TOCS 16, [5] BAVIER , A., B OWMAN , M., C HUN , B., C ULLER , D., 2 (1998), 133–169. K ARLIN , S., P ETERSON , L., ROSCOE , T., S PALINK , T., [24] M AC C ORMICK , J., M URPHY, N., NAJORK , M., AND WAWRZONIAK , M. Operating system support for T HEKKATH , C. A., AND Z HOU , L. Boxwood: Abstrac- planetary-scale network services. In Proc. of the 1st NSDI tions as the foundation for storage infrastructure. In Proc. (Mar. 2004), pp. 253–266. of the 6th OSDI (Dec. 2004), pp. 105–120. [6] B ENTLEY, J. L., AND M C I LROY, M. D. Data compres- [25] M C C ARTHY, J. Recursive functions of symbolic expres- sion using long common strings. In Data Compression sions and their computation by machine. CACM 3, 4 (Apr. Conference (1999), pp. 287–295. 1960), 184–195. [7] B LOOM , B. H. Space/time trade-offs in hash coding with [26] O’N EIL , P., C HENG , E., G AWLICK , D., AND O’N EIL , allowable errors. CACM 13, 7 (1970), 422–426. E. The log-structured merge-tree (LSM-tree). Acta Inf. [8] B URROWS , M. The Chubby lock service for loosely- 33, 4 (1996), 351–385. coupled distributed systems. In Proc. of the 7th OSDI [27] ORACLE . COM . (Nov. 2006). database/clustering/index.html. Product page. [9] C HANDRA , T., G RIESEMER , R., AND R EDSTONE , J. [28] P IKE , R., D ORWARD , S., G RIESEMER , R., AND Q UIN - Paxos made live — An engineering perspective. In Proc. LAN , S. Interpreting the data: Parallel analysis with of PODC (2007). Sawzall. Scientific Programming Journal 13, 4 (2005), 227–298. [10] C OMER , D. Ubiquitous B-tree. Computing Surveys 11, 2 (June 1979), 121–137. [29] R ATNASAMY, S., F RANCIS , P., H ANDLEY, M., K ARP, R., AND S HENKER , S. A scalable content-addressable [11] C OPELAND , G. P., A LEXANDER , W., B OUGHTER , network. In Proc. of SIGCOMM (Aug. 2001), pp. 161– E. E., AND K ELLER , T. W. Data placement in Bubba. In 172. Proc. of SIGMOD (1988), pp. 99–108. [30] ROWSTRON , A., AND D RUSCHEL , P. Pastry: Scal- [12] D EAN , J., AND G HEMAWAT, S. MapReduce: Simplified able, distributed object location and routing for large- data processing on large clusters. In Proc. of the 6th OSDI scale peer-to-peer systems. In Proc. of Middleware 2001 (Dec. 2004), pp. 137–150. (Nov. 2001), pp. 329–350. [13] D E W ITT, D., K ATZ , R., O LKEN , F., S HAPIRO , L., [31] SENSAGE . COM . S TONEBRAKER , M., AND W OOD , D. Implementation Product page. techniques for main memory database systems. In Proc. [32] S TOICA , I., M ORRIS , R., K ARGER , D., K AASHOEK , of SIGMOD (June 1984), pp. 1–8. M. F., AND BALAKRISHNAN , H. Chord: A scalable [14] D E W ITT, D. J., AND G RAY, J. Parallel database sys- peer-to-peer lookup service for Internet applications. In tems: The future of high performance database systems. Proc. of SIGCOMM (Aug. 2001), pp. 149–160. CACM 35, 6 (June 1992), 85–98. [33] S TONEBRAKER , M. The case for shared nothing. [15] F RENCH , C. D. One size fits all database architectures Database Engineering Bulletin 9, 1 (Mar. 1986), 4–9. do not work for DSS. In Proc. of SIGMOD (May 1995), [34] S TONEBRAKER , M., A BADI , D. J., BATKIN , A., C HEN , pp. 449–450. X., C HERNIACK , M., F ERREIRA , M., L AU , E., L IN , [16] G AWLICK , D., AND K INKADE , D. Varieties of concur- A., M ADDEN , S., O’N EIL , E., O’N EIL , P., R ASIN , rency control in IMS/VS fast path. Database Engineering A., T RAN , N., AND Z DONIK , S. C-Store: A column- Bulletin 8, 2 (1985), 3–10. oriented DBMS. In Proc. of VLDB (Aug. 2005), pp. 553– [17] G HEMAWAT, S., G OBIOFF , H., AND L EUNG , S.-T. The 564. Google file system. In Proc. of the 19th ACM SOSP (Dec. [35] S TONEBRAKER , M., AOKI , P. M., D EVINE , R., 2003), pp. 29–43. L ITWIN , W., AND O LSON , M. A. Mariposa: A new ar- [18] G RAY, J. Notes on database operating systems. In Oper- chitecture for distributed data. In Proc. of the Tenth ICDE ating Systems — An Advanced Course, vol. 60 of Lecture (1994), IEEE Computer Society, pp. 54–65. Notes in Computer Science. Springer-Verlag, 1978. [36] SYBASE . COM . [19] G REER , R. Daytona and the fourth-generation language servers/sybaseiq. Product page. Cymbal. In Proc. of SIGMOD (1999), pp. 525–526. [37] Z HAO , B. Y., K UBIATOWICZ , J., AND J OSEPH , A. D. Tapestry: An infrastructure for fault-tolerant wide-area [20] H AGMANN , R. Reimplementing the Cedar file system location and routing. Tech. Rep. UCB/CSD-01-1141, CS using logging and group commit. In Proc. of the 11th Division, UC Berkeley, Apr. 2001. SOSP (Dec. 1987), pp. 155–162. [38] Z UKOWSKI , M., B ONCZ , P. A., N ES , N., AND H EMAN , [21] H ARTMAN , J. H., AND O USTERHOUT, J. K. The Zebra S. MonetDB/X100 — A DBMS in the CPU cache. IEEE striped network file system. In Proc. of the 14th SOSP Data Eng. Bull. 28, 2 (2005), 17–22. (Asheville, NC, 1993), pp. 29–43. To appear in OSDI 2006 14