Hadoop

978-1-4244-7153-9/10/$26.00 ©2010 IEEE
The Hadoop Distributed File System
Konstantin Shvachko, Hairong Kuang, Sanjay Radia, Robert Chansler
Yahoo!
Sunnyvale, California USA
{Shv, Hairong, SRadia, Chansler}@Yahoo-Inc.com
Abstract—The Hadoop Distributed File System (HDFS) is
designed to store very large data sets reliably, and to stream
those data sets at high bandwidth to user applications. In a large
cluster, thousands of servers both host directly attached storage
and execute user application tasks. By distributing storage and
computation across many servers, the resource can grow with
demand while remaining economical at every size. We describe
the architecture of HDFS and report on experience using HDFS
to manage 25 petabytes of enterprise data at Yahoo!.
Keywords: Hadoop, HDFS, distributed file system
I. INTRODUCTION AND RELATED WORK
Hadoop [1][16][19] provides a distributed file system and a
framework for the analysis and transformation of very large
data sets using the MapReduce [3] paradigm. An important
characteristic of Hadoop is the partitioning of data and compu-
tation across many (thousands) of hosts, and executing applica-
tion computations in parallel close to their data. A Hadoop
cluster scales computation capacity, storage capacity and IO
bandwidth by simply adding commodity servers. Hadoop clus-
ters at Yahoo! span 25 000 servers, and store 25 petabytes of
application data, with the largest cluster being 3500 servers.
One hundred other organizations worldwide report using
Hadoop.
HDFS
Distributed file system
Subject of this paper!
MapReduce Distributed computation framework
HBase Column-oriented table service
Pig
Dataflow language and parallel execution
framework
Hive Data warehouse infrastructure
ZooKeeper Distributed coordination service
Chukwa System for collecting management data
Avro Data serialization system
Table 1. Hadoop project components
Hadoop is an Apache project; all components are available
via the Apache open source license. Yahoo! has developed and
contributed to 80% of the core of Hadoop (HDFS and MapRe-
duce). HBase was originally developed at Powerset, now a
department at Microsoft. Hive [15] was originated and devel-
developed at Facebook. Pig [4], ZooKeeper [6], and Chukwa
were originated and developed at Yahoo! Avro was originated
at Yahoo! and is being co-developed with Cloudera.
HDFS is the file system component of Hadoop. While the
interface to HDFS is patterned after the UNIX file system,
faithfulness to standards was sacrificed in favor of improved
performance for the applications at hand.
HDFS stores file system metadata and application data
separately. As in other distributed file systems, like PVFS
[2][14], Lustre [7] and GFS [5][8], HDFS stores metadata on a
dedicated server, called the NameNode. Application data are
stored on other servers called DataNodes. All servers are fully
connected and communicate with each other using TCP-based
protocols.
Unlike Lustre and PVFS, the DataNodes in HDFS do not
use data protection mechanisms such as RAID to make the data
durable. Instead, like GFS, the file content is replicated on mul-
tiple DataNodes for reliability. While ensuring data durability,
this strategy has the added advantage that data transfer band-
width is multiplied, and there are more opportunities for locat-
ing computation near the needed data.
Several distributed file systems have or are exploring truly
distributed implementations of the namespace. Ceph [17] has a
cluster of namespace servers (MDS) and uses a dynamic sub-
tree partitioning algorithm in order to map the namespace tree
to MDSs evenly. GFS is also evolving into a distributed name-
space implementation [8]. The new GFS will have hundreds of
namespace servers (masters) with 100 million files per master.
Lustre [7] has an implementation of clustered namespace on its
roadmap for Lustre 2.2 release. The intent is to stripe a direc-
tory over multiple metadata servers (MDS), each of which con-
tains a disjoint portion of the namespace. A file is assigned to a
particular MDS using a hash function on the file name.
II. ARCHITECTURE
A. NameNode
The HDFS namespace is a hierarchy of files and directo-
ries. Files and directories are represented on the NameNode by
inodes, which record attributes like permissions, modification
and access times, namespace and disk space quotas. The file
content is split into large blocks (typically 128 megabytes, but
user selectable file-by-file) and each block of the file is inde-
pendently replicated at multiple DataNodes (typically three, but
user selectable file-by-file). The NameNode maintains the
namespace tree and the mapping of file blocks to DataNodes

2
(the physical location of file data). An HDFS client wanting to
read a file first contacts the NameNode for the locations of data
blocks comprising the file and then reads block contents from
the DataNode closest to the client. When writing data, the cli-
ent requests the NameNode to nominate a suite of three
DataNodes to host the block replicas. The client then writes
data to the DataNodes in a pipeline fashion. The current design
has a single NameNode for each cluster. The cluster can have
thousands of DataNodes and tens of thousands of HDFS clients
per cluster, as each DataNode may execute multiple application
tasks concurrently.
HDFS keeps the entire namespace in RAM. The inode data
and the list of blocks belonging to each file comprise the meta-
data of the name system called the image. The persistent record
of the image stored in the local host’s native files system is
called a checkpoint. The NameNode also stores the modifica-
tion log of the image called the journal in the local host’s na-
tive file system. For improved durability, redundant copies of
the checkpoint and journal can be made at other servers. Dur-
ing restarts the NameNode restores the namespace by reading
the namespace and replaying the journal. The locations of
block replicas may change over time and are not part of the
persistent checkpoint.
B. DataNodes
Each block replica on a DataNode is represented by two
files in the local host’s native file system. The first file contains
the data itself and the second file is block’s metadata including
checksums for the block data and the block’s generation stamp.
The size of the data file equals the actual length of the block
and does not require extra space to round it up to the nominal
block size as in traditional file systems. Thus, if a block is half
full it needs only half of the space of the full block on the local
drive.
During startup each DataNode connects to the NameNode
and performs a handshake. The purpose of the handshake is to
verify the namespace ID and the software version of the
DataNode. If either does not match that of the NameNode the
DataNode automatically shuts down.
The namespace ID is assigned to the file system instance
when it is formatted. The namespace ID is persistently stored
on all nodes of the cluster. Nodes with a different namespace
ID will not be able to join the cluster, thus preserving the integ-
rity of the file system.
The consistency of software versions is important because
incompatible version may cause data corruption or loss, and on
large clusters of thousands of machines it is easy to overlook
nodes that did not shut down properly prior to the software
upgrade or were not available during the upgrade.
A DataNode that is newly initialized and without any
namespace ID is permitted to join the cluster and receive the
cluster’s namespace ID.
After the handshake the DataNode registers with the
NameNode. DataNodes persistently store their unique storage
IDs. The storage ID is an internal identifier of the DataNode,
which makes it recognizable even if it is restarted with a differ-
ent IP address or port. The storage ID is assigned to the
DataNode when it registers with the NameNode for the first
time and never changes after that.
A DataNode identifies block replicas in its possession to the
NameNode by sending a block report. A block report contains
the block id, the generation stamp and the length for each block
replica the server hosts. The first block report is sent immedi-
ately after the DataNode registration. Subsequent block reports
are sent every hour and provide the NameNode with an up-to-
date view of where block replicas are located on the cluster.
During normal operation DataNodes send heartbeats to the
NameNode to confirm that the DataNode is operating and the
block replicas it hosts are available. The default heartbeat in-
terval is three seconds. If the NameNode does not receive a
heartbeat from a DataNode in ten minutes the NameNode con-
siders the DataNode to be out of service and the block replicas
hosted by that DataNode to be unavailable. The NameNode
then schedules creation of new replicas of those blocks on other
DataNodes.
Heartbeats from a DataNode also carry information about
total storage capacity, fraction of storage in use, and the num-
ber of data transfers currently in progress. These statistics are
used for the NameNode’s space allocation and load balancing
decisions.
The NameNode does not directly call DataNodes. It uses
replies to heartbeats to send instructions to the DataNodes. The
instructions include commands to:
• replicate blocks to other nodes;
• remove local block replicas;
• re-register or to shut down the node;
• send an immediate block report.
These commands are important for maintaining the overall
system integrity and therefore it is critical to keep heartbeats
frequent even on big clusters. The NameNode can process
thousands of heartbeats per second without affecting other
NameNode operations.
C. HDFS Client
User applications access the file system using the HDFS
client, a code library that exports the HDFS file system inter-
face.
Similar to most conventional file systems, HDFS supports
operations to read, write and delete files, and operations to cre-
ate and delete directories. The user references files and directo-
ries by paths in the namespace. The user application generally
does not need to know that file system metadata and storage are
on different servers, or that blocks have multiple replicas.
When an application reads a file, the HDFS client first asks
the NameNode for the list of DataNodes that host replicas of
the blocks of the file. It then contacts a DataNode directly and
requests the transfer of the desired block. When a client writes,
it first asks the NameNode to choose DataNodes to host repli-
cas of the first block of the file. The client organizes a pipeline
from node-to-node and sends the data. When the first block is
filled, the client requests new DataNodes to be chosen to host
replicas of the next block. A new pipeline is organized, and the

3
Figure 1. An HDFS client creates a new file by giving its path to the NameNode. For each block of the file, the NameNode returns
a list of DataNodes to host its replicas. The client then pipelines data to the chosen DataNodes, which eventually confirm the
creation of the block replicas to the NameNode.
client sends the further bytes of the file. Each choice of
DataNodes is likely to be different. The interactions among the
client, the NameNode and the DataNodes are illustrated in
Fig. 1.
Unlike conventional file systems, HDFS provides an API
that exposes the locations of a file blocks. This allows applica-
tions like the MapReduce framework to schedule a task to
where the data are located, thus improving the read perform-
ance. It also allows an application to set the replication factor
of a file. By default a file’s replication factor is three. For criti-
cal files or files which are accessed very often, having a higher
replication factor improves their tolerance against faults and
increase their read bandwidth.
D. Image and Journal
The namespace image is the file system metadata that de-
scribes the organization of application data as directories and
files. A persistent record of the image written to disk is called a
checkpoint. The journal is a write-ahead commit log for
changes to the file system that must be persistent. For each
client-initiated transaction, the change is recorded in the jour-
nal, and the journal file is flushed and synched before the
change is committed to the HDFS client. The checkpoint file is
never changed by the NameNode; it is replaced in its entirety
when a new checkpoint is created during restart, when re-
quested by the administrator, or by the CheckpointNode de-
scribed in the next section. During startup the NameNode ini-
tializes the namespace image from the checkpoint, and then
replays changes from the journal until the image is up-to-date
with the last state of the file system. A new checkpoint and
empty journal are written back to the storage directories before
the NameNode starts serving clients.
If either the checkpoint or the journal is missing, or be-
comes corrupt, the namespace information will be lost partly or
entirely. In order to preserve this critical information HDFS can
be configured to store the checkpoint and journal in multiple
storage directories. Recommended practice is to place the di-
rectories on different volumes, and for one storage directory to
be on a remote NFS server. The first choice prevents loss from
single volume failures, and the second choice protects against
failure of the entire node. If the NameNode encounters an error
writing the journal to one of the storage directories it automati-
cally excludes that directory from the list of storage directories.
The NameNode automatically shuts itself down if no storage
directory is available.
The NameNode is a multithreaded system and processes
requests simultaneously from multiple clients. Saving a trans-
action to disk becomes a bottleneck since all other threads need
to wait until the synchronous flush-and-sync procedure initi-
ated by one of them is complete. In order to optimize this
process the NameNode batches multiple transactions initiated
by different clients. When one of the NameNode’s threads ini-
tiates a flush-and-sync operation, all transactions batched at
that time are committed together. Remaining threads only need
to check that their transactions have been saved and do not
need to initiate a flush-and-sync operation.
E. CheckpointNode
The NameNode in HDFS, in addition to its primary role
serving client requests, can alternatively execute either of two
other roles, either a CheckpointNode or a BackupNode. The
role is specified at the node startup.
The CheckpointNode periodically combines the existing
checkpoint and journal to create a new checkpoint and an
empty journal. The CheckpointNode usually runs on a different
host from the NameNode since it has the same memory re-
quirements as the NameNode. It downloads the current check-
point and journal files from the NameNode, merges them lo-
cally, and returns the new checkpoint back to the NameNode.

4
Creating periodic checkpoints is one way to protect the file
system metadata. The system can start from the most recent
checkpoint if all other persistent copies of the namespace im-
age or journal are unavailable.
Creating a checkpoint lets the NameNode truncate the tail
of the journal when the new checkpoint is uploaded to the
NameNode. HDFS clusters run for prolonged periods of time
without restarts during which the journal constantly grows. If
the journal grows very large, the probability of loss or corrup-
tion of the journal file increases. Also, a very large journal ex-
tends the time required to restart the NameNode. For a large
cluster, it takes an hour to process a week-long journal. Good
practice is to create a daily checkpoint.
F. BackupNode
A recently introduced feature of HDFS is the BackupNode.
Like a CheckpointNode, the BackupNode is capable of creating
periodic checkpoints, but in addition it maintains an in-
memory, up-to-date image of the file system namespace that is
always synchronized with the state of the NameNode.
The BackupNode accepts the journal stream of namespace
transactions from the active NameNode, saves them to its own
storage directories, and applies these transactions to its own
namespace image in memory. The NameNode treats the
BackupNode as a journal store the same as it treats journal files
in its storage directories. If the NameNode fails, the
BackupNode’s image in memory and the checkpoint on disk is
a record of the latest namespace state.
The BackupNode can create a checkpoint without down-
loading checkpoint and journal files from the active
NameNode, since it already has an up-to-date namespace im-
age in its memory. This makes the checkpoint process on the
BackupNode more efficient as it only needs to save the name-
space into its local storage directories.
The BackupNode can be viewed as a read-only NameNode.
It contains all file system metadata information except for
block locations. It can perform all operations of the regular
NameNode that do not involve modification of the namespace
or knowledge of block locations. Use of a BackupNode pro-
vides the option of running the NameNode without persistent
storage, delegating responsibility for the namespace state per-
sisting to the BackupNode.
G. Upgrades, File System Snapshots
During software upgrades the possibility of corrupting the
system due to software bugs or human mistakes increases. The
purpose of creating snapshots in HDFS is to minimize potential
damage to the data stored in the system during upgrades.
The snapshot mechanism lets administrators persistently
save the current state of the file system, so that if the upgrade
results in data loss or corruption it is possible to rollback the
upgrade and return HDFS to the namespace and storage state as
they were at the time of the snapshot.
The snapshot (only one can exist) is created at the cluster
administrator’s option whenever the system is started. If a
snapshot is requested, the NameNode first reads the checkpoint
and journal files and merges them in memory. Then it writes
the new checkpoint and the empty journal to a new location, so
that the old checkpoint and journal remain unchanged.
During handshake the NameNode instructs DataNodes
whether to create a local snapshot. The local snapshot on the
DataNode cannot be created by replicating the data files direc-
tories as this will require doubling the storage capacity of every
DataNode on the cluster. Instead each DataNode creates a copy
of the storage directory and hard links existing block files into
it. When the DataNode removes a block it removes only the
hard link, and block modifications during appends use the
copy-on-write technique. Thus old block replicas remain un-
touched in their old directories.
The cluster administrator can choose to roll back HDFS to
the snapshot state when restarting the system. The NameNode
recovers the checkpoint saved when the snapshot was created.
DataNodes restore the previously renamed directories and initi-
ate a background process to delete block replicas created after
the snapshot was made. Having chosen to roll back, there is no
provision to roll forward. The cluster administrator can recover
the storage occupied by the snapshot by commanding the sys-
tem to abandon the snapshot, thus finalizing the software up-
grade.
System evolution may lead to a change in the format of the
NameNode’s checkpoint and journal files, or in the data repre-
sentation of block replica files on DataNodes. The layout ver-
sion identifies the data representation formats, and is persis-
tently stored in the NameNode’s and the DataNodes’ storage
directories. During startup each node compares the layout ver-
sion of the current software with the version stored in its stor-
age directories and automatically converts data from older for-
mats to the newer ones. The conversion requires the mandatory
creation of a snapshot when the system restarts with the new
software layout version.
HDFS does not separate layout versions for the NameNode
and DataNodes because snapshot creation must be an all-
cluster effort rather than a node-selective event. If an upgraded
NameNode due to a software bug purges its image then back-
ing up only the namespace state still results in total data loss, as
the NameNode will not recognize the blocks reported by
DataNodes, and will order their deletion. Rolling back in this
case will recover the metadata, but the data itself will be lost. A
coordinated snapshot is required to avoid a cataclysmic de-
struction.
III. FILE I/O OPERATIONS AND REPLICA MANGEMENT
A. File Read and Write
An application adds data to HDFS by creating a new file
and writing the data to it. After the file is closed, the bytes writ-
ten cannot be altered or removed except that new data can be
added to the file by reopening the file for append. HDFS im-
plements a single-writer, multiple-reader model.
The HDFS client that opens a file for writing is granted a
lease for the file; no other client can write to the file. The writ-
ing client periodically renews the lease by sending a heartbeat
to the NameNode. When the file is closed, the lease is revoked.

5
The lease duration is bound by a soft limit and a hard limit.
Until the soft limit expires, the writer is certain of exclusive
access to the file. If the soft limit expires and the client fails to
close the file or renew the lease, another client can preempt the
lease. If after the hard limit expires (one hour) and the client
has failed to renew the lease, HDFS assumes that the client has
quit and will automatically close the file on behalf of the writer,
and recover the lease. The writer's lease does not prevent other
clients from reading the file; a file may have many concurrent
readers.
An HDFS file consists of blocks. When there is a need for a
new block, the NameNode allocates a block with a unique
block ID and determines a list of DataNodes to host replicas of
the block. The DataNodes form a pipeline, the order of which
minimizes the total network distance from the client to the last
DataNode. Bytes are pushed to the pipeline as a sequence of
packets. The bytes that an application writes first buffer at the
client side. After a packet buffer is filled (typically 64 KB), the
data are pushed to the pipeline. The next packet can be pushed
to the pipeline before receiving the acknowledgement for the
previous packets. The number of outstanding packets is limited
by the outstanding packets window size of the client.
After data are written to an HDFS file, HDFS does not pro-
vide any guarantee that data are visible to a new reader until the
file is closed. If a user application needs the visibility guaran-
tee, it can explicitly call the hflush operation. Then the current
packet is immediately pushed to the pipeline, and the hflush
operation will wait until all DataNodes in the pipeline ac-
knowledge the successful transmission of the packet. All data
written before the hflush operation are then certain to be visible
to readers.
Figure 2. Data pipeline during block construction
If no error occurs, block construction goes through three
stages as shown in Fig. 2 illustrating a pipeline of three
DataNodes (DN) and a block of five packets. In the picture,
bold lines represent data packets, dashed lines represent ac-
knowledgment messages, and thin lines represent control mes-
sages to setup and close the pipeline. Vertical lines represent
activity at the client and the three DataNodes where time pro-
ceeds from top to bottom. From t0 to t1 is the pipeline setup
stage. The interval t1 to t2 is the data streaming stage, where t1
is the time when the first data packet gets sent and t2 is the time
that the acknowledgment to the last packet gets received. Here
an hflush operation transmits the second packet. The hflush
indication travels with the packet data and is not a separate
operation. The final interval t2 to t3 is the pipeline close stage
for this block.
In a cluster of thousands of nodes, failures of a node (most
commonly storage faults) are daily occurrences. A replica
stored on a DataNode may become corrupted because of faults
in memory, disk, or network. HDFS generates and stores
checksums for each data block of an HDFS file. Checksums are
verified by the HDFS client while reading to help detect any
corruption caused either by client, DataNodes, or network.
When a client creates an HDFS file, it computes the checksum
sequence for each block and sends it to a DataNode along with
the data. A DataNode stores checksums in a metadata file sepa-
rate from the block’s data file. When HDFS reads a file, each
block’s data and checksums are shipped to the client. The client
computes the checksum for the received data and verifies that
the newly computed checksums matches the checksums it re-
ceived. If not, the client notifies the NameNode of the corrupt
replica and then fetches a different replica of the block from
another DataNode.
When a client opens a file to read, it fetches the list of
blocks and the locations of each block replica from the
NameNode. The locations of each block are ordered by their
distance from the reader. When reading the content of a block,
the client tries the closest replica first. If the read attempt fails,
the client tries the next replica in sequence. A read may fail if
the target DataNode is unavailable, the node no longer hosts a
replica of the block, or the replica is found to be corrupt when
checksums are tested.
HDFS permits a client to read a file that is open for writing.
When reading a file open for writing, the length of the last
block still being written is unknown to the NameNode. In this
case, the client asks one of the replicas for the latest length be-
fore starting to read its content.
The design of HDFS I/O is particularly optimized for batch
processing systems, like MapReduce, which require high
throughput for sequential reads and writes. However, many
efforts have been put to improve its read/write response time in
order to support applications like Scribe that provide real-time
data streaming to HDFS, or HBase that provides random, real-
time access to large tables.
B. Block Placement
For a large cluster, it may not be practical to connect all
nodes in a flat topology. A common practice is to spread the
nodes across multiple racks. Nodes of a rack share a switch,
and rack switches are connected by one or more core switches.
Communication between two nodes in different racks has to go
through multiple switches. In most cases, network bandwidth

6
between nodes in the same rack is greater than network band-
width between nodes in different racks. Fig. 3 describes a clus-
ter with two racks, each of which contains three nodes.
Figure 3. Cluster topology example
HDFS estimates the network bandwidth between two nodes
by their distance. The distance from a node to its parent node is
assumed to be one. A distance between two nodes can be cal-
culated by summing up their distances to their closest common
ancestor. A shorter distance between two nodes means that the
greater bandwidth they can utilize to transfer data.
HDFS allows an administrator to configure a script that re-
turns a node’s rack identification given a node’s address. The
NameNode is the central place that resolves the rack location of
each DataNode. When a DataNode registers with the
NameNode, the NameNode runs a configured script to decide
which rack the node belongs to. If no such a script is config-
ured, the NameNode assumes that all the nodes belong to a
default single rack.
The placement of replicas is critical to HDFS data reliabil-
ity and read/write performance. A good replica placement pol-
icy should improve data reliability, availability, and network
bandwidth utilization. Currently HDFS provides a configurable
block placement policy interface so that the users and research-
ers can experiment and test any policy that’s optimal for their
applications.
The default HDFS block placement policy provides a
tradeoff between minimizing the write cost, and maximizing
data reliability, availability and aggregate read bandwidth.
When a new block is created, HDFS places the first replica on
the node where the writer is located, the second and the third
replicas on two different nodes in a different rack, and the rest
are placed on random nodes with restrictions that no more than
one replica is placed at one node and no more than two replicas
are placed in the same rack when the number of replicas is less
than twice the number of racks. The choice to place the second
and third replicas on a different rack better distributes the block
replicas for a single file across the cluster. If the first two repli-
cas were placed on the same rack, for any file, two-thirds of its
block replicas would be on the same rack.
After all target nodes are selected, nodes are organized as a
pipeline in the order of their proximity to the first replica. Data
are pushed to nodes in this order. For reading, the NameNode
first checks if the client’s host is located in the cluster. If yes,
block locations are returned to the client in the order of its
closeness to the reader. The block is read from DataNodes in
this preference order. (It is usual for MapReduce applications
to run on cluster nodes, but as long as a host can connect to the
NameNode and DataNodes, it can execute the HDFS client.)
This policy reduces the inter-rack and inter-node write traf-
fic and generally improves write performance. Because the
chance of a rack failure is far less than that of a node failure,
this policy does not impact data reliability and availability
guarantees. In the usual case of three replicas, it can reduce the
aggregate network bandwidth used when reading data since a
block is placed in only two unique racks rather than three.
The default HDFS replica placement policy can be summa-
rized as follows:
1. No Datanode contains more than one replica of
any block.
2. No rack contains more than two replicas of the
same block, provided there are sufficient racks on
the cluster.
C. Replication management
The NameNode endeavors to ensure that each block always
has the intended number of replicas. The NameNode detects
that a block has become under- or over-replicated when a block
report from a DataNode arrives. When a block becomes over
replicated, the NameNode chooses a replica to remove. The
NameNode will prefer not to reduce the number of racks that
host replicas, and secondly prefer to remove a replica from the
DataNode with the least amount of available disk space. The
goal is to balance storage utilization across DataNodes without
reducing the block’s availability.
When a block becomes under-replicated, it is put in the rep-
lication priority queue. A block with only one replica has the
highest priority, while a block with a number of replicas that is
greater than two thirds of its replication factor has the lowest
priority. A background thread periodically scans the head of the
replication queue to decide where to place new replicas. Block
replication follows a similar policy as that of the new block
placement. If the number of existing replicas is one, HDFS
places the next replica on a different rack. In case that the block
has two existing replicas, if the two existing replicas are on the
same rack, the third replica is placed on a different rack; other-
wise, the third replica is placed on a different node in the same
rack as an existing replica. Here the goal is to reduce the cost of
creating new replicas.
The NameNode also makes sure that not all replicas of a
block are located on one rack. If the NameNode detects that a
block’s replicas end up at one rack, the NameNode treats the
block as under-replicated and replicates the block to a different
rack using the same block placement policy described above.
After the NameNode receives the notification that the replica is
created, the block becomes over-replicated. The NameNode
then will decides to remove an old replica because the over-
replication policy prefers not to reduce the number of racks.
D. Balancer
HDFS block placement strategy does not take into account
DataNode disk space utilization. This is to avoid placing
new—more likely to be referenced—data at a small subset of

DN00
Rack 1
/

Rack 0
DN01 DN02 DN10 DN11 DN12

7
the DataNodes. Therefore data might not always be placed uni-
formly across DataNodes. Imbalance also occurs when new
nodes are added to the cluster.
The balancer is a tool that balances disk space usage on an
HDFS cluster. It takes a threshold value as an input parameter,
which is a fraction in the range of (0, 1). A cluster is balanced
if for each DataNode, the utilization of the node (ratio of used
space at the node to total capacity of the node) differs from the
utilization of the whole cluster (ratio of used space in the clus-
ter to total capacity of the cluster) by no more than the thresh-
old value.
The tool is deployed as an application program that can be
run by the cluster administrator. It iteratively moves replicas
from DataNodes with higher utilization to DataNodes with
lower utilization. One key requirement for the balancer is to
maintain data availability. When choosing a replica to move
and deciding its destination, the balancer guarantees that the
decision does not reduce either the number of replicas or the
number of racks.
The balancer optimizes the balancing process by minimiz-
ing the inter-rack data copying. If the balancer decides that a
replica A needs to be moved to a different rack and the destina-
tion rack happens to have a replica B of the same block, the
data will be copied from replica B instead of replica A.
A second configuration parameter limits the bandwidth
consumed by rebalancing operations. The higher the allowed
bandwidth, the faster a cluster can reach the balanced state, but
with greater competition with application processes.
E. Block Scanner
Each DataNode runs a block scanner that periodically scans
its block replicas and verifies that stored checksums match the
block data. In each scan period, the block scanner adjusts the
read bandwidth in order to complete the verification in a con-
figurable period. If a client reads a complete block and check-
sum verification succeeds, it informs the DataNode. The
DataNode treats it as a verification of the replica.
The verification time of each block is stored in a human
readable log file. At any time there are up to two files in top-
level DataNode directory, current and prev logs. New verifica-
tion times are appended to current file. Correspondingly each
DataNode has an in-memory scanning list ordered by the rep-
lica’s verification time.
Whenever a read client or a block scanner detects a corrupt
block, it notifies the NameNode. The NameNode marks the
replica as corrupt, but does not schedule deletion of the replica
immediately. Instead, it starts to replicate a good copy of the
block. Only when the good replica count reaches the replication
factor of the block the corrupt replica is scheduled to be re-
moved. This policy aims to preserve data as long as possible.
So even if all replicas of a block are corrupt, the policy allows
the user to retrieve its data from the corrupt replicas.
F. Decommissioing
The cluster administrator specifies which nodes can join the
cluster by listing the host addresses of nodes that are permitted
to register and the host addresses of nodes that are not permit-
ted to register. The administrator can command the system to
re-evaluate these include and exclude lists. A present member
of the cluster that becomes excluded is marked for decommis-
sioning. Once a DataNode is marked as decommissioning, it
will not be selected as the target of replica placement, but it
will continue to serve read requests. The NameNode starts to
schedule replication of its blocks to other DataNodes. Once the
NameNode detects that all blocks on the decommissioning
DataNode are replicated, the node enters the decommissioned
state. Then it can be safely removed from the cluster without
jeopardizing any data availability.
G. Inter-Cluster Data Copy
When working with large datasets, copying data into and
out of a HDFS cluster is daunting. HDFS provides a tool called
DistCp for large inter/intra-cluster parallel copying. It is a
MapReduce job; each of the map tasks copies a portion of the
source data into the destination file system. The MapReduce
framework automatically handles parallel task scheduling, error
detection and recovery.
IV. PRACTICE AT YAHOO!
Large HDFS clusters at Yahoo! include about 3500 nodes.
A typical cluster node has:
· 2 quad core Xeon processors @ 2.5ghz
· Red Hat Enterprise Linux Server Release 5.1
· Sun Java JDK 1.6.0_13-b03
· 4 directly attached SATA drives (one terabyte each)
· 16G RAM
· 1-gigabit Ethernet
Seventy percent of the disk space is allocated to HDFS. The
remainder is reserved for the operating system (Red Hat
Linux), logs, and space to spill the output of map tasks.
(MapReduce intermediate data are not stored in HDFS.) Forty
nodes in a single rack share an IP switch. The rack switches are
connected to each of eight core switches. The core switches
provide connectivity between racks and to out-of-cluster re-
sources. For each cluster, the NameNode and the BackupNode
hosts are specially provisioned with up to 64GB RAM; applica-
tion tasks are never assigned to those hosts. In total, a cluster of
3500 nodes has 9.8 PB of storage available as blocks that are
replicated three times yielding a net 3.3 PB of storage for user
applications. As a convenient approximation, one thousand
nodes represent one PB of application storage. Over the years
that HDFS has been in use (and into the future), the hosts se-
lected as cluster nodes benefit from improved technologies.
New cluster nodes always have faster processors, bigger disks
and larger RAM. Slower, smaller nodes are retired or relegated
to clusters reserved for development and testing of Hadoop.
The choice of how to provision a cluster node is largely an is-
sue of economically purchasing computation and storage.
HDFS does not compel a particular ratio of computation to
storage, or set a limit on the amount of storage attached to a
cluster node.
On an example large cluster (3500 nodes), there are about
60 million files. Those files have 63 million blocks. As each

8
block typically is replicated three times, every data node hosts
54 000 block replicas. Each day user applications will create
two million new files on the cluster. The 25 000 nodes in
Hadoop clusters at Yahoo! provide 25 PB of on-line data stor-
age. At the start of 2010, this is a modest—but growing—
fraction of the data processing infrastructure at Yahoo!. Yahoo!
began to investigate MapReduce programming with a distrib-
uted file system in 2004. The Apache Hadoop project was
founded in 2006. By the end of that year, Yahoo! had adopted
Hadoop for internal use and had a 300-node cluster for devel-
opment. Since then HDFS has become integral to the back of-
fice at Yahoo!. The flagship application for HDFS has been the
production of the Web Map, an index of the World Wide Web
that is a critical component of search (75 hours elapsed time,
500 terabytes of MapReduce intermediate data, 300 terabytes
total output). More applications are moving to Hadoop, espe-
cially those that analyze and model user behavior.
Becoming a key component of Yahoo!’s technology suite
meant tackling technical problems that are the difference be-
tween being a research project and being the custodian of many
petabytes of corporate data. Foremost are issues of robustness
and durability of data. But also important are economical per-
formance, provisions for resource sharing among members of
the user community, and ease of administration by the system
operators.
A. Durability of Data
Replication of data three times is a robust guard against loss
of data due to uncorrelated node failures. It is unlikely Yahoo!
has ever lost a block in this way; for a large cluster, the prob-
ability of losing a block during one year is less than .005. The
key understanding is that about 0.8 percent of nodes fail each
month. (Even if the node is eventually recovered, no effort is
taken to recover data it may have hosted.) So for the sample
large cluster as described above, a node or two is lost each day.
That same cluster will re-create the 54 000 block replicas
hosted on a failed node in about two minutes. (Re-replication is
fast because it is a parallel problem that scales with the size of
the cluster.) The probability of several nodes failing within two
minutes such that all replicas of some block are lost is indeed
small.
Correlated failure of nodes is a different threat. The most
commonly observed fault in this regard is the failure of a rack
or core switch. HDFS can tolerate losing a rack switch (each
block has a replica on some other rack). Some failures of a core
switch can effectively disconnect a slice of the cluster from
multiple racks, in which case it is probable that some blocks
will become unavailable. In either case, repairing the switch
restores unavailable replicas to the cluster. Another kind of
correlated failure is the accidental or deliberate loss of electri-
cal power to the cluster. If the loss of power spans racks, it is
likely that some blocks will become unavailable. But restoring
power may not be a remedy because one-half to one percent of
the nodes will not survive a full power-on restart. Statistically,
and in practice, a large cluster will lose a handful of blocks
during a power-on restart. (The strategy of deliberately restart-
ing one node at a time over a period of weeks to identify nodes
that will not survive a restart has not been tested.)
In addition to total failures of nodes, stored data can be
corrupted or lost. The block scanner scans all blocks in a large
cluster each fortnight and finds about 20 bad replicas in the
process.
B. Caring for the Commons
As the use of HDFS has grown, the file system itself has
had to introduce means to share the resource within a large and
diverse user community. The first such feature was a permis-
sions framework closely modeled on the Unix permissions
scheme for file and directories. In this framework, files and
directories have separate access permissions for the owner, for
other members of the user group associated with the file or
directory, and for all other users. The principle differences be-
tween Unix (POSIX) and HDFS are that ordinary files in
HDFS have neither “execute” permissions nor “sticky” bits.
In the present permissions framework, user identity is
weak: you are who your host says you are. When accessing
HDFS, the application client simply queries the local operating
system for user identity and group membership. A stronger
identity model is under development. In the new framework,
the application client must present to the name system creden-
tials obtained from a trusted source. Different credential ad-
ministrations are possible; the initial implementation will use
Kerberos. The user application can use the same framework to
confirm that the name system also has a trustworthy identity.
And the name system also can demand credentials from each of
the data nodes participating in the cluster.
The total space available for data storage is set by the num-
ber of data nodes and the storage provisioned for each node.
Early experience with HDFS demonstrated a need for some
means to enforce the resource allocation policy across user
communities. Not only must fairness of sharing be enforced,
but when a user application might involve thousands of hosts
writing data, protection against application inadvertently ex-
hausting resources is also important. For HDFS, because the
system metadata are always in RAM, the size of the namespace
(number of files and directories) is also a finite resource. To
manage storage and namespace resources, each directory may
be assigned a quota for the total space occupied by files in the
sub-tree of the namespace beginning at that directory. A sepa-
rate quota may also be set for the total number of files and di-
rectories in the sub-tree.
While the architecture of HDFS presumes most applications
will stream large data sets as input, the MapReduce program-
ming framework can have a tendency to generate many small
output files (one from each reduce task) further stressing the
namespace resource. As a convenience, a directory sub-tree can
be collapsed into a single Hadoop Archive file. A HAR file is
similar to a familiar tar, JAR, or Zip file, but file system opera-
tion can address the individual files for the archive, and a HAR
file can be used transparently as the input to a MapReduce job.
C. Benchmarks
A design goal of HDFS is to provide very high I/O band-
width for large data sets. There are three kinds of measure-
ments that test that goal.

9
• What is bandwidth observed from a contrived bench-
mark?
• What bandwidth is observed in a production cluster
with a mix of user jobs?
• What bandwidth can be obtained by the most carefully
constructed large-scale user application?
The statistics reported here were obtained from clusters of
at least 3500 nodes. At this scale, total bandwidth is linear with
the number of nodes, and so the interesting statistic is the
bandwidth per node. These benchmarks are available as part of
the Hadoop codebase.
The DFSIO benchmark measures average throughput for
read, write and append operations. DFSIO is an application
available as part of the Hadoop distribution. This MapReduce
program reads/writes/appends random data from/to large files.
Each map task within the job executes the same operation on a
distinct file, transfers the same amount of data, and reports its
transfer rate to the single reduce task. The reduce task then
summarizes the measurements. The test is run without conten-
tion from other applications, and the number of map tasks is
chosen to be proportional to the cluster size. It is designed to
measure performance only during data transfer, and excludes
the overheads of task scheduling, startup, and the reduce task.
• DFSIO Read: 66 MB /s per node
• DFSIO Write: 40 MB /s per node
For a production cluster, the number of bytes read and writ-
ten is reported to a metrics collection system. These averages
are taken over a few weeks and represent the utilization of the
cluster by jobs from hundreds of individual users. On average
each node was occupied by one or two application tasks at any
moment (fewer than the number of processor cores available).
• Busy Cluster Read: 1.02 MB/s per node
• Busy Cluster Write: 1.09 MB/s per node
Table 2. Sort benchmark for one terabyte and one petabyte of
data. Each data record is 100 bytes with a 10-byte key. The
test program is a general sorting procedure that is not special-
ized for the record size. In the terabyte sort, the block replica-
tion factor was set to one, a modest advantage for a short test.
In the petabyte sort, the replication factor was set to two so
that the test would confidently complete in case of a (not un-
expected) node failure.
At the beginning of 2009, Yahoo! participated in the Gray
Sort competition [9]. The nature of this task stresses the sys-
tem’s ability to move data from and to the file system (it really
isn't about sorting). The competitive aspect means that the re-
sults in Table 2 are about the best a user application can
achieve with the current design and hardware. The I/O rate in
the last column is the combination of reading the input and
writing the output from and to HDFS. In the second row, while
the rate for HDFS is reduced, the total I/O per node will be
about double because for the larger (petabyte!) data set, the
MapReduce intermediates must also be written to and read
from disk. In the smaller test, there is no need to spill the
MapReduce intermediates to disk; they are buffered the mem-
ory of the tasks.
Large clusters require that the HDFS NameNode support
the number of client operations expected in a large cluster. The
NNThroughput benchmark is a single node process which
starts the NameNode application and runs a series of client
threads on the same node. Each client thread performs the same
NameNode operation repeatedly by directly calling the Name-
Node method implementing this operation. The benchmark
measures the number of operations per second performed by
the NameNode. The benchmark is designed to avoid communi-
cation overhead caused by RPC connections and serialization,
and therefore runs clients locally rather than remotely from
different nodes. This provides the upper bound of pure
NameNode performance.
Operation Throughput (ops/s)
Open file for read 126 100
Create file 5600
Rename file 8300
Delete file 20 700
DataNode Heartbeat 300 000
Blocks report (blocks/s) 639 700
Table 3. NNThroughput benchmark
V. FUTURE WORK
This section presents some of the future work that the
Hadoop team at Yahoo is considering; Hadoop being an open
source project implies that new features and changes are de-
cided by the Hadoop development community at large.
The Hadoop cluster is effectively unavailable when its
NameNode is down. Given that Hadoop is used primarily as a
batch system, restarting the NameNode has been a satisfactory
recovery means. However, we have taken steps towards auto-
mated failover. Currently a BackupNode receives all transac-
tions from the primary NameNode. This will allow a failover to
a warm or even a hot BackupNode if we send block reports to
both the primary NameNode and BackupNode. A few Hadoop
users outside Yahoo! have experimented with manual failover.
Our plan is to use Zookeeper, Yahoo’s distributed consensus
technology to build an automated failover solution.
Scalability of the NameNode [13] has been a key struggle.
Because the NameNode keeps all the namespace and block
locations in memory, the size of the NameNode heap has lim-
ited the number of files and also the number of blocks address-
able. The main challenge with the NameNode has been that
when its memory usage is close to the maximum the
NameNode becomes unresponsive due to Java garbage collec-
tion and sometimes requires a restart. While we have encour-
HDFS I/0 Bytes/s
Bytes
(TB)
Nodes Maps Reduces Time Aggregate
(GB)
Per
Node
(MB)
1 1460 8000 2700 62 s 32 22.1
1000 3658 80 000 20 000 58 500 s 34.2 9.35

10
aged our users to create larger files, this has not happened since
it would require changes in application behavior. We have
added quotas to manage the usage and have provided an ar-
chive tool. However these do not fundamentally address the
scalability problem.
Our near-term solution to scalability is to allow multiple
namespaces (and NameNodes) to share the physical storage
within a cluster. We are extending our block IDs to be prefixed
by block pool identifiers. Block pools are analogous to LUNs
in a SAN storage system and a namespace with its pool of
blocks is analogous as a file system volume.
This approach is fairly simple and requires minimal
changes to the system. It offers a number of advantages besides
scalability: it isolates namespaces of different sets of applica-
tions and improves the overall availability of the cluster. It also
generalizes the block storage abstraction to allow other services
to use the block storage service with perhaps a different name-
space structure. We plan to explore other approaches to scaling
such as storing only partial namespace in memory and truly
distributed implementation of the NameNode in the future. In
particular, our assumption that applications will create a small
number of large files was flawed. As noted earlier, changing
application behavior is hard. Furthermore, we are seeing new
classes of applications for HDFS that need to store a large
number of smaller files.
The main drawback of multiple independent namespaces is
the cost of managing them, especially if the number of name-
spaces is large. We are also planning to use application or job
centric namespaces rather than cluster centric namespaces—
this is analogous to the per-process namespaces that are used to
deal with remote execution in distributed systems in the late
80s and early 90s [10][11][12].
Currently our clusters are less than 4000 nodes. We believe
we can scale to much larger clusters with the solutions outlined
above. However, we believe it is prudent to have multiple clus-
ters rather than a single large cluster (say three 6000-node clus-
ters rather than a single 18 000-node cluster) as it allows much
improved availability and isolation. To that end we are plan-
ning to provide greater cooperation between clusters. For ex-
ample caching remotely accessed files or reducing the replica-
tion factor of blocks when files sets are replicated across clus-
ters.
VI. ACKNOWLEDGMENT
We would like to thank all members of the HDFS team at
Yahoo! present and past for their hard work building the file
system. We would like to thank all Hadoop committers and
collaborators for their valuable contributions. Corinne Chandel
drew illustrations for this paper.
REFERENCES
[1] Apache Hadoop. http://hadoop.apache.org/
[2] P. H. Carns, W. B. Ligon III, R. B. Ross, and R. Thakur. “PVFS: A
parallel file system for Linux clusters,” in Proc. of 4th Annual Linux
Showcase and Conference, 2000, pp. 317–327.
[3] J. Dean, S. Ghemawat, “MapReduce: Simplified Data Processing on
Large Clusters,” In Proc. of the 6th Symposium on Operating Systems
Design and Implementation, San Francisco CA, Dec. 2004.
[4] A. Gates, O. Natkovich, S. Chopra, P. Kamath, S. Narayanam, C.
Olston, B. Reed, S. Srinivasan, U. Srivastava. “Building a High-Level
Dataflow System on top of MapReduce: The Pig Experience,” In Proc.
of Very Large Data Bases, vol 2 no. 2, 2009, pp. 1414–1425
[5] S. Ghemawat, H. Gobioff, S. Leung. “The Google file system,” In Proc.
of ACM Symposium on Operating Systems Principles, Lake George,
NY, Oct 2003, pp 29–43.
[6] F. P. Junqueira, B. C. Reed. “The life and times of a zookeeper,” In
Proc. of the 28th ACM Symposium on Principles of Distributed
Computing, Calgary, AB, Canada, August 10–12, 2009.
[7] Lustre File System. http://www.lustre.org
[8] M. K. McKusick, S. Quinlan. “GFS: Evolution on Fast-forward,” ACM
Queue, vol. 7, no. 7, New York, NY. August 2009.
[9] O. O'Malley, A. C. Murthy. Hadoop Sorts a Petabyte in 16.25 Hours and
a Terabyte in 62 Seconds. May 2009.
http://developer.yahoo.net/blogs/hadoop/2009/05/hadoop_sorts_a_petab
yte_in_162.html
[10] R. Pike, D. Presotto, K. Thompson, H. Trickey, P. Winterbottom, “Use
of Name Spaces in Plan9,” Operating Systems Review, 27(2), April
1993, pages 72–76.
[11] S. Radia, "Naming Policies in the spring system," In Proc. of 1st IEEE
Workshop on Services in Distributed and Networked Environments,
June 1994, pp. 164–171.
[12] S. Radia, J. Pachl, “The Per-Process View of Naming and Remote
Execution,” IEEE Parallel and Distributed Technology, vol. 1, no. 3,
August 1993, pp. 71–80.
[13] K. V. Shvachko, “HDFS Scalability: The limits to growth,” ;login:.
April 2010, pp. 6–16.
[14] W. Tantisiriroj, S. Patil, G. Gibson. “Data-intensive file systems for
Internet services: A rose by any other name ...” Technical Report CMU-
PDL-08-114, Parallel Data Laboratory, Carnegie Mellon University,
Pittsburgh, PA, October 2008.
[15] A. Thusoo, J. S. Sarma, N. Jain, Z. Shao, P. Chakka, S. Anthony, H. Liu,
P. Wyckoff, R. Murthy, “Hive – A Warehousing Solution Over a Map-
Reduce Framework,” In Proc. of Very Large Data Bases, vol. 2 no. 2,
August 2009, pp. 1626-1629.
[16] J. Venner, Pro Hadoop. Apress, June 22, 2009.
[17] S. Weil, S. Brandt, E. Miller, D. Long, C. Maltzahn, “Ceph: A Scalable,
High-Performance Distributed File System,” In Proc. of the 7th
Symposium on Operating Systems Design and Implementation, Seattle,
WA, November 2006.
[18] B. Welch, M. Unangst, Z. Abbasi, G. Gibson, B. Mueller, J. Small, J.
Zelenka, B. Zhou, “Scalable Performance of the Panasas Parallel file
System”, In Proc. of the 6th
USENIX Conference on File and Storage
Technologies, San Jose, CA, February 2008
[19] T. White, Hadoop: The Definitive Guide. O'Reilly Media, Yahoo! Press,
June 5, 2009.

Hadoop

More Related Content

What's hot (20)

Similar to Hadoop (20)

More from Shahbaz Sidhu (20)

Recently uploaded (20)

Hadoop