<s>
Apache	B-Application
Hadoop	I-Application
(	O
)	O
is	O
a	O
collection	O
of	O
open-source	B-Application
software	I-Application
utilities	O
that	O
facilitates	O
using	O
a	O
network	O
of	O
many	O
computers	O
to	O
solve	O
problems	O
involving	O
massive	O
amounts	O
of	O
data	O
and	O
computation	O
.	O
</s>
<s>
It	O
provides	O
a	O
software	B-Architecture
framework	I-Architecture
for	O
distributed	B-Application
storage	I-Application
and	O
processing	O
of	O
big	B-Application
data	I-Application
using	O
the	O
MapReduce	B-Operating_System
programming	O
model	O
.	O
</s>
<s>
Hadoop	B-Application
was	O
originally	O
designed	O
for	O
computer	B-Architecture
clusters	I-Architecture
built	O
from	O
commodity	B-Device
hardware	I-Device
,	O
which	O
is	O
still	O
the	O
common	O
use	O
.	O
</s>
<s>
All	O
the	O
modules	O
in	O
Hadoop	B-Application
are	O
designed	O
with	O
a	O
fundamental	O
assumption	O
that	O
hardware	O
failures	O
are	O
common	O
occurrences	O
and	O
should	O
be	O
automatically	O
handled	O
by	O
the	O
framework	O
.	O
</s>
<s>
The	O
core	O
of	O
Apache	B-Application
Hadoop	I-Application
consists	O
of	O
a	O
storage	O
part	O
,	O
known	O
as	O
Hadoop	B-Application
Distributed	B-Protocol
File	I-Protocol
System	I-Protocol
(	O
HDFS	O
)	O
,	O
and	O
a	O
processing	O
part	O
which	O
is	O
a	O
MapReduce	B-Operating_System
programming	O
model	O
.	O
</s>
<s>
Hadoop	B-Application
splits	O
files	O
into	O
large	O
blocks	O
and	O
distributes	O
them	O
across	O
nodes	O
in	O
a	O
cluster	O
.	O
</s>
<s>
It	O
then	O
transfers	O
packaged	B-Language
code	I-Language
into	O
nodes	O
to	O
process	O
the	O
data	O
in	O
parallel	O
.	O
</s>
<s>
This	O
approach	O
takes	O
advantage	O
of	O
data	B-General_Concept
locality	I-General_Concept
,	O
where	O
nodes	O
manipulate	O
the	O
data	O
they	O
have	O
access	O
to	O
.	O
</s>
<s>
This	O
allows	O
the	O
dataset	O
to	O
be	O
processed	B-Architecture
faster	O
and	O
more	O
efficiently	O
than	O
it	O
would	O
be	O
in	O
a	O
more	O
conventional	O
supercomputer	B-Operating_System
architecture	I-Operating_System
that	O
relies	O
on	O
a	O
parallel	B-Application
file	I-Application
system	I-Application
where	O
computation	O
and	O
data	O
are	O
distributed	O
via	O
high-speed	O
networking	O
.	O
</s>
<s>
The	O
base	O
Apache	B-Application
Hadoop	I-Application
framework	O
is	O
composed	O
of	O
the	O
following	O
modules	O
:	O
</s>
<s>
Hadoop	B-Application
Common	O
–	O
contains	O
libraries	O
and	O
utilities	O
needed	O
by	O
other	O
Hadoop	B-Application
modules	O
;	O
</s>
<s>
Hadoop	B-Application
Distributed	B-Protocol
File	I-Protocol
System	I-Protocol
(	O
HDFS	O
)	O
–	O
a	O
distributed	O
file-system	O
that	O
stores	O
data	O
on	O
commodity	O
machines	O
,	O
providing	O
very	O
high	O
aggregate	O
bandwidth	O
across	O
the	O
cluster	O
;	O
</s>
<s>
Hadoop	B-Application
YARN	I-Application
–	O
(	O
introduced	O
in	O
2012	O
)	O
a	O
platform	O
responsible	O
for	O
managing	O
computing	O
resources	O
in	O
clusters	O
and	O
using	O
them	O
for	O
scheduling	O
users	O
 '	O
applications	O
;	O
</s>
<s>
Hadoop	B-Application
MapReduce	B-Operating_System
–	O
an	O
implementation	O
of	O
the	O
MapReduce	B-Operating_System
programming	O
model	O
for	O
large-scale	O
data	O
processing	O
.	O
</s>
<s>
The	O
term	O
Hadoop	B-Application
is	O
often	O
used	O
for	O
both	O
base	O
modules	O
and	O
sub-modules	O
and	O
also	O
the	O
ecosystem	O
,	O
or	O
collection	O
of	O
additional	O
software	O
packages	O
that	O
can	O
be	O
installed	O
on	O
top	O
of	O
or	O
alongside	O
Hadoop	B-Application
,	O
such	O
as	O
Apache	B-Language
Pig	I-Language
,	O
Apache	B-General_Concept
Hive	I-General_Concept
,	O
Apache	B-Language
HBase	I-Language
,	O
Apache	B-Application
Phoenix	I-Application
,	O
Apache	B-Language
Spark	I-Language
,	O
Apache	B-Language
ZooKeeper	I-Language
,	O
Apache	B-Language
Impala	I-Language
,	O
Apache	B-Language
Flume	I-Language
,	O
Apache	B-Language
Sqoop	I-Language
,	O
Apache	B-Language
Oozie	I-Language
,	O
and	O
Apache	B-General_Concept
Storm	I-General_Concept
.	O
</s>
<s>
Apache	B-Application
Hadoop	I-Application
's	O
MapReduce	B-Operating_System
and	O
HDFS	O
components	O
were	O
inspired	O
by	O
Google	B-Application
papers	O
on	O
MapReduce	B-Operating_System
and	O
Google	B-Application
File	I-Application
System	I-Application
.	O
</s>
<s>
The	O
Hadoop	B-Application
framework	O
itself	O
is	O
mostly	O
written	O
in	O
the	O
Java	B-Language
programming	I-Language
language	I-Language
,	O
with	O
some	O
native	O
code	O
in	O
C	B-Language
and	O
command	B-Application
line	I-Application
utilities	O
written	O
as	O
shell	B-Language
scripts	I-Language
.	O
</s>
<s>
Though	O
MapReduce	B-Operating_System
Java	B-Language
code	I-Language
is	O
common	O
,	O
any	O
programming	O
language	O
can	O
be	O
used	O
with	O
Hadoop	B-Application
Streaming	O
to	O
implement	O
the	O
map	O
and	O
reduce	O
parts	O
of	O
the	O
user	O
's	O
program	O
.	O
</s>
<s>
Other	O
projects	O
in	O
the	O
Hadoop	B-Application
ecosystem	O
expose	O
richer	O
user	O
interfaces	O
.	O
</s>
<s>
According	O
to	O
its	O
co-founders	O
,	O
Doug	O
Cutting	O
and	O
Mike	O
Cafarella	O
,	O
the	O
genesis	O
of	O
Hadoop	B-Application
was	O
the	B-Application
Google	I-Application
File	B-Protocol
System	I-Protocol
paper	O
that	O
was	O
published	O
in	O
October	O
2003	O
.	O
</s>
<s>
This	O
paper	O
spawned	O
another	O
one	O
from	O
Google	B-Application
"	O
MapReduce	B-Operating_System
:	O
Simplified	O
Data	O
Processing	O
on	O
Large	O
Clusters	O
"	O
.	O
</s>
<s>
Development	O
started	O
on	O
the	O
Apache	B-Language
Nutch	I-Language
project	O
,	O
but	O
was	O
moved	O
to	O
the	O
new	O
Hadoop	B-Application
subproject	O
in	O
January	O
2006	O
.	O
</s>
<s>
Doug	O
Cutting	O
,	O
who	O
was	O
working	O
at	O
Yahoo	B-Application
!	I-Application
</s>
<s>
The	O
initial	O
code	O
that	O
was	O
factored	O
out	O
of	O
Nutch	B-Language
consisted	O
of	O
about	O
5,000	O
lines	O
of	O
code	O
for	O
HDFS	O
and	O
about	O
6,000	O
lines	O
of	O
code	O
for	O
MapReduce	B-Operating_System
.	O
</s>
<s>
In	O
March	O
2006	O
,	O
Owen	O
O'Malley	O
was	O
the	O
first	O
committer	O
to	O
add	O
to	O
the	O
Hadoop	B-Application
project	O
;	O
Hadoop	B-Application
0.1.0	O
was	O
released	O
in	O
April	O
2006	O
.	O
</s>
<s>
The	O
very	O
first	O
design	O
document	O
for	O
the	O
Hadoop	B-Application
Distributed	B-Protocol
File	I-Protocol
System	I-Protocol
was	O
written	O
by	O
Dhruba	O
Borthakur	O
in	O
2007	O
.	O
</s>
<s>
Hadoop	B-Application
consists	O
of	O
the	O
Hadoop	B-Application
Common	O
package	O
,	O
which	O
provides	O
file	B-Protocol
system	I-Protocol
and	O
operating	B-General_Concept
system	I-General_Concept
level	O
abstractions	O
,	O
a	O
MapReduce	B-Operating_System
engine	O
(	O
either	O
MapReduce/MR1	O
or	O
YARN/MR2	O
)	O
and	O
the	O
Hadoop	B-Application
Distributed	B-Protocol
File	I-Protocol
System	I-Protocol
(	O
HDFS	O
)	O
.	O
</s>
<s>
The	O
Hadoop	B-Application
Common	O
package	O
contains	O
the	O
Java	B-Language
Archive	I-Language
(	O
JAR	B-Language
)	O
files	O
and	O
scripts	O
needed	O
to	O
start	O
Hadoop	B-Application
.	O
</s>
<s>
For	O
effective	O
scheduling	O
of	O
work	O
,	O
every	O
Hadoop-compatible	O
file	B-Protocol
system	I-Protocol
should	O
provide	O
location	O
awareness	O
,	O
which	O
is	O
the	O
name	O
of	O
the	O
rack	O
,	O
specifically	O
the	O
network	O
switch	O
where	O
a	O
worker	O
node	O
is	O
.	O
</s>
<s>
Hadoop	B-Application
applications	O
can	O
use	O
this	O
information	O
to	O
execute	O
code	O
on	O
the	O
node	O
where	O
the	O
data	O
is	O
,	O
and	O
,	O
failing	O
that	O
,	O
on	O
the	O
same	O
rack/switch	O
to	O
reduce	O
backbone	O
traffic	O
.	O
</s>
<s>
HDFS	O
uses	O
this	O
method	O
when	O
replicating	B-Architecture
data	O
for	O
data	B-Architecture
redundancy	I-Architecture
across	O
multiple	O
racks	O
.	O
</s>
<s>
A	O
small	O
Hadoop	B-Application
cluster	O
includes	O
a	O
single	O
master	O
and	O
multiple	O
worker	O
nodes	O
.	O
</s>
<s>
Hadoop	B-Application
requires	O
the	O
Java	B-Language
Runtime	I-Language
Environment	O
(	O
JRE	O
)	O
1.6	O
or	O
higher	O
.	O
</s>
<s>
The	O
standard	O
startup	O
and	O
shutdown	O
scripts	O
require	O
that	O
Secure	B-Protocol
Shell	I-Protocol
(	O
SSH	B-Protocol
)	O
be	O
set	O
up	O
between	O
nodes	O
in	O
the	O
cluster	O
.	O
</s>
<s>
In	O
a	O
larger	O
cluster	O
,	O
HDFS	O
nodes	O
are	O
managed	O
through	O
a	O
dedicated	O
NameNode	O
server	O
to	O
host	O
the	O
file	B-Protocol
system	I-Protocol
index	O
,	O
and	O
a	O
secondary	O
NameNode	O
that	O
can	O
generate	O
snapshots	O
of	O
the	O
namenode	O
's	O
memory	O
structures	O
,	O
thereby	O
preventing	O
file-system	O
corruption	O
and	O
loss	O
of	O
data	O
.	O
</s>
<s>
When	O
Hadoop	B-Application
MapReduce	B-Operating_System
is	O
used	O
with	O
an	O
alternate	O
file	B-Protocol
system	I-Protocol
,	O
the	O
NameNode	O
,	O
secondary	O
NameNode	O
,	O
and	O
DataNode	O
architecture	O
of	O
HDFS	O
are	O
replaced	O
by	O
the	O
file-system-specific	O
equivalents	O
.	O
</s>
<s>
The	O
Hadoop	B-Application
distributed	B-Protocol
file	I-Protocol
system	I-Protocol
(	O
HDFS	O
)	O
is	O
a	O
distributed	O
,	O
scalable	O
,	O
and	O
portable	O
file	B-Protocol
system	I-Protocol
written	O
in	O
Java	B-Language
for	O
the	O
Hadoop	B-Application
framework	O
.	O
</s>
<s>
Some	O
consider	O
it	O
to	O
instead	O
be	O
a	O
data	O
store	O
due	O
to	O
its	O
lack	O
of	O
POSIX	O
compliance	O
,	O
but	O
it	O
does	O
provide	O
shell	O
commands	O
and	O
Java	B-Language
application	O
programming	O
interface	O
(	O
API	O
)	O
methods	B-Language
that	O
are	O
similar	O
to	O
other	O
file	O
systems	O
.	O
</s>
<s>
A	O
Hadoop	B-Application
instance	O
is	O
divided	O
into	O
HDFS	O
and	O
MapReduce	B-Operating_System
.	O
</s>
<s>
HDFS	O
is	O
used	O
for	O
storing	O
the	O
data	O
and	O
MapReduce	B-Operating_System
is	O
used	O
for	O
processing	O
data	O
.	O
</s>
<s>
The	O
master	O
node	O
can	O
track	O
files	O
,	O
manage	O
the	O
file	B-Protocol
system	I-Protocol
and	O
has	O
the	O
metadata	O
of	O
all	O
of	O
the	O
stored	O
data	O
within	O
it	O
.	O
</s>
<s>
In	O
particular	O
,	O
the	O
name	O
node	O
contains	O
the	O
details	O
of	O
the	O
number	O
of	O
blocks	O
,	O
locations	O
of	O
the	O
data	O
node	O
that	O
the	O
data	O
is	O
stored	O
in	O
,	O
where	O
the	O
replications	B-Architecture
are	O
stored	O
,	O
and	O
other	O
details	O
.	O
</s>
<s>
In	O
this	O
way	O
when	O
Name	O
Node	O
does	O
not	O
receive	O
a	O
heartbeat	O
from	O
a	O
data	O
node	O
for	O
2	O
minutes	O
,	O
it	O
will	O
take	O
that	O
data	O
node	O
as	O
dead	O
and	O
starts	O
the	O
process	O
of	O
block	O
replications	B-Architecture
on	O
some	O
other	O
Data	O
node	O
.	O
</s>
<s>
Secondary	O
Name	O
Node	O
:	O
This	O
is	O
only	O
to	O
take	O
care	O
of	O
the	O
checkpoints	O
of	O
the	O
file	B-Protocol
system	I-Protocol
metadata	O
which	O
is	O
in	O
the	O
Name	O
Node	O
.	O
</s>
<s>
Job	O
Tracker	O
:	O
Job	O
Tracker	O
receives	O
the	O
requests	O
for	O
Map	B-Operating_System
Reduce	I-Operating_System
execution	O
from	O
the	O
client	O
.	O
</s>
<s>
Hadoop	B-Application
cluster	O
has	O
nominally	O
a	O
single	O
namenode	O
plus	O
a	O
cluster	O
of	O
datanodes	O
,	O
although	O
redundancy	O
options	O
are	O
available	O
for	O
the	O
namenode	O
due	O
to	O
its	O
criticality	O
.	O
</s>
<s>
The	O
file	B-Protocol
system	I-Protocol
uses	O
TCP/IP	B-Protocol
sockets	B-Protocol
for	O
communication	O
.	O
</s>
<s>
Clients	O
use	O
remote	B-Operating_System
procedure	I-Operating_System
calls	I-Operating_System
(	O
RPC	O
)	O
to	O
communicate	O
with	O
each	O
other	O
.	O
</s>
<s>
It	O
achieves	O
reliability	O
by	O
replicating	B-Architecture
the	O
data	O
across	O
multiple	O
hosts	O
,	O
and	O
hence	O
theoretically	O
does	O
not	O
require	O
redundant	B-Architecture
array	I-Architecture
of	I-Architecture
independent	I-Architecture
disks	I-Architecture
(	O
RAID	B-Architecture
)	O
storage	O
on	O
hosts	O
(	O
but	O
to	O
increase	O
input-output	O
(	O
I/O	O
)	O
performance	O
some	O
RAID	B-Architecture
configurations	O
are	O
still	O
useful	O
)	O
.	O
</s>
<s>
With	O
the	O
default	O
replication	B-Architecture
value	O
,	O
3	O
,	O
data	O
is	O
stored	O
on	O
three	O
nodes	O
:	O
two	O
on	O
the	O
same	O
rack	O
,	O
and	O
one	O
on	O
a	O
different	O
rack	O
.	O
</s>
<s>
Data	O
nodes	O
can	O
talk	O
to	O
each	O
other	O
to	O
rebalance	O
data	O
,	O
to	O
move	O
copies	O
around	O
,	O
and	O
to	O
keep	O
the	O
replication	B-Architecture
of	O
data	O
high	O
.	O
</s>
<s>
HDFS	O
is	O
not	O
fully	O
POSIX-compliant	O
,	O
because	O
the	O
requirements	O
for	O
a	O
POSIX	O
file-system	O
differ	O
from	O
the	O
target	O
goals	O
of	O
a	O
Hadoop	B-Application
application	O
.	O
</s>
<s>
In	O
May	O
2012	O
,	O
high-availability	O
capabilities	O
were	O
added	O
to	O
HDFS	O
,	O
letting	O
the	O
main	O
metadata	O
server	O
called	O
the	O
NameNode	O
manually	O
fail-over	B-General_Concept
onto	O
a	O
backup	O
.	O
</s>
<s>
The	O
project	O
has	O
also	O
started	O
developing	O
automatic	O
fail-overs	B-General_Concept
.	O
</s>
<s>
The	O
HDFS	O
file	B-Protocol
system	I-Protocol
includes	O
a	O
so-called	O
secondary	O
namenode	O
,	O
a	O
misleading	O
term	O
that	O
some	O
might	O
incorrectly	O
interpret	O
as	O
a	O
backup	O
namenode	O
when	O
the	O
primary	O
namenode	O
goes	O
offline	O
.	O
</s>
<s>
For	O
example	O
:	O
if	O
node	O
A	O
contains	O
data	O
(	O
a	O
,	O
b	O
,	O
c	B-Language
)	O
and	O
node	O
X	O
contains	O
data	O
(	O
x	O
,	O
y	O
,	O
z	O
)	O
,	O
the	O
job	O
tracker	O
schedules	O
node	O
A	O
to	O
perform	O
map	O
or	O
reduce	O
tasks	O
on	O
(	O
a	O
,	O
b	O
,	O
c	B-Language
)	O
and	O
node	O
X	O
would	O
be	O
scheduled	O
to	O
perform	O
map	O
or	O
reduce	O
tasks	O
on	O
(	O
x	O
,	O
y	O
,	O
z	O
)	O
.	O
</s>
<s>
When	O
Hadoop	B-Application
is	O
used	O
with	O
other	O
file	O
systems	O
,	O
this	O
advantage	O
is	O
not	O
always	O
available	O
.	O
</s>
<s>
This	O
can	O
have	O
a	O
significant	O
impact	O
on	O
job-completion	O
times	O
as	O
demonstrated	O
with	O
data-intensive	B-Operating_System
jobs	O
.	O
</s>
<s>
HDFS	O
was	O
designed	O
for	O
mostly	O
immutable	O
files	O
and	O
may	O
not	O
be	O
suitable	O
for	O
systems	O
requiring	O
concurrent	B-Operating_System
write	O
operations	O
.	O
</s>
<s>
HDFS	O
can	O
be	O
mounted	B-Application
directly	O
with	O
a	O
Filesystem	B-Application
in	I-Application
Userspace	I-Application
(	O
FUSE	B-Application
)	O
virtual	B-Application
file	I-Application
system	I-Application
on	O
Linux	B-Application
and	O
some	O
other	O
Unix	B-Application
systems	I-Application
.	O
</s>
<s>
File	O
access	O
can	O
be	O
achieved	O
through	O
the	O
native	O
Java	B-Language
API	O
,	O
the	O
Thrift	B-Protocol
API	O
(	O
generates	O
a	O
client	O
in	O
a	O
number	O
of	O
languages	O
e.g.	O
</s>
<s>
C++	O
,	O
Java	B-Language
,	O
Python	O
,	O
PHP	O
,	O
Ruby	O
,	O
Erlang	O
,	O
Perl	O
,	O
Haskell	O
,	O
C#	O
,	O
Cocoa	B-Operating_System
,	O
Smalltalk	O
,	O
and	O
OCaml	B-Language
)	O
,	O
the	O
command-line	B-Application
interface	I-Application
,	O
the	O
HDFS-UI	O
web	B-Application
application	I-Application
over	O
HTTP	B-Protocol
,	O
or	O
via	O
3rd-party	O
network	O
client	O
libraries	O
.	O
</s>
<s>
HDFS	O
is	O
designed	O
for	O
portability	O
across	O
various	O
hardware	O
platforms	O
and	O
for	O
compatibility	O
with	O
a	O
variety	O
of	O
underlying	O
operating	B-General_Concept
systems	I-General_Concept
.	O
</s>
<s>
The	O
HDFS	O
design	O
introduces	O
portability	O
limitations	O
that	O
result	O
in	O
some	O
performance	O
bottlenecks	O
,	O
since	O
the	O
Java	B-Language
implementation	O
cannot	O
use	O
features	O
that	O
are	O
exclusive	O
to	O
the	O
platform	O
on	O
which	O
HDFS	O
is	O
running	O
.	O
</s>
<s>
Monitoring	O
end-to-end	O
performance	O
requires	O
tracking	O
metrics	O
from	O
datanodes	O
,	O
namenodes	O
,	O
and	O
the	O
underlying	O
operating	B-General_Concept
system	I-General_Concept
.	O
</s>
<s>
There	O
are	O
currently	O
several	O
monitoring	O
platforms	O
to	O
track	O
HDFS	O
performance	O
,	O
including	O
Hortonworks	O
,	O
Cloudera	B-General_Concept
,	O
and	O
Datadog	O
.	O
</s>
<s>
Hadoop	B-Application
works	O
directly	O
with	O
any	O
distributed	B-Protocol
file	I-Protocol
system	I-Protocol
that	O
can	O
be	O
mounted	B-Application
by	O
the	O
underlying	O
operating	B-General_Concept
system	I-General_Concept
by	O
simply	O
using	O
a	O
file://	O
URL	O
;	O
however	O
,	O
this	O
comes	O
at	O
a	O
price	O
–	O
the	O
loss	O
of	O
locality	O
.	O
</s>
<s>
To	O
reduce	O
network	O
traffic	O
,	O
Hadoop	B-Application
needs	O
to	O
know	O
which	O
servers	O
are	O
closest	O
to	O
the	O
data	O
,	O
information	O
that	O
Hadoop-specific	O
file	B-Protocol
system	I-Protocol
bridges	O
can	O
provide	O
.	O
</s>
<s>
In	O
May	O
2011	O
,	O
the	O
list	O
of	O
supported	O
file	O
systems	O
bundled	O
with	O
Apache	B-Application
Hadoop	I-Application
were	O
:	O
</s>
<s>
HDFS	O
:	O
Hadoop	B-Application
's	O
own	O
rack-aware	O
file	B-Protocol
system	I-Protocol
.	O
</s>
<s>
This	O
is	O
designed	O
to	O
scale	O
to	O
tens	O
of	O
petabytes	O
of	O
storage	O
and	O
runs	O
on	O
top	O
of	O
the	O
file	O
systems	O
of	O
the	O
underlying	O
operating	B-General_Concept
systems	I-General_Concept
.	O
</s>
<s>
Apache	B-Application
Hadoop	I-Application
Ozone	O
:	O
HDFS-compatible	O
object	O
store	O
targeting	O
optimized	O
for	O
billions	O
of	O
small	O
files	O
.	O
</s>
<s>
FTP	B-Protocol
file	B-Protocol
system	I-Protocol
:	O
This	O
stores	O
all	O
its	O
data	O
on	O
remotely	O
accessible	O
FTP	B-Protocol
servers	I-Protocol
.	O
</s>
<s>
Amazon	B-General_Concept
S3	I-General_Concept
(	O
Simple	B-General_Concept
Storage	I-General_Concept
Service	I-General_Concept
)	O
object	O
storage	O
:	O
This	O
is	O
targeted	O
at	O
clusters	O
hosted	O
on	O
the	O
Amazon	B-General_Concept
Elastic	I-General_Concept
Compute	I-General_Concept
Cloud	I-General_Concept
server-on-demand	O
infrastructure	O
.	O
</s>
<s>
There	O
is	O
no	O
rack-awareness	O
in	O
this	O
file	B-Protocol
system	I-Protocol
,	O
as	O
it	O
is	O
all	O
remote	O
.	O
</s>
<s>
Windows	O
Azure	O
Storage	O
Blobs	O
(	O
WASB	O
)	O
file	B-Protocol
system	I-Protocol
:	O
This	O
is	O
an	O
extension	O
of	O
HDFS	O
that	O
allows	O
distributions	O
of	O
Hadoop	B-Application
to	O
access	O
data	O
in	O
Azure	O
blob	O
stores	O
without	O
moving	O
the	O
data	O
permanently	O
into	O
the	O
cluster	O
.	O
</s>
<s>
A	O
number	O
of	O
third-party	O
file	B-Protocol
system	I-Protocol
bridges	O
have	O
also	O
been	O
written	O
,	O
none	O
of	O
which	O
are	O
currently	O
in	O
Hadoop	B-Application
distributions	O
.	O
</s>
<s>
However	O
,	O
some	O
commercial	O
distributions	O
of	O
Hadoop	B-Application
ship	O
with	O
an	O
alternative	O
file	B-Protocol
system	I-Protocol
as	O
the	O
defaultspecifically	O
IBM	O
and	O
MapR	B-General_Concept
.	O
</s>
<s>
In	O
2009	O
,	O
IBM	O
discussed	O
running	O
Hadoop	B-Application
over	O
the	O
IBM	B-Application
General	I-Application
Parallel	I-Application
File	I-Application
System	I-Application
.	O
</s>
<s>
In	O
April	O
2010	O
,	O
Parascale	O
published	O
the	O
source	O
code	O
to	O
run	O
Hadoop	B-Application
against	O
the	O
Parascale	O
file	B-Protocol
system	I-Protocol
.	O
</s>
<s>
In	O
April	O
2010	O
,	O
Appistry	O
released	O
a	O
Hadoop	B-Application
file	B-Protocol
system	I-Protocol
driver	O
for	O
use	O
with	O
its	O
own	O
CloudIQ	O
Storage	O
product	O
.	O
</s>
<s>
In	O
June	O
2010	O
,	O
HP	O
discussed	O
a	O
location-aware	O
IBRIX	B-General_Concept
Fusion	I-General_Concept
file	B-Protocol
system	I-Protocol
driver	O
.	O
</s>
<s>
In	O
May	O
2011	O
,	O
MapR	B-General_Concept
Technologies	I-General_Concept
Inc	I-General_Concept
.	I-General_Concept
announced	O
the	O
availability	O
of	O
an	O
alternative	O
file	B-Protocol
system	I-Protocol
for	O
Hadoop	B-Application
,	O
MapR	B-Application
FS	I-Application
,	O
which	O
replaced	O
the	O
HDFS	O
file	B-Protocol
system	I-Protocol
with	O
a	O
full	O
random-access	O
read/write	O
file	B-Protocol
system	I-Protocol
.	O
</s>
<s>
Atop	O
the	O
file	O
systems	O
comes	O
the	O
MapReduce	B-Operating_System
Engine	O
,	O
which	O
consists	O
of	O
one	O
JobTracker	O
,	O
to	O
which	O
client	O
applications	O
submit	O
MapReduce	B-Operating_System
jobs	O
.	O
</s>
<s>
With	O
a	O
rack-aware	O
file	B-Protocol
system	I-Protocol
,	O
the	O
JobTracker	O
knows	O
which	O
node	O
contains	O
the	O
data	O
,	O
and	O
which	O
other	O
machines	O
are	O
nearby	O
.	O
</s>
<s>
The	O
TaskTracker	O
on	O
each	O
node	O
spawns	O
a	O
separate	O
Java	B-Language
virtual	I-Language
machine	I-Language
(	O
JVM	B-Language
)	O
process	O
to	O
prevent	O
the	O
TaskTracker	O
itself	O
from	O
failing	O
if	O
the	O
running	O
job	O
crashes	O
its	O
JVM	B-Language
.	O
</s>
<s>
The	O
Job	O
Tracker	O
and	O
TaskTracker	O
status	O
and	O
information	O
is	O
exposed	O
by	O
Jetty	B-Language
and	O
can	O
be	O
viewed	O
from	O
a	O
web	O
browser	O
.	O
</s>
<s>
There	O
is	O
no	O
consideration	O
of	O
the	O
current	O
system	B-Operating_System
load	I-Operating_System
of	O
the	O
allocated	O
machine	O
,	O
and	O
hence	O
its	O
actual	O
availability	O
.	O
</s>
<s>
If	O
one	O
TaskTracker	O
is	O
very	O
slow	O
,	O
it	O
can	O
delay	O
the	O
entire	O
MapReduce	B-Operating_System
job	O
especially	O
towards	O
the	O
end	O
,	O
when	O
everything	O
can	O
end	O
up	O
waiting	O
for	O
the	O
slowest	O
task	O
.	O
</s>
<s>
By	O
default	O
Hadoop	B-Application
uses	O
FIFO	B-Operating_System
scheduling	O
,	O
and	O
optionally	O
5	O
scheduling	O
priorities	O
to	O
schedule	O
jobs	O
from	O
a	O
work	O
queue	O
.	O
</s>
<s>
The	O
fair	O
scheduler	O
was	O
developed	O
by	O
Facebook	B-Application
.	I-Application
</s>
<s>
Jobs	O
are	O
grouped	O
into	O
pools	B-General_Concept
.	O
</s>
<s>
Pools	B-General_Concept
have	O
to	O
specify	O
the	O
minimum	O
number	O
of	O
map	O
slots	O
,	O
reduce	O
slots	O
,	O
as	O
well	O
as	O
a	O
limit	O
on	O
the	O
number	O
of	O
running	O
jobs	O
.	O
</s>
<s>
The	O
capacity	O
scheduler	O
was	O
developed	O
by	O
Yahoo	B-Application
.	O
</s>
<s>
There	O
is	O
no	O
preemption	B-Operating_System
once	O
a	O
job	O
is	O
running	O
.	O
</s>
<s>
The	O
biggest	O
difference	O
between	O
Hadoop	B-Application
1	O
and	O
Hadoop	B-Application
2	O
is	O
the	O
addition	O
of	O
YARN	B-Application
(	O
Yet	O
Another	O
Resource	O
Negotiator	O
)	O
,	O
which	O
replaced	O
the	O
MapReduce	B-Operating_System
engine	O
in	O
the	O
first	O
version	O
of	O
Hadoop	B-Application
.	O
</s>
<s>
YARN	B-Application
strives	O
to	O
allocate	O
resources	O
to	O
various	O
applications	O
effectively	O
.	O
</s>
<s>
There	O
are	O
important	O
features	O
provided	O
by	O
Hadoop	B-Application
3	O
.	O
</s>
<s>
For	O
example	O
,	O
while	O
there	O
is	O
one	O
single	O
namenode	O
in	O
Hadoop	B-Application
2	O
,	O
Hadoop	B-Application
3	O
enables	O
having	O
multiple	O
name	O
nodes	O
,	O
which	O
solves	O
the	O
single	O
point	O
of	O
failure	O
problem	O
.	O
</s>
<s>
In	O
Hadoop	B-Application
3	O
,	O
there	O
are	O
containers	O
working	O
in	O
principle	O
of	O
Docker	B-Language
,	O
which	O
reduces	O
time	O
spent	O
on	O
application	O
development	O
.	O
</s>
<s>
One	O
of	O
the	O
biggest	O
changes	O
is	O
that	O
Hadoop	B-Application
3	O
decreases	O
storage	O
overhead	O
with	O
erasure	O
coding	O
.	O
</s>
<s>
Also	O
,	O
Hadoop	B-Application
3	O
permits	O
usage	O
of	O
GPU	O
hardware	O
within	O
the	O
cluster	O
,	O
which	O
is	O
a	O
very	O
substantial	O
benefit	O
to	O
execute	O
deep	O
learning	O
algorithms	O
on	O
a	O
Hadoop	B-Application
cluster	O
.	O
</s>
<s>
The	O
HDFS	O
is	O
not	O
restricted	O
to	O
MapReduce	B-Operating_System
jobs	O
.	O
</s>
<s>
The	O
list	O
includes	O
the	O
HBase	B-Language
database	O
,	O
the	O
Apache	B-Application
Mahout	I-Application
machine	O
learning	O
system	O
,	O
and	O
the	O
Apache	B-General_Concept
Hive	I-General_Concept
data	O
warehouse	O
.	O
</s>
<s>
Theoretically	O
,	O
Hadoop	B-Application
could	O
be	O
used	O
for	O
any	O
workload	O
that	O
is	O
batch-oriented	O
rather	O
than	O
real-time	O
,	O
is	O
very	O
data-intensive	B-Operating_System
,	O
and	O
benefits	O
from	O
parallel	B-Operating_System
processing	I-Operating_System
.	O
</s>
<s>
It	O
can	O
also	O
be	O
used	O
to	O
complement	O
a	O
real-time	O
system	O
,	O
such	O
as	O
lambda	B-General_Concept
architecture	I-General_Concept
,	O
Apache	B-General_Concept
Storm	I-General_Concept
,	O
Flink	B-Application
,	O
and	O
Spark	B-Language
Streaming	O
.	O
</s>
<s>
Commercial	O
applications	O
of	O
Hadoop	B-Application
include	O
:	O
</s>
<s>
On	O
19	O
February	O
2008	O
,	O
Yahoo	B-Application
!	I-Application
</s>
<s>
Inc	O
.	O
launched	O
what	O
they	O
claimed	O
was	O
the	O
world	O
's	O
largest	O
Hadoop	B-Application
production	O
application	O
.	O
</s>
<s>
The	O
Yahoo	B-Application
!	I-Application
</s>
<s>
Search	O
Webmap	O
is	O
a	O
Hadoop	B-Application
application	O
that	O
runs	O
on	O
a	O
Linux	B-Architecture
cluster	I-Architecture
with	O
more	O
than	O
10,000	O
cores	B-Architecture
and	O
produced	O
data	O
that	O
was	O
used	O
in	O
every	O
Yahoo	B-Application
!	I-Application
</s>
<s>
There	O
are	O
multiple	O
Hadoop	B-Application
clusters	O
at	O
Yahoo	B-Application
!	I-Application
</s>
<s>
and	O
no	O
HDFS	O
file	O
systems	O
or	O
MapReduce	B-Operating_System
jobs	O
are	O
split	O
across	O
multiple	O
data	O
centers	O
.	O
</s>
<s>
Every	O
Hadoop	B-Application
cluster	O
node	O
bootstraps	O
the	O
Linux	B-Application
image	O
,	O
including	O
the	O
Hadoop	B-Application
distribution	O
.	O
</s>
<s>
Work	O
that	O
the	O
clusters	O
perform	O
is	O
known	O
to	O
include	O
the	O
index	O
calculations	O
for	O
the	O
Yahoo	B-Application
!	I-Application
</s>
<s>
In	O
June	O
2009	O
,	O
Yahoo	B-Application
!	I-Application
</s>
<s>
made	O
the	O
source	O
code	O
of	O
its	O
Hadoop	B-Application
version	O
available	O
to	O
the	O
open-source	B-Application
community	O
.	O
</s>
<s>
In	O
2010	O
,	O
Facebook	B-Application
claimed	O
that	O
they	O
had	O
the	O
largest	O
Hadoop	B-Application
cluster	O
in	O
the	O
world	O
with	O
21	O
PB	O
of	O
storage	O
.	O
</s>
<s>
,	O
Hadoop	B-Application
adoption	O
had	O
become	O
widespread	O
:	O
more	O
than	O
half	O
of	O
the	O
Fortune	O
50	O
companies	O
used	O
Hadoop	B-Application
.	O
</s>
<s>
Hadoop	B-Application
can	O
be	O
deployed	O
in	O
a	O
traditional	O
onsite	O
datacenter	O
as	O
well	O
as	O
in	B-Architecture
the	I-Architecture
cloud	I-Architecture
.	O
</s>
<s>
The	B-Architecture
cloud	I-Architecture
allows	O
organizations	O
to	O
deploy	O
Hadoop	B-Application
without	O
the	O
need	O
to	O
acquire	O
hardware	O
or	O
specific	O
setup	O
expertise	O
.	O
</s>
<s>
A	O
number	O
of	O
companies	O
offer	O
commercial	O
implementations	O
or	O
support	O
for	O
Hadoop	B-Application
.	O
</s>
<s>
The	O
Apache	O
Software	O
Foundation	O
has	O
stated	O
that	O
only	O
software	O
officially	O
released	O
by	O
the	O
Apache	B-Application
Hadoop	I-Application
Project	O
can	O
be	O
called	O
Apache	B-Application
Hadoop	I-Application
or	O
Distributions	O
of	O
Apache	B-Application
Hadoop	I-Application
.	O
</s>
<s>
The	O
naming	O
of	O
products	O
and	O
derivative	O
works	O
from	O
other	O
vendors	O
and	O
the	O
term	O
"	O
compatible	O
"	O
are	O
somewhat	O
controversial	O
within	O
the	O
Hadoop	B-Application
developer	O
community	O
.	O
</s>
<s>
Some	O
papers	O
influenced	O
the	O
birth	O
and	O
growth	O
of	O
Hadoop	B-Application
and	O
big	B-Application
data	I-Application
processing	O
.	O
</s>
<s>
Jeffrey	O
Dean	O
,	O
Sanjay	O
Ghemawat	O
(	O
2004	O
)	O
,	O
Google	B-Application
.	I-Application
</s>
<s>
This	O
paper	O
inspired	O
Doug	O
Cutting	O
to	O
develop	O
an	O
open-source	B-Application
implementation	O
of	O
the	O
Map-Reduce	B-Operating_System
framework	O
.	O
</s>
<s>
He	O
named	O
it	O
Hadoop	B-Application
,	O
after	O
his	O
son	O
's	O
toy	O
elephant	O
.	O
</s>
<s>
(	O
2006	O
)	O
,	O
Google	B-Application
.	I-Application
</s>
