<s>
Dask	B-Operating_System
is	O
an	O
open-source	B-Application
Python	B-Language
library	O
for	O
parallel	B-Operating_System
computing	I-Operating_System
.	O
</s>
<s>
Dask	B-Operating_System
scales	O
Python	B-Language
code	I-Language
from	O
multi-core	O
local	O
machines	O
to	O
large	O
distributed	O
clusters	O
in	O
the	O
cloud	O
.	O
</s>
<s>
Dask	B-Operating_System
provides	O
a	O
familiar	O
user	B-Application
interface	I-Application
by	O
mirroring	O
the	O
APIs	B-General_Concept
of	O
other	O
libraries	O
in	O
the	O
PyData	O
ecosystem	O
including	O
:	O
Pandas	B-Application
,	O
scikit-learn	B-Application
and	O
NumPy	B-Application
.	O
</s>
<s>
It	O
also	O
exposes	O
low-level	O
APIs	B-General_Concept
that	O
help	O
programmers	O
run	O
custom	O
algorithms	O
in	O
parallel	O
.	O
</s>
<s>
Dask	B-Operating_System
was	O
created	O
by	O
Matthew	O
Rocklin	O
in	O
December	O
2014	O
and	O
has	O
over	O
9.8k	O
stars	O
and	O
500	O
contributors	O
on	O
GitHub	B-Application
.	O
</s>
<s>
Dask	B-Operating_System
is	O
used	O
by	O
retail	O
,	O
financial	O
,	O
governmental	O
organizations	O
,	O
as	O
well	O
as	O
life	O
science	O
and	O
geophysical	O
institutes	O
.	O
</s>
<s>
Walmart	O
,	O
Wayfair	O
,	O
JDA	O
,	O
GrubHub	O
,	O
General	O
Motors	O
,	O
NVIDIA	O
,	O
Harvard	O
Medical	O
School	O
,	O
Capital	O
One	O
and	O
NASA	O
are	O
among	O
the	O
organizations	O
that	O
use	O
Dask	B-Operating_System
.	O
</s>
<s>
Dask	B-Operating_System
has	O
two	O
parts	O
:	O
</s>
<s>
Dask	B-Operating_System
's	O
high-level	O
parallel	O
collections	O
–	O
DataFrames	O
,	O
Bags	O
,	O
and	O
Arrays	O
–	O
operate	O
in	O
parallel	O
on	O
datasets	O
that	O
may	O
not	O
fit	O
into	O
memory	O
.	O
</s>
<s>
Dask	B-Operating_System
’s	O
task	O
scheduler	O
executes	O
task	O
graphs	O
in	O
parallel	O
.	O
</s>
<s>
Dask	B-Operating_System
supports	O
several	O
user	B-Application
interfaces	I-Application
called	O
high-level	O
and	O
low-level	O
collections	O
:	O
</s>
<s>
Under	O
the	O
hood	O
,	O
each	O
of	O
these	O
user	B-Application
interfaces	I-Application
adopts	O
the	O
same	O
parallel	B-Operating_System
computing	I-Operating_System
machinery	O
.	O
</s>
<s>
Dask	B-Operating_System
's	O
high-level	O
collections	O
are	O
the	O
natural	O
entry	O
point	O
for	O
users	O
who	O
are	O
interested	O
in	O
scaling	O
up	O
their	O
pandas	B-Application
,	O
NumPy	B-Application
or	O
scikit-learn	B-Application
workload	O
.	O
</s>
<s>
Dask	B-Operating_System
’s	O
DataFrame	O
,	O
Array	O
and	O
Dask-ML	O
are	O
alternatives	O
to	O
Pandas	B-Application
DataFrame	O
,	O
Numpy	B-Application
Array	O
and	O
scikit-learn	B-Application
respectively	O
with	O
slight	O
variations	O
to	O
the	O
original	O
interfaces	O
.	O
</s>
<s>
Dask	B-Operating_System
Array	O
is	O
a	O
high-level	O
collection	O
that	O
parallelizes	O
array-based	O
workloads	O
and	O
maintains	O
the	O
familiar	O
NumPy	B-Application
API	B-General_Concept
,	O
such	O
as	O
slicing	O
,	O
arithmetic	O
,	O
reductions	O
,	O
mathematics	O
,	O
etc.	O
,	O
making	O
it	O
easy	O
for	O
Numpy	B-Application
users	O
to	O
scale	O
up	O
array	O
operations	O
.	O
</s>
<s>
A	O
Dask	B-Operating_System
array	O
comprises	O
many	O
smaller	O
n-dimensional	O
Numpy	B-Application
arrays	O
and	O
uses	O
a	O
blocked	O
algorithm	O
to	O
enable	O
computation	O
on	O
larger-than-memory	O
arrays	O
.	O
</s>
<s>
During	O
an	O
operation	O
,	O
Dask	B-Operating_System
translates	O
the	O
array	O
operation	O
into	O
a	O
task	O
graph	O
,	O
breaks	O
up	O
large	O
Numpy	B-Application
arrays	O
into	O
multiple	O
smaller	O
chunks	O
,	O
and	O
executes	O
the	O
work	O
on	O
each	O
chunk	O
in	O
parallel	O
.	O
</s>
<s>
Dask	B-Operating_System
DataFrame	O
is	O
a	O
high-level	O
collection	O
that	O
parallelizes	O
DataFrame	O
based	O
workloads	O
.	O
</s>
<s>
A	O
Dask	B-Operating_System
DataFrame	O
comprises	O
many	O
smaller	O
Pandas	B-Application
DataFrames	O
partitioned	O
along	O
the	O
index	O
.	O
</s>
<s>
It	O
maintains	O
the	O
familiar	O
Pandas	B-Application
API	B-General_Concept
,	O
making	O
it	O
easy	O
for	O
Pandas	B-Application
users	O
to	O
scale	O
up	O
DataFrame	O
workloads	O
.	O
</s>
<s>
During	O
a	O
DataFrame	O
operation	O
,	O
Dask	B-Operating_System
creates	O
a	O
task	O
graph	O
and	O
triggers	O
operations	O
on	O
the	O
constituent	O
DataFrames	O
in	O
a	O
manner	O
that	O
reduces	O
memory	O
footprint	O
and	O
increases	O
parallelism	B-Operating_System
through	O
sharing	O
and	O
deleting	O
of	O
intermediate	O
results	O
.	O
</s>
<s>
Dask	B-Operating_System
Bag	O
is	O
an	O
unordered	O
collection	O
of	O
repeated	O
objects	O
,	O
a	O
hybrid	O
between	O
a	O
set	O
and	O
a	O
list	O
.	O
</s>
<s>
Dask	B-Operating_System
Bag	O
is	O
used	O
to	O
parallelize	O
computation	O
of	O
semi-structured	O
or	O
unstructured	O
data	O
,	O
such	O
as	O
JSON	O
records	O
,	O
text	O
data	O
,	O
log	O
files	O
or	O
user-defined	O
Python	B-Language
objects	O
using	O
operations	O
such	O
as	O
filter	O
,	O
fold	O
,	O
map	O
and	O
groupby	O
.	O
</s>
<s>
Dask	B-Operating_System
Bags	O
can	O
be	O
created	O
from	O
an	O
existing	O
Python	B-Language
iterable	O
or	O
can	O
load	O
data	O
directly	O
from	O
text	O
files	O
and	O
binary	O
files	O
in	O
the	O
Avro	O
format	O
.	O
</s>
<s>
The	O
Dask	B-Operating_System
low-level	O
interface	O
allows	O
for	O
more	O
customization	O
.	O
</s>
<s>
It	O
is	O
suitable	O
for	O
data	O
that	O
does	O
not	O
fall	O
within	O
the	O
scope	O
of	O
a	O
Dask	B-Operating_System
DataFrame	O
,	O
Bag	O
or	O
Array	O
.	O
</s>
<s>
Dask	B-Operating_System
has	O
the	O
following	O
low-level	O
collections	O
:	O
</s>
<s>
Dask	B-Operating_System
delayed	O
is	O
an	O
interface	O
used	O
to	O
parallelize	O
generic	O
Python	B-Language
code	I-Language
that	O
does	O
not	O
fit	O
into	O
high	O
level	O
collections	O
like	O
Dask	B-Operating_System
Array	O
or	O
Dask	B-Operating_System
DataFrame	O
.	O
</s>
<s>
Python	B-Language
functions	O
decorated	O
with	O
Dask	B-Operating_System
delayed	O
adopt	O
a	O
lazy	O
evaluation	O
strategy	O
by	O
deferring	O
execution	O
and	O
generating	O
a	O
task	O
graph	O
with	O
the	O
function	O
and	O
its	O
arguments	O
.	O
</s>
<s>
The	O
Python	B-Language
function	O
will	O
only	O
execute	O
when	O
.compute	O
is	O
invoked	O
.	O
</s>
<s>
Dask	B-Operating_System
delayed	O
can	O
be	O
used	O
as	O
a	O
function	O
dask.delayed	O
or	O
as	O
a	O
decorator	O
@dask	O
.delayed	O
.	O
</s>
<s>
Dask	B-Operating_System
Futures	O
,	O
an	O
immediate	O
(	O
non-lazy	O
)	O
alternative	O
to	O
Dask	B-Operating_System
delayed	O
,	O
provides	O
a	O
real-time	O
task	O
framework	O
that	O
extends	O
Python	B-Language
’s	O
concurrent.futures	O
interface	O
,	O
which	O
provides	O
a	O
high-level	O
interface	O
for	O
asynchronous	O
execution	O
of	O
callables	O
.	O
</s>
<s>
For	O
example	O
,	O
users	O
can	O
run	O
Dask	B-Operating_System
array/bag/dataframe	O
to	O
load	O
and	O
pre-process	O
data	O
,	O
then	O
switch	O
to	O
Dask	B-Operating_System
delayed	O
for	O
a	O
custom	O
algorithm	O
that	O
is	O
specific	O
to	O
their	O
domain	O
,	O
then	O
switch	O
back	O
to	O
Dask	B-Operating_System
array/dataframe	O
to	O
clean	O
up	O
and	O
store	O
results	O
.	O
</s>
<s>
Dask	B-Operating_System
’s	O
high	O
and	O
low-level	O
collections	O
create	O
a	O
directed	O
acyclic	O
graph	O
of	O
tasks	O
,	O
which	O
represents	O
the	O
relationship	O
between	O
computation	O
tasks	O
.	O
</s>
<s>
A	O
node	O
in	O
a	O
task	O
graph	O
represents	O
a	O
Python	B-Language
function	O
that	O
performs	O
a	O
unit	O
of	O
computation	O
and	O
an	O
edge	O
represents	O
the	O
data	O
dependency	O
between	O
the	O
upstream	O
and	O
downstream	O
task	O
.	O
</s>
<s>
After	O
a	O
task	O
graph	O
is	O
generated	O
,	O
the	O
task	O
scheduler	O
manages	O
the	O
workflow	O
based	O
on	O
the	O
given	O
task	O
graph	O
by	O
assigning	O
tasks	O
to	O
workers	O
in	O
a	O
manner	O
that	O
improves	O
parallelism	B-Operating_System
and	O
respects	O
the	O
data	O
dependencies	O
.	O
</s>
<s>
Dask	B-Operating_System
provides	O
two	O
families	O
of	O
schedulers	O
:	O
single-machine	O
scheduler	O
and	O
distributed	O
scheduler	O
.	O
</s>
<s>
Local	O
threadsA	O
threaded	O
scheduler	O
leverages	O
Python	B-Language
’s	O
concurrent.futures.ThreadPoolExecuter	O
to	O
execute	O
computations	O
.	O
</s>
<s>
Due	O
to	O
Python	B-Language
’s	O
Global	O
Interpreter	O
Lock	O
,	O
local	O
threads	O
provide	O
parallelism	B-Operating_System
only	O
when	O
the	O
computation	O
is	O
primarily	O
non-Python	O
code	O
,	O
which	O
is	O
the	O
case	O
for	O
Pandas	B-Application
DataFrame	O
,	O
Numpy	B-Application
arrays	O
or	O
other	O
Python/C/C	O
++	O
based	O
projects	O
.	O
</s>
<s>
Local	O
processA	O
multiprocessing	O
scheduler	O
leverages	O
Python	B-Language
’s	O
concurrent.futures.ProcessPoolExecutor	O
to	O
execute	O
computations	O
.	O
</s>
<s>
This	O
allows	O
bypassing	O
of	O
issues	O
with	O
Python	B-Language
’s	O
Global	O
Interpretable	O
Lock	O
and	O
provides	O
parallelism	B-Operating_System
for	O
computation	O
tasks	O
with	O
primarily	O
Python	B-Language
code	I-Language
.	O
</s>
<s>
Single	O
threadA	O
single	O
threaded	O
scheduler	O
executes	O
computation	O
with	O
no	O
parallelism	B-Operating_System
.	O
</s>
<s>
Dask	B-Operating_System
’s	O
distributed	O
scheduler	O
can	O
be	O
set	O
up	O
on	O
a	O
local	O
machine	O
or	O
scale	O
out	O
on	O
a	O
cluster	O
.	O
</s>
<s>
Dask	B-Operating_System
can	O
work	O
with	O
resource	O
managers	O
,	O
such	O
as	O
Hadoop	B-Application
YARN	I-Application
,	O
Kubernetes	B-General_Concept
,	O
or	O
PBS	O
,	O
Slurm	B-Operating_System
,	O
SGD	O
and	O
LSF	O
for	O
High	O
Performance	O
Computing	O
(	O
HPC	O
)	O
clusters	O
.	O
</s>
<s>
Dask-ML	O
is	O
compatible	O
with	O
scikit-learn	B-Application
’	O
s	O
estimator	O
API	B-General_Concept
of	O
fit	O
,	O
transform	O
and	O
predict	O
and	O
is	O
well	O
integrated	O
with	O
machine	O
learning	O
and	O
deep	O
learning	O
frameworks	O
such	O
XGBoost	B-Language
,	O
LightGBM	O
,	O
PyTorch	B-Algorithm
,	O
Keras	B-Algorithm
and	O
TensorFlow	B-Language
through	O
scikit-learn	B-Application
compatible	O
wrappers	O
.	O
</s>
<s>
Selected	O
scikit-learn	B-Application
estimators	O
and	O
utilities	O
can	O
be	O
parallelized	B-Operating_System
through	O
executing	O
jobs	O
across	O
multiple	O
CPU	O
cores	O
using	O
the	O
Joblib	O
library	O
.	O
</s>
<s>
Dask	B-Operating_System
offers	O
an	O
alternative	O
Joblib	O
backend	O
which	O
is	O
useful	O
for	O
scaling	O
of	O
Joblib-backed	O
scikit-learn	B-Application
algorithms	O
out	O
to	O
a	O
cluster	O
of	O
machines	O
for	O
compute	O
constrained	O
workloads	O
.	O
</s>
<s>
For	O
memory	O
constrained	O
workloads	O
,	O
Dask	B-Operating_System
offers	O
alternatives	O
,	O
such	O
as	O
Parallel	O
Meta-estimators	O
for	O
parallelizing	O
and	O
scaling	O
out	O
tasks	O
that	O
are	O
not	O
parallelized	B-Operating_System
within	O
scikit-learn	B-Application
and	O
Incremental	O
Hyperparameter	O
Optimization	O
for	O
scaling	O
hyper-parameter	O
search	O
and	O
parallelized	B-Operating_System
estimators	O
.	O
</s>
<s>
XGBoost	B-Language
and	O
LightGBM	O
are	O
popular	O
algorithms	O
that	O
are	O
based	O
on	O
Gradient	B-Algorithm
Boosting	I-Algorithm
and	O
both	O
are	O
integrated	O
with	O
Dask	B-Operating_System
for	O
distributed	O
learning	O
.	O
</s>
<s>
Dask	B-Operating_System
does	O
not	O
power	O
XGBoost	B-Language
or	O
LightGBM	O
,	O
rather	O
it	O
facilitates	O
setting	O
up	O
of	O
the	O
cluster	O
,	O
scheduler	O
,	O
and	O
workers	O
required	O
then	O
hands	O
off	O
the	O
data	O
to	O
the	O
machine	O
learning	O
framework	O
to	O
perform	O
distributed	O
training	O
.	O
</s>
<s>
Training	O
an	O
XGBoost	B-Language
model	O
with	O
Dask	B-Operating_System
,	O
a	O
Dask	B-Operating_System
cluster	O
is	O
composed	O
of	O
a	O
central	O
scheduler	O
and	O
multiple	O
distributed	O
workers	O
,	O
is	O
accomplished	O
by	O
spinning	O
up	O
an	O
XGBoost	B-Language
scheduler	O
in	O
the	O
same	O
process	O
running	O
the	O
Dask	B-Operating_System
central	O
scheduler	O
and	O
XGBoost	B-Language
worker	O
in	O
the	O
same	O
process	O
running	O
the	O
Dask	B-Operating_System
workers	O
.	O
</s>
<s>
Dask	B-Operating_System
workers	O
then	O
hand	O
over	O
the	O
Pandas	B-Application
DataFrame	O
to	O
the	O
local	O
XGBoost	B-Language
worker	O
for	O
distributed	O
training	O
.	O
</s>
<s>
Skorch	O
is	O
a	O
scikit-learn	B-Application
compatible	O
wrapper	O
for	O
PyTorch	B-Algorithm
,	O
which	O
enables	O
Dask-ML	O
to	O
be	O
used	O
together	O
with	O
PyTorch	B-Algorithm
.	O
</s>
<s>
SciKeras	O
is	O
an	O
scikit-learn	B-Application
compatible	O
wrapper	O
for	O
Keras	B-Algorithm
models	O
which	O
enables	O
Dask-ML	O
to	O
be	O
used	O
with	O
Keras	B-Algorithm
.	O
</s>
<s>
Walmart	O
uses	O
Dask	B-Operating_System
for	O
forecasting	O
the	O
demand	O
for	O
500,000,000	O
store-item	O
combinations	O
.	O
</s>
<s>
Using	O
RAPIDS	O
and	O
XGBoost	B-Language
,	O
supported	O
by	O
Dask	B-Operating_System
,	O
they	O
have	O
reached	O
100x	O
acceleration	O
.	O
</s>
<s>
Blue	O
Yonder	O
uses	O
Dask	B-Operating_System
to	O
process	O
terabytes	O
of	O
data	O
on	O
a	O
daily	O
basis	O
.	O
</s>
<s>
They	O
can	O
write	O
Pandas-like	O
code	O
in	O
Dask	B-Operating_System
,	O
which	O
can	O
then	O
be	O
pushed	O
directly	O
to	O
production	O
.	O
</s>
<s>
Grubhub	O
uses	O
Dask	B-Operating_System
alongside	O
TensorFlow	B-Language
for	O
pre-processing	O
and	O
ETL	O
.	O
</s>
<s>
Dask	B-Operating_System
allows	O
them	O
to	O
continue	O
working	O
in	O
Python	B-Language
and	O
get	O
the	O
functionalities	O
needed	O
.	O
</s>
<s>
Dask	B-Operating_System
is	O
used	O
for	O
high	O
resolution	O
,	O
4-dimensional	O
,	O
cellular	O
imagery	O
by	O
Harvard	O
Medical	O
School	O
,	O
Howard	O
Hughes	O
Medical	O
Institute	O
,	O
Chan	O
Zuckerberg	O
Initiative	O
,	O
and	O
the	O
UC	O
Berkeley	O
Advanced	O
Bioimaging	O
Center	O
.	O
</s>
<s>
Dask	B-Operating_System
helps	O
them	O
scale	O
their	O
data	O
analysis	O
workflows	O
with	O
its	O
API	B-General_Concept
that	O
resembles	O
NumPy	B-Application
,	O
Pandas	B-Application
,	O
and	O
scikit-learn	B-Application
code	O
.	O
</s>
<s>
Dask	B-Operating_System
is	O
also	O
used	O
at	O
the	O
Novartis	O
Institute	O
for	O
Biomedical	O
Research	O
to	O
scale	O
machine	O
learning	O
prototypes	O
.	O
</s>
<s>
Dask	B-Operating_System
is	O
used	O
in	O
Climate	O
Science	O
,	O
Energy	O
,	O
Hydrology	O
,	O
Meteorology	O
,	O
and	O
Satellite	O
Imaging	O
by	O
companies	O
such	O
as	O
NASA	O
,	O
LANL	O
,	O
PANGEO	O
:	O
Earth	O
Science	O
and	O
the	O
UK	O
Meteorology	O
Office	O
.	O
</s>
<s>
Dask	B-Operating_System
is	O
integrated	O
into	O
many	O
libraries	O
,	O
such	O
as	O
Pangeo	O
and	O
xarray	O
;	O
time	O
series	O
software	O
like	O
Prophet	O
and	O
tsfresh	O
;	O
ETL/ML	O
software	O
like	O
scikit-learn	B-Application
,	O
RAPIDS	O
,	O
and	O
XGBoost	B-Language
;	O
workflow	O
management	O
tools	O
like	O
Apache	B-Application
Airflow	I-Application
and	O
Prefect	O
.	O
</s>
<s>
Dask	B-Operating_System
was	O
originally	O
developed	O
at	O
Continuum	O
Analytics	O
,	O
a	O
for-profit	O
Python	B-Language
consulting	O
company	O
that	O
eventually	O
became	O
Anaconda	B-Application
,	O
Inc.	O
,	O
the	O
creator	O
of	O
many	O
open-source	B-Application
packages	O
and	O
the	O
Anaconda	B-Application
Python	B-Language
distribution	O
.	O
</s>
<s>
Dask	B-Operating_System
grew	O
out	O
of	O
the	O
Blaze	O
project	O
,	O
a	O
DARPA	O
funded	O
project	O
to	O
accelerate	O
computation	O
in	O
open	O
source	O
.	O
</s>
<s>
Blaze	O
was	O
an	O
ambitious	O
project	O
that	O
tried	O
to	O
redefine	O
computation	O
,	O
storage	O
,	O
compression	O
,	O
and	O
data	O
science	O
APIs	B-General_Concept
for	O
Python	B-Language
,	O
led	O
originally	O
by	O
Travis	O
Oliphant	O
and	O
Peter	O
Wang	O
,	O
the	O
co-founders	O
of	O
Anaconda	B-Application
.	O
</s>
<s>
Instead	O
of	O
rewriting	O
a	O
software	O
ecosystem	O
,	O
Dask	B-Operating_System
’s	O
team	O
intended	O
to	O
augment	O
the	O
existing	O
one	O
with	O
the	O
right	O
component	O
.	O
</s>
<s>
With	O
this	O
idea	O
in	O
mind	O
,	O
on	O
December	O
21	O
,	O
2014	O
Matthew	O
Rocklin	O
created	O
Dask	B-Operating_System
.	O
</s>
<s>
The	O
purpose	O
of	O
Dask	B-Operating_System
was	O
originally	O
to	O
parallelize	O
NumPy	B-Application
so	O
that	O
it	O
could	O
use	O
one	O
full	O
workstation	O
computer	O
,	O
which	O
was	O
common	O
in	O
finance	O
shops	O
at	O
the	O
time	O
.	O
</s>
<s>
The	O
first	O
projects	O
to	O
really	O
adopt	O
Dask	B-Operating_System
were	O
Xarray	O
(	O
commonly	O
used	O
in	O
geo	O
sciences	O
)	O
and	O
Scikit-Image	O
(	O
commonly	O
used	O
in	O
image	O
processing	O
)	O
.	O
</s>
<s>
Dask	B-Operating_System
was	O
integrated	O
into	O
Xarray	O
within	O
a	O
few	O
months	O
of	O
being	O
created	O
.	O
</s>
<s>
It	O
provided	O
Dask	B-Operating_System
with	O
its	O
first	O
user	O
community	O
,	O
which	O
remains	O
to	O
this	O
day	O
.	O
</s>
<s>
Understanding	O
that	O
there	O
is	O
demand	O
for	O
a	O
lightweight	O
parallelism	B-Operating_System
solution	O
for	O
Pandas	B-Application
DataFrames	O
and	O
machine	O
learning	O
tools	O
,	O
such	O
as	O
scikit-learn	B-Application
,	O
Dask	B-Operating_System
quickly	O
evolved	O
to	O
support	O
other	O
projects	O
as	O
well	O
.	O
</s>
<s>
Since	O
2018	O
,	O
other	O
teams	O
and	O
institutions	O
in	O
academia	O
,	O
tech	O
companies	O
,	O
and	O
large	O
corporations	O
such	O
as	O
NASA	O
,	O
UK	O
Met	O
Office	O
,	O
Blue	O
Yonder	O
and	O
NVIDIA	O
,	O
became	O
interested	O
in	O
Dask	B-Operating_System
and	O
began	O
integrating	O
it	O
into	O
their	O
systems	O
.	O
</s>
<s>
Dask	B-Operating_System
received	O
support	O
from	O
a	O
diverse	O
set	O
of	O
sources	O
:	O
the	O
US	O
Government	O
(	O
DARPA	O
grant	O
)	O
,	O
the	O
Gordon	O
and	O
Betty	O
Moore	O
Foundation	O
,	O
Anaconda	B-Application
,	O
NSF	O
,	O
NASA	O
(	O
US	O
research	O
grants	O
with	O
collaborations	O
like	O
Pangeo	O
)	O
and	O
Nvidia	O
.	O
</s>
<s>
In	O
2020	O
,	O
Matthew	O
Rocklin	O
founded	O
Coiled	O
Computing	O
,	O
Inc	O
.	O
to	O
provide	O
further	O
support	O
for	O
Dask	B-Operating_System
development	O
and	O
enable	O
companies	O
to	O
deploy	O
Dask	B-Operating_System
clusters	O
in	O
the	O
cloud	O
.	O
</s>
