Accep ed o publica ion in he P oceedings o he IEEE 18 h In e na ional Con e ence on Cloud Compu ing (CLOUD), July 2025
Se e less Da a Analy ics (Finally) B idging he
Gap: In oducing he ORTZI DATAFRAME
Ge m´
an T. Eizagui e∗, Ma c Hos au†, Ma c S´
anchez-A igas
Depa men o Compu e Enginee ing and Ma hema ics
Uni e si a Ro i a i Vi gili, Ta agona
{ge man elmo.eizagui e,ma c.hos au,ma c.sanchez}@u .ca
Abs ac —Se e less echnologies ha e simpli ied dis ibu ed
compu ing by s eamlining esou ce managemen and o e ing
ou -o - he-box usabili y o cloud esou ces. Howe e , se e less
compu ing has ye o ully pe mea e he b oade da a analy ics
communi y. One o he majo easons causing his slow adop ion
is he lack o a handy se e less in e ace o seamlessly unning
ecu en wo kloads in he cloud. To ill his gap, we in oduce
in his wo k he missing piece in se e less analy ics: he ORTZI
DATAFRAME, a p ac ical and in ui i e p og amming abs ac ion
ha mi o s pandas Da aF ames, so ha use s can e o lessly
un hei local, single- h eaded Py hon code a scale in he cloud.
Needless o say, such a powe ul abs ac ion is ce ainly useless i
no backed by a se e less analy ics sys em ha can ope a e o e
i in pa allel. Fo his eason, ano he majo con ibu ion o his
pape is a ully- ledged sys em ha can un jobs in pa allel ac oss
he cloud con inuum using he no el ORTZI DATAFRAMES. The
new sys em le e ages he speci ic capabili ies o each se e less
backend wi hou use in e en ion.
Ou e alua ion demons a es ha ORTZI enables explo a ion
o he nuanced ade-o s o he e ogeneous backends wi h min-
imal p og amming changes and o e head. By ha nessing he
seamless na u e o ORTZI, we op imize jobs h ough s a egic
backend selec ion, s ill deli e ing a use - iendly open sou ce
amewo k o p og amme s wi hou cloud expe ise.
Index Te ms—Dis ibu ed compu ing, se e less compu ing,
cloud compu ing, da a analy ics, p og amming models
I. INTRODUCTION
Cloud compu ing o e s i ually unlimi ed compu ing ca-
pabili ies on demand, smoo hly scaling esou ces o mee use
equi emen s. Beyond b oad scalabili y, he cloud p o ides a
ich ecosys em o specialized se ices ailo ed o di e se appli-
ca ion needs. Howe e , ope a ional asks and he s eep lea ning
cu e equi ed o mas e cloud en i onmen s equen ly de e
gene al p og amme s om ully le e aging cloud esou ces.
To add ess cloud usabili y challenges, se e less compu ing
simpli ies he in e ac ion be ween p og amme s and in as-
© 2024 IEEE. Pe sonal use o his ma e ial is pe mi ed. Pe mission om
IEEE mus be ob ained o all o he uses, in any cu en o u u e media,
including ep in ing/ epublishing his ma e ial o ad e ising o p omo ional
pu poses, c ea ing new collec i e wo ks, o esale o edis ibu ion o se e s
o lis s, o euse o any copy igh ed componen o his wo k in o he wo ks.
We app ecia e all anonymous e iewe s a CLOUD ’25, who p o ided
insigh ul eedback. This wo k has been pa ly unded by he EU Ho izon
p og amme unde he g an ag eemen s: 101092644 (Nea Da a), 101092646
(CloudSkin), 101093110 (EXTRACT) and 101086248 (CLOUDSTARS).
Ge m´
an T. Eizagui e is ecipien o a p e-doc o al FPU g an om he Spanish
Minis y o Uni e si ies ( e . FPU21/00630). Ma c S´
anchez-A igas is a Se a
H´
un e Fellow.
uc u e by abs ac ing esou ce managemen away. Se e less
echnologies, such as cloud unc ions and objec s o age,
p o ide apid scalabili y, simpli ied esou ce p o isioning, and
as deploymen imes h ough simple APIs. These ea u es
educe he complexi y o cloud de elopmen , and enable o
eques p ecisely he esou ces needed, exac ly when needed.
The long p omise o se e less ex ends o many domains,
and pa icula ly o da a analy ics [1]. Hos ing da a analy ics
jobs in he cloud can o en be ime-consuming, and se e less
compu ing alle ia es his hu dle. Ideally, an analys wi hou
cloud expe ise should be able o launch da a analy ics jobs in
he cloud s aigh away, wi hou wo ying abou he unde lying
se up. Howe e , he goal o ully se e less da a analy ics
is in eali y a om being achie ed, p ima ily due o h ee
un esol ed challenges. These a e:
1. No gene al abs ac ion o se e less da a analy ics. The
s ong syne gy be ween se e less compu ing and da a analy -
ics has led o ui ul de elopmen o p o o ypes. Howe e , o
he bes o ou knowledge, he e is no a simple and gene al-
pu pose in e ace o un se e less da a analy ics e o lessly
in he cloud. In sho , he exis ing open sou ce amewo ks
can be g ouped in o h ee classes:
•P o o ypes wi h no gene al p og amming abs ac ions:
Jol eon [2], Di o [3] o Min low [4];
•Tools wi h usable da a abs ac ions, bu a limi ed se o
ope a o s: PyW en [5], Li hops [6], and See [7];
•Tools wi h usable p og amming abs ac ions, bu ailo ed
o a speci ic sys em/backend equi ing manual se up:
Pixels-Tu bo [8], Wukong [9], Flin [10] o Dex e [11].
2. The lack o open-sou ce se e less analy ics sys ems. As
o oday, mos o he s a e-o - he-a se e less da a analy ics
sys ems ha e no been open-sou ced. Un o una ely, he lis is
a he long, bu some o he mos ele an ones a e Locus [12],
Lambada [13], SONIC [14], Cae us [15], and Sma pick [16].
Code inaccessibili y o ces o he esea che s o independen ly
ec ea e hese ideas based on in-pape e e ences. Fo example,
Jol eon ( e)implemen s Cae us’ scheduling 1. Al hough such
ein e p e a ions a e ine i able, hey hinde he es ablishmen
o gene al s anda ds o se e less da a analy ics sys ems.
3. The cloud con inuum is nowadays a de ac o s anda d.
He e ogeneous compu e backend pools a e no an excep ion
1h ps://gi hub.com/pkusys/Jol eon/blob/wo k low/schedule .py
bu a ule. Mode n analy ics sys ems combine on-p emise clus-
e s, cloud esou ces and e en high-pe o mance lap ops. Ye ,
mos se e less amewo ks [2]–[7], [9], [11], [13] o e look
his exis ing compu ing powe , op ing ins ead o o load en i e
wo kloads o homogeneous se e less se ices.
We a gue ha se e less amewo ks should no be con ined
o pu e se e less backends. Ins ead, hey should deli e a
se e less expe ience while le e aging all a ailable esou ces,
enabling cos and pe o mance op imiza ions. Fo example,
amewo ks could p io i ize al eady p o isioned esou ces
while exploi ing cloud unc ions when necessa y.
O loading wo kloads o cloud unc ions has been p e i-
ously p oposed, o example, o o e lap pa ial wo kloads wi h
VM deploymen imes [8] o o handle sudden load spikes [17].
We belie e hese app oaches should be ex ended o gene al
da a analy ics p og amming, g an ing use s on-demand and
managed access o he lexibili y o cloud unc ions.
Ye , his holis ic ision o le e aging he cloud con inuum
emains la gely un ealized in p ac ical e ms. Many con em-
po a y se e less amewo ks s ill na owly ocus on speci ic
cloud se ices, while e en es ablished da a p ocessing sys ems
like Spa k [18], Ray [19], o Dask [20] ypically lack ou -
o - he-box con inuum suppo and demand complex manual
con igu a ion e en o basic homogeneous clus e se ups. This
highligh s a signi ican gap in p o iding uly seamless access
o oday’s di e se compu ing en i onmen s.
Ou con ibu ion
We in oduce ORTZI, a se e less da a analy ics amewo k
designed o he e ogeneous compu ing en i onmen s. Wi h i s
amilia pandas-like DATAFRAME abs ac ion, ORTZI enables
de elope s o i ially w i e cus om da a analy ics jobs, which
a e au oma ically pa allelized ac oss he a ailable compu ing
backends. Mo eo e , ORTZI o ches a es end- o-end wo kload
execu ion, le e aging he unique capabili ies o each compu e
backend o op imize job pe o mance and da a exchanges. The
key con ibu ions o his wo k a e as ollows:
•An open-sou ce, gene al-pu pose p og amming ab-
s ac ion o se e less da a analy ics, ea u ing a
comp ehensi e se o buil -in p imi i es, anging om
simple map o complex so and g oupBy ope a o s.
•An adap i e se e less a chi ec u e o conduc seamless
job pa alleliza ion ac oss he cloud con inuum, dynami-
cally adjus ing o he a chi ec u al ai s o each backend.
•Amanaged da a exchange sys em o complex da a
analy ics, ocusing on decoupled and he e ogeneous
compu ing en i onmen s.
ORTZI deli e s pe o mance compa able o s a e-o - he-a
se e less amewo ks in a pu e cloud unc ions deploymen ,
wi h only a 6.78% o e head. By anspa en ly le e aging he -
e ogeneous backends, ORTZI deli e s a 25.57% pe o mance
imp o emen o e Jol eon [2] a compa able cos and educes
la ency by 70.59% compa ed o EMR Se e less [21].
II. ORTZI OVERVIEW
ORTZI is a p og amming pla o m o da a analy ics ha
au oma ically p o isions and le e ages compu ing esou ces
wi h minimal c eden ial se up. This mi o s he expe ience o
comme cial, se e less da a analy ics p oduc s like Coiled [22]
o Anyscale [23], bu as a ully open-sou ce solu ion wi h
seamless compa ibili y ac oss he cloud con inuum.
The ORTZI DATAFRAME.ORTZI is buil on he ORTZI
DATAFRAME, a p ac ical abs ac ion wi h a Dask [20]- and
pandas [24]-like API. This ensu es amilia i y and ease o
adop ion, o e ing an expe ience simila o local p og amming.
Lis ing 1shows an example o da a analy ics p og amming
wi h he ORTZI DATAFRAME.
In e nally, a ORTZI DATAFRAME is di ided ow-wise in o
pa i ions. Each pa i ion is a sepa a e ins ance o a Pola s [25]
Da aF ame. ORTZI au oma ically dis ibu es he pa i ions
ac oss he unde lying compu e backend o enable he pa allel
p ocessing o an ORTZI DATAFRAME. Also, ORTZI manages
he pe sis ence o pa i ions in a anspa en way o he use ,
and s o es a pa i ion in memo y, on disk, and in an ex e nal
s o age sys em, depending on he compu ing needs.
In e ms o da a mu abili y, ORTZI DATAFRAMES suppo
wo main ypes o ans o ma ions: (a) na ow ans o ma ions;
and (b) wide ans o ma ions. Na ow ans o ma ions equi e
only one- o-one dependencies be ween pa i ions and include
p imi i es o apply a unc ion o he da a om all he pa i ions
in pa allel, such as map and il e . The pa alleliza ion o
such ope a ions is s aigh o wa d due o he absence o c oss-
dependencies.
A a mo e complex aspec a e ORTZI wide ans o ma ions,
as hey equi e shu ling pa i ions [7]. Wide ans o ma ions,
such as so and g oupBy, in ol e all- o-all communica ion
pa e ns. ORTZI abs ac s he complexi ies o hese pa e ns
om he use by au oma ically handling he pa allel exchange
o pa i ions be ween he compu e nodes.
1 om o zi impo Da aF ame
2 om o zi.backends impo BackendType
3 om some_lib impo enc yp _ unc, some_ unc
4 om numpy impo mean
5
6sd = Da aF ame. om_cs (
7"s3://my-bucke /da a.cs ",
8on_p emise=T ue
9)
10 sd = sd .so (by="col1")
11 sd = sd .map(enc yp _ unc)
12 sd = sd .se _scaleup(BackendType.CLOUD_FUNCTION)
13 sd = sd .map(some_ unc)
14 esul = sd . educe(mean)
Lis ing 1: Se e less da a analy ics in he cloud con inuum
using ORTZI abs ac ions.
To ma e ialize esul s and p esen esul s o he use , ORTZI
implemen s a se o ac ions, such as collec and w i e.
Ac ions de ine compu a ion bounda ies. In o he wo ds, when
ORTZI inds an ac ion, i uns all he pending ans o ma ions
o ha poin . This is isible in line 15 o Lis ing 1, whe e
he call o he educe ac ion igge s he execu ion o all he
ans o ma ions om line 6 o line 13.
ORTZI DATAFRAMES o e speci ic unc ionali ies ailo ed
o p og amming in he cloud con inuum. Fo example, he
p og amme can en o ce he execu ion o ce ain ope a ions on
p op ie a y de ices using he on_p emise lag. This ea u e
is especially use ul o p i acy-sensi i e da a ha canno be
o loaded o he cloud in i s aw o ma . Addi ionally, use s can
choose o scale esou ces (i necessa y) using cloud unc ions,
cloud VMs, o bo h wi h he se _scaleup unc ion.
Concep De ini ion
Task The smalles uni o execu ion, ypically p ocessing
a subse o he da a.
S age A g ouping o unc ionally homogeneous asks ha
sha e dependencies and can un concu en ly.
Exchange The da a ans e be ween wo o mo e s ages, in-
ol ing da a shu ling o edis ibu ion among asks.
Pa i ion Logical subse o he da a being analyzed, p ocessed
by a single ask.
TABLE I: Key concep s in ORTZI’s design.
Job compila ion. ORTZI in e nals a e designed a ound widely
used pa e ns in dis ibu ed da a analy ics. Table Iin oduces
he ounda ional concep s in eg al o ORTZI’s a chi ec u e.
ORTZI DATAFRAMES adop a lazy execu ion model o job
compu a ion, inspi ed by Spa k RDDs [26]. T ans o ma ions,
such as map and so , a e de e ed un il an ac ion, such as
collec , is in oked.
As he use adds ans o ma ions h ough he in e ace, he
in e nal applica ion cons uc s a logical plan, which is a high-
le el ep esen a ion o he ope a ions equi ed o p ocess and
analyze he da a. The logical plan consis s o s ages o ganized
as a di ec ed acyclic g aph (DAG). Na ow ans o ma ions a e
pipelined wi hin he same s age, while wide ans o ma ions
de ine s age bounda ies.
A he end o each s age, asks w i e hei pa i ions o
in e media e s o age. These pa i ions a e shu led, wi h each
ask in he subsequen s age pulling i s inpu pa i ions. We
e e o his p ocess o shu ling pa i ions as an exchange.
Figu e 1a illus a es he s eps in gene a ing he logical plan
o a so ope a o , as he one in Lis ing 1. So ing is
implemen ed in a dis ibu ed manne ; he e o e, i comp ises
wo s ages connec ed by a da a exchange.
Upon in oking an ac ion, he applica ion au oma ically
compiles he inal physical plan based on he logical plan.
In ORTZI, he physical plan is o med by b eaking down
he logical plan in o execu able asks, de ining pa i ions, and
es ablishing dependencies. Tasks a e e sa ile, meaning ha
hey can be execu ed in any unde lying backend.
In ou cu en implemen a ion, he numbe o asks in
na ow ans o ma ions is de e mined using a s a ic pa i ion
size heu is ic. Fo wide ans o ma ions, howe e , he numbe
o pa i ions pe s age is in e ed using he See esou ce
p o isioning algo i hm [7]. We depic a physical plan example
o he so ope a o in Figu e 1a.
Adap i e execu ion. To execu e he job, we schedule asks
c oss a ailable esou ces. A wo ke is de ined as a p ocess
esponsible o ecei ing and execu ing hese asks. Wo ke s
op imize da a exchanges by dynamically adap ing o he un-
de lying s o age backend, enabling seamless da a exchanges
ac oss epheme al, pe sis en and ex e nal s o age.
Wo ke s u ilize sha ed memo y de ices o exchanges be-
ween colloca ed asks, as illus a ed in Figu e 1b. When
sha ed memo y is una ailable—common in mos comme cial
cloud unc ion pla o ms [1]—wo ke s seamlessly all back o
disk o local exchanges. Fo exchanges be ween dis ibu ed
asks, we anspa en ly u ilize a ex e nal s o age se ice,
eachable om he en i e ORTZI a chi ec u e.
A pa i ion may eside in he local s o age o one wo ke
bu be equi ed by a emo e wo ke . In such cases, ORTZI
au oma ically signals he local wo ke o push he pa i ion o
ex e nal s o age, making i accessible o all. We e e o his
ype o exchange, which in ol es mul iple s o age backends,
as a hyb id exchange, as illus a ed in Figu e 1d.
d .so ()
1
2
S age 0
Exchange
S age 1
T0.0
T0.1
T0.2
T1.0
T1.1
T1.2
3
(a)
Wo ke
Wo ke
Wo ke
T0.0
T1.0
T0.1
T1.1
T0.2
T1.2
(b)
Wo ke
Wo ke
Wo ke
T0.0
T1.0
T0.1
T1.1
T0.2
T1.2
(c)
Wo ke
Wo ke
Wo ke
T0.2
T1.2
T0.0
T1.0
T0.1
T1.1
(d)
Fig. 1: Example o an ORTZI execu ion low. (a) T ansla ion
o a ORTZI DATAFRAME so ope a o in o a physical plan
and Job compila ion in ORTZI: 1 Applica ion p og am, 2
Logical plan gene a ion: DAG cons uc ion, 3 Physical plan
gene a ion: ask and dependency de ini ion. The subsequen
images show di e en da a exchange mechanisms: (b) Local
exchange (in-memo y), (c) Ex e nal exchange (objec s o age),
and (d) Hyb id exchange.
ORTZI ully manages he execu ion low, allowing he p o-
g amme o emain agnos ic beyond hei in e ac ion wi h he
DATAFRAME. We choose cloud objec s o age as he ex e nal
s o age backend o he cu en e sion o he sys em, as i
p o ides a cos -e ec i e, always-on se ice [12]. Howe e ,
ORTZI’s modula design acili a es he po en ial in eg a ion
o al e na i e ex e nal s o age sys ems, including se e less
cache p o o ypes [27]–[29] ha could po en ially o e highe
pe o mance.
III. DESIGN
We design ORTZI a chi ec u e o ensu e b oad compa ibili y
in adap i e execu ion. The sys em comp ises h ee main com-
ponen s: he applica ion p og am, a cen alized d i e , and a
pool o dis ibu ed execu o s ha hos he wo ke s. Figu e 2
illus a es he a chi ec u e.
Execu o
Execu o
Schedule
Applica ion
Se e lessDa a ame(). un()
D i e
Pa i ion
manage
Execu o
se e
Task
schedule
Disk
Pa i ion
manage
Wo ke s I/O
handle
Task
egis e
Cloud unc ion
Execu o
Execu o
se e
Task
schedule
Memo y
Pa i ion
manage
Wo ke s I/O
handle
Task
egis e
Execu o (s)
Cloud i ual ins ance
Kube ne es node
Execu o (s)
Task
me ada a
Objec
s o age
Clien machine
Execu o
Resou ce
manage
Fig. 2: ORTZI a chi ec u e diag am.
A. Applica ion P og am
The applica ion p og am uns on he clien machine and is
abs ac ed h ough he ORTZI DATAFRAME, wi h which he
use di ec ly in e ac s. I is esponsible o job compila ion
and d i e con igu a ion, managing he d i e ’s li ecycle. The
applica ion delega es esou ce p o isioning o he d i e , by
ans e ing he physical plan along wi h esou ce con igu a-
ion pa ame e s.
B. D i e
The d i e o ches a es job execu ion and esou ce manage-
men using dedica ed modules, ou lined as ollows.
Resou ce Manage . Launches cloud unc ions and manages
he li ecycle o cloud VMs and/o pods wi hin an exis ing
Kube ne es [30] clus e . I dynamically p o isions new cloud
esou ces based on job-speci ic compu a ional needs, such as
he numbe o CPUs and equi ed memo y.
Cu en ly, ORTZI pe mi s he speci ica ion o esou ce allo-
ca ion a s age-le el g anula i y. By de aul , i eage ly scales
compu e esou ces o ensu e he numbe o alloca ed CPUs
aligns wi h he coun o execu able, independen asks a any
gi en ime. Fu u e i e a ions o he amewo k will inco po a e
mo e in elligen esou ce p o isioning s a egies.
Cen al Task Schedule . Ins an ia es execu o s on p o isioned
esou ces and dis ibu es asks ac oss he execu o pool.
Cen al Pa i ion Manage . Main ains a egis y o all sys em
pa i ions, acking hei hos execu o s. I no i ies execu o s
o emo e al eady consumed pa i ions, p e en ing s o age
bloa om o phaned da a—especially use ul in cloud unc ions
whe e s o age esou ces a e limi ed.
C. Execu o s
Execu o s a e ansien ins ances unning on p o isioned
esou ces, asked wi h execu ing assigned ope a ions. They
communica e solely wi h he d i e and he ex e nal s o age
sys em. Wi hin each execu o , ask execu ion is pa allelized
ac oss a pool o wo ke p ocesses. All execu o s sha e he
ollowing componen s, wi h a ia ions only in he echnical
adap a ions equi ed o he speci ic backend.
Execu o Se e . Es ablishes a pe sis en connec ion wi h he
d i e and manages he ou ing o eques s o and om he
execu o modules.
Local Task Schedule . Reques s asks and assigns hem o
he a ailable wo ke s.
Local Pa i ion Manage . T acks locally s o ed pa i ions
and emo es hem om local s o age—ei he memo y o
disk—when no longe equi ed. I also spills pa i ions o
ex e nal s o age upon eques .
In complex jobs wi h a ying esou ce needs, a wo ke
o en execu es mul iple asks wi hin he same s age, whe e
he unc ion code emains consis en . To op imize his, each
wo ke ini ializes an in-memo y ask egis e a s a up wi h
he unc ion code o each s age. The global ask schedule
only has o send me ada a and a gumen s pe ask a un ime,
educing bo h ne wo k o e head and dese ializa ion cos s.
IV. IMPLEMENTATION
We build ORTZI in Py hon, wi h app oxima ely 8k lines o
code. The esou ce manage in ORTZI is based on Li hops [6]
and le e ages Docke con aine s o un execu o s ac oss di -
e en backends. In e nally, ORTZI uses Pola s [25] o da a
s uc u es and PyA ow [31] o da a (de)se ializa ion.
We u ilize asyncio co ou ines o pe o m I/O eques s
wi hin wo ke s, which enables concu ency a low con ex
swi ching o e head. To enhance I/O pe o mance, we apply
ile consolida ion in exchanges, as desc ibed in Ri le [32].
Execu o s and d i e s communica e using gRPC [33]. The
d i e hos s a gRPC se e , whe e execu o s no i y ask
comple ion and pa i ion managemen eques s. Execu o s no
deployed in cloud unc ions also un hei own gRPC se e
o handle no i ica ions om he ask schedule . Howe e , due
o he unadd essable na u e o cloud unc ions, deploying a
gRPC se e on hem is no easible. Ins ead, cloud unc ion
execu o s es ablish a pe sis en s eaming connec ion wi h he
d i e .
To synch onize and communica e be ween componen s o
he execu o we ely on Py hon mul ip ocessing queues.
Howe e , cloud unc ions do no ha e access o sha ed mem-
o y, and hus canno ins an ia e such da a s uc u es. We
o e come his limi a ion by implemen ing a cus om abs ac ion
o mul ip ocessing queues o e pipes o cloud unc ion
execu o s.
ORTZI is a ully open-sou ce p ojec and is a ailable a
h ps://gi hub.com/GEizagui e/o zi-CLOUD2025.
V. EVALUATION
Se up. We use an AWS EC2 [34]m4.10xla ge ins ance
and AWS Lambda [35] unc ions, each alloca ed 1,769MB
o memo y, co esponding o one single CPU alloca ion 2.
To ensu e ai compa isons, we limi he memo y o he
Docke con aine unning he VM execu o o ma ch he o al
agg ega e memo y o he cloud unc ions in he equi alen
se up, and assign each wo ke o a dedica ed CPU. We use
AWS S3 [36] o he cloud objec s o age. We use 5 eplicas
pe con igu a ion and deploy all esou ces in us-eas -1.
We cen e ou e alua ion on h ee key insigh s gained om
de eloping ORTZI.
R1. The ORTZI DATAFRAME p o ides seamless back-
end compa ibili y wi h minimal code modi ica ions.
R2. Exchange pe o mance can be signi ican ly op i-
mized based on backend selec ion and job ai s.
R3. Ou a chi ec u e is pe o man on benchma k ana-
ly ics jobs.
Wo kloads. Ou e alua ion includes wo widely used da a
analy ics benchma ks.
•Te aSo is a common job o assess exchange pe o -
mance in dis ibu ed sys ems [37]. I consis s o wo
s ages wi h an all- o-all communica ion pa e n. I is da a-
p ese ing, meaning ha he DATAFRAME size emains
he same h oughou he execu ion, imposing signi ican
p essu e on he in e media e communica ion.
•TPC-DS (Que y 95) is a da a analy ics benchma k
comp ising eigh s ages wi h complex dependencies, in-
cluding ga he communica ion pa e ns and join ans-
o ma ions. We use a 10GB inpu in ou expe imen s.
A. Se e less backends: no a ule o humb
To e alua e he impac o backend selec ion, we un Te a-
So on homogeneous deploymen s using cloud VMs and
cloud unc ions. We se he numbe o asks pe s age o ma ch
he numbe o wo ke s. Figu e 3p esen s execu ion imes o
1GB, 2GB, and 5GB Te aSo ac oss he es ed con igu a ions.
In he VM se up, we deploy a single execu o wi h he
speci ied numbe o wo ke s. We e alua e pe o mance using
h ee s o age backends: disk, objec s o age, and memo y.
In cloud unc ions we launch a sepa a e cloud unc ion o
each execu o , each unc ion unning a single wo ke . We es
cloud unc ions exclusi ely wi h objec s o age, as hei sha ed
memo y is limi ed and disk space is es ic ed.
Cloud unc ions ou pe o m VMs in s ong scaling. The
assump ion ha sha ed memo y always p o ides he bes
pe o mance should be app oached wi h cau ion. A low
pa allelism le els, co-loca ed wo ke s wi h in-memo y com-
munica ion pe o m be e . Howe e , as pa allelism inc eases,
2h ps://docs.aws.amazon.com/lambda/la es /dg/con igu a ion-memo y.h ml
dis ibu ed wo ke s exhibi supe io s ong scaling, achie ing
op imal execu ion imes a much highe le els.
This ad an age s ems om wo key ac o s: (1) he o e head
o managing mul iple wo ke s on he same machine and
(2) esou ce con en ion ac oss ne wo k, disk, and memo y
when all wo ke s sha e a de ice. This insigh is pa icula ly
ele an in scena ios wi h memo y cons ain s and in ensi e
da a exchanges, necessi a ing al e na i e s o age backends.
Monoli hs emain complex. Execu ing se e less wo kloads
on a single se e s ill p o ides aluable insigh s. While
memo y exchanges o e he bes pe o mance, hei scaling
beha io closely esembles ha o pe sis en s o age, likely
due o (de)se ializa ion and concu ency o e heads.
Disk and objec s o age pe o m simila ly, especially as
inpu sizes inc ease. Objec s o age is p e e able o deploy-
men s wi h mul iple disagg ega ed execu o s, as i enables
ubiqui ous access wi hou equi ing he d i e o manage
pa i ions. Howe e , local disk exchanges can help educe
ne wo k conges ion and imp o e scalabili y wi h la ge inpu s.
Ou esul s show ha backend and esou ce p o isioning
decisions a e complex and signi ican ly impac he execu ion
ime o da a-in ensi e jobs (R2). The selec ion o he op imal
backend eme ges as a mul i ace ed p oblem, p esen ing an
oppo uni y o u u e esea ch.
ORTZI makes backend po abili y easy. To showcase he
e sa ili y and simplici y o ORTZI DATAFRAMES ac oss
cloud con inuum esou ces, we use he same applica ion code
o all backend e alua ions (R1). All expe imen s in his
sec ion employ he so p imi i e om Lis ing 1. Be o e
in oking so , he se _scale_ou unc ion is called o
selec ei he cloud unc ions o VMs.
B. A se e less expe ience a low cos
The ORTZI DATAFRAME inco po a es mul iple manage-
men laye s o ensu e se e less usabili y. To e alua e i s e i-
ciency and disca d excessi e o e head, in Table II we compa e
i s execu ion ime in he Te aSo benchma k agains See [7],
a se e less so ope a o buil di ec ly on Li hops [6]— he
same esou ce p o isioning engine used by ORTZI.
F amewo k Execu ion Time (s) O e head %
See [7] 18.45 –
ORTZI 19.17 6.78
TABLE II: See and ORTZI execu ion ime in a 5GB Te aSo .
ORTZI in oduces minimal managemen o e head, in-
cu ing only a 6.78% la ency inc ease compa ed o See ,
which ope a es di ec ly on cloud unc ions wi hou gene al-
pu pose usabili y enhancemen s. This compa able pe o mance
is achie ed despi e ORTZI’s addi ional abs ac ions, such as
he execu o ,wo ke , and i s mul iple o ches a ion laye s.
C. ORTZI wi hin he S a e-o - he-A
We compa e he execu ion ime o ORTZI agains Jol eon
[2], an au oma ic esou ce p o isioning amewo k o se e -
less da a analy ics. We con igu e Jol eon o mee a 30-second
4 8 12 16 20 24
Numbe o wo ke s
10
20
30
Execu ion ime (s)
EC2 - S3
EC2 - EBS
EC2 - Memo y
Lambda - S3
(a) 1GB inpu
4 8 12 16 20 24 28
Numbe o wo ke s
20
40
Execu ion ime (s)
(b) 2GB inpu
12 16 20 24 28 32 36 40
Numbe o wo ke s
25
50
75
Execu ion ime (s)
(c) 5GB inpu
Fig. 3: Execu ion ime o Te aso benchma k ac oss a ying inpu sizes and numbe s o pa allel wo ke s ( asks). We compa e
ou di e en se e less con igu a ions: co-loca ed wo ke s in an AWS EC2 ins ance using sha ed memo y, an EBS (HDD)
olume, o AWS S3 o he da a exchange, and dis ibu ed wo ke s in AWS Lambda ins ances using AWS S3.
SLO while minimizing cos and eplica e i s exac same
esou ce con igu a ion in ORTZI. Addi ionally, we compa e
bo h sys ems o EMR Se e less [21], he se e less Spa k
p oduc in AWS. Figu e 4p esen s he la ency and cos
esul s. O e all, ORTZI deli e s pe o mance on pa wi h
o supe io o equi alen p o o ypes (R3), demons a ing ha
an open-sou ce, gene al-pu pose amewo k can compe e wi h
comme cial solu ions and specialized esea ch p o o ypes.
We deli e a pe o mance-compe i i e a chi ec u e. ORTZI
achie es la ency and cos compa able o Jol eon when using
iden ical esou ce con igu a ions. This esul is expec ed, as
bo h ely on AWS Lambda and S3. Rela i e o AWS EMR
Se e less, ORTZI imp o es pe o mance by an imp essi e
70.59%. This gain may be a ibu ed o slow scaling imes
o EMR Se e less, as p e iously epo ed in [38].
Le e aging hyb id backends ansla es in o pe o mance
gains. To illus a e he bene i s o cloud con inuum a chi ec-
u es (R2), we main ain Jol eon’s speci ica ions (i.e., numbe
o CPUs and memo y) bu modi y he esou ce alloca ion
s a egy: we o load scan ans o ma ions o cloud unc ions
while unning he emaining compu a ions on a cloud VM. We
e e o his con igu a ion as hyb id in Figu e 4.
The hyb id deploymen cu s la ency by 25.57% e sus
Jol eon’s cloud- unc ion app oach. This speedup comes mainly
om swapping AWS S3 da a exchanges wi h as e in-memo y
ans e s be ween ou pos s ages, especially hose wi h low
pa allelism, hus educing I/O ime.
20 40 60 80
Execu ion ime (s)
0
1
2
Cos ($0.01)
Jol eon
EMR Se e less
O zi (cloud unc ions)
O zi (hyb id)
Fig. 4: TPC-DS la ency in Jol eon [2], EMR Se e less [21]
and ORTZI (using a cloud unc ion and a hyb id deploymen ).
Fu he esea ch could explo e he ad an ages o hyb id,
con inuum-na i e deploymen s by inco po a ing he e ogeneous
ha dwa e (like GPUs and FPGAs) o ligh weigh IoT de-
ices. Gi en his pape ’s ocus on demons a ing he design
o ORTZI, hese speci ic explo a ions all ou side i s immedia e
pu iew bu o e signi ican a enues o u u e in es iga ion.
VI. RELATED WORK
Ea ly se e less amewo ks [39]–[41] in oduced abs ac-
ions o o loading local compu a ions o cloud unc ions,
pa ing he way o he i s se e less da a analy ics ope a o s
[5], [6]. Howe e , hese solu ions o e limi ed buil -in ope -
a o s, especially hose wi h added complexi y such as so .
ORTZI DATAFRAMES p o ide a comp ehensi e amewo k o
da a analy ics, including buil -in wide ans o ma ions.
Fully- ledged se e less da a analy ics se ices a e a ailable
oday as comme cial solu ions om cloud p o ide s [21], [42],
[43] and specialized companies [22], [44], [45]. Howe e ,
hese o e ings incu addi ional cos s o esou ce manage-
men . In esea ch, e o s ins ead ei he ocus exclusi ely on
cloud unc ions [9], [10] o equi e homogeneous esou ce
pools [?]. Building se e less amewo ks o e hyb id back-
ends has also been explo ed in p io wo k [46] and app oached
h ough a ious p oposals [8], [16], [47]. Exis ing p ojec s,
howe e , ei he do no deli e ou -o - he-box usabili y o a e
limi ed o a ew backends. The ORTZI DATAFRAME is he i s
ully open-sou ce, se e less da a analy ics abs ac ion o p o-
ide seamless esou ce managemen in he cloud con inuum.
Da a exchanges in se e less a chi ec u es ha e been e-
isi ed mul iple imes [4], [7], [12], [13], [48], including in-
memo y exchanges o co-loca ed asks [4], [38], [49]. Ex en-
si e esea ch has also ocused on al e na i e communica ion
me hods o cloud unc ions [14], [50]. Howe e , exis ing
p oposals a e no s ill in eg a ed wi hin b oade da a analy ics
amewo ks, limi ing hei lexibili y and applicabili y. Simila
limi a ions apply o domain-speci ic amewo ks [51], [52].
VII. CONCLUSION
We p esen ORTZI, a se e less da a analy ics amewo k
designed o execu e p og amma ic wo kloads ac oss he e o-
geneous backends. ORTZI o e s an adap i e a chi ec u e
suppo ing cloud unc ions, cloud i ual ins ances, and on-
p emise clus e s, u ilizing sha ed memo y, disk and objec
s o age o exchanges. We demons a e ha ORTZI e ec i ely
na iga es compu e and s o age backends wi h minimal p o-
g amming e o .
Despi e being open-sou ce and gene al-pu pose, ORTZI
achie es compe i i e pe o mance compa ed o bo h comme -
cial and esea ch p oduc s. ORTZI is con inually e ol ing and
is ac i ely used in ongoing esea ch p ojec s on cloud-edge
analy ics and sma esou ce alloca ion.
REFERENCES
[1] E. Jonas, J. Schleie -Smi h, V. S eekan i, C.-C. Tsai, A. Khandelwal,
Q. Pu, V. Shanka , J. Ca ei a, K. K au h, N. Yadwadka , J. E.
Gonzalez, R. A. Popa, I. S oica, and D. A. Pa e son, “Cloud
P og amming Simpli ied: A Be keley View on Se e less Compu ing,”
2019. [Online]. A ailable: h ps://a xi .o g/abs/1902.03383
[2] Z. Zhang, C. Jin, and X. Jin, “Jol eon: Unleashing he P omise o
Se e less o Se e less Wo k lows,” in 21s USENIX Symposium on
Ne wo ked Sys ems Design and Implemen a ion (NSDI 24). San a
Cla a, CA: USENIX Associa ion, Ap . 2024, pp. 167–183. [Online].
A ailable: h ps://dl.acm.o g/doi/10.5555/3691825.3691835
[3] C. Jin, Z. Zhang, X. Xiang, S. Zou, G. Huang, X. Liu, and X. Jin, “Di o:
E icien Se e less Analy ics wi h Elas ic Pa allelism,” in P oceedings
o he ACM SIGCOMM 2023 Con e ence, se . ACM SIGCOMM ’23.
New Yo k, NY, USA: Associa ion o Compu ing Machine y, 2023, p.
406–419. [Online]. A ailable: h ps://doi.o g/10.1145/3603269.3604816
[4] T. Li, Y. Li, W. Zhu, Y. Xu, and J. C. S. Lui, “MinFlow:
High-pe o mance and Cos -e icien Da a Passing o I/O-in ensi e
S a e ul Se e less Analy ics,” in 22nd USENIX Con e ence on
File and S o age Technologies, se . FAST ’24. San a Cla a, CA:
USENIX Associa ion, Feb. 2024, pp. 311–327. [Online]. A ailable:
h ps://dl.acm.o g/doi/10.5555/3650697.3650716
[5] E. Jonas, Q. Pu, S. Venka a aman, I. S oica, and B. Rech , “Occupy
he cloud: dis ibu ed compu ing o he 99%,” in P oceedings o he
2017 Symposium on Cloud Compu ing, se . SoCC ’17. New Yo k,
NY, USA: Associa ion o Compu ing Machine y, 2017, p. 445–451.
[Online]. A ailable: h ps://doi.o g/10.1145/3127479.3128601
[6] J. Samp´
e, G. Ve nik, M. S´
anchez-A igas, and P. Ga c´
ıa-L´
opez,
“Se e less Da a Analy ics in he IBM Cloud,” in P oceedings o
he 19 h In e na ional Middlewa e Con e ence Indus ial T ack, se .
Middlewa e Indus ial T ack ’18. New Yo k, NY, USA: Associa ion
o Compu ing Machine y, 2018, p. 1–8. [Online]. A ailable:
h ps://doi.o g/10.1145/3284028.328402
[7] M. S´
anchez-A igas and G. T. Eizagui e, “A see knows bes : op imized
objec s o age shu ling o se e less analy ics,” in P oceedings o he
23 d ACM/IFIP In e na ional Middlewa e Con e ence, se . Middlewa e
’22. New Yo k, NY, USA: Associa ion o Compu ing Machine y,
2022, p. 148–160. [Online]. A ailable: h ps://doi.o g/10.1145/3528535.
3565241
[8] H. Bian, T. Sha, and A. Ailamaki, “Using Cloud Func ions as
Accele a o o Elas ic Da a Analy ics,” P oc. ACM Manag. Da a, ol. 1,
no. 2, Jun. 2023. [Online]. A ailable: h ps://doi.o g/10.1145/3589306
[9] B. Ca e , J. Zhang, A. Wang, A. Anwa , P. Wu, and Y. Cheng,
“Wukong: a scalable and locali y-enhanced amewo k o se e less
pa allel compu ing,” in P oceedings o he 11 h ACM Symposium
on Cloud Compu ing, se . SoCC ’20. New Yo k, NY, USA:
Associa ion o Compu ing Machine y, 2020, p. 1–15. [Online].
A ailable: h ps://doi.o g/10.1145/3419111.3421286
[10] Y. Kim and J. Lin, “Se e less Da a Analy ics wi h Flin ,” in P oceedings
o he IEEE 11 h In e na ional Con e ence on Cloud Compu ing, se .
CLOUD ’18, Los Alami os, CA, USA, Jul. 2018, pp. 451–455.
[Online]. A ailable: h ps://doi.o g/10.1109/CLOUD.2018.00063
[11] A. M. Nes o o , D. Ma ´
on, A. Gu ie ez-To e, C. Wang, C. Misale,
A. Yousse , D. Ca e a, and J. L. Be al, “Dex e : A Pe o mance-
Cos E icien Resou ce Alloca ion Manage o Se e less Da a
Analy ics,” in P oceedings o he 25 h In e na ional Middlewa e
Con e ence, se . Middlewa e ’24. New Yo k, NY, USA: Associa ion
o Compu ing Machine y, 2024, p. 117–130. [Online]. A ailable:
h ps://doi.o g/10.1145/3652892.370075
[12] Q. Pu, S. Venka a aman, and I. S oica, “Shu ling, Fas and Slow:
Scalable Analy ics on Se e less In as uc u e,” in 16 h USENIX
Symposium on Ne wo ked Sys ems Design and Implemen a ion (NSDI
19). Bos on, MA: USENIX Associa ion, Feb. 2019, pp. 193–206.
[Online]. A ailable: h ps://dl.acm.o g/doi/10.5555/3323234.3323251
[13] I. M¨
ulle , R. Ma oqu´
ın, and G. Alonso, “Lambada: In e ac i e Da a
Analy ics on Cold Da a Using Se e less Cloud In as uc u e,” in
P oceedings o he 2020 ACM SIGMOD In e na ional Con e ence on
Managemen o Da a, se . SIGMOD ’20. New Yo k, NY, USA:
Associa ion o Compu ing Machine y, 2020, p. 115–130. [Online].
A ailable: h ps://doi.o g/10.1145/3318464.3389758
[14] A. Mahgoub, K. Shanka , S. Mi a, A. Klimo ic, S. Cha e ji,
and S. Bagchi, “SONIC: Applica ion-awa e Da a Passing o
Chained Se e less Applica ions,” in 2021 USENIX Annual Technical
Con e ence (USENIX ATC 21). USENIX Associa ion, Jul. 2021, pp.
285–301. [Online]. A ailable: h ps://www.usenix.o g/con e ence/a c21/
p esen a ion/mahgoub
[15] H. Zhang, Y. Tang, A. Khandelwal, J. Chen, and I. S oica,
“Cae us: NIMBLE Task Scheduling o Se e less Analy ics,” in
18 h USENIX Symposium on Ne wo ked Sys ems Design and
Implemen a ion, se . NSDI ’21. USENIX Associa ion, Ap . 2021,
pp. 653–669. [Online]. A ailable: h ps://www.usenix.o g/con e ence/
nsdi21/p esen a ion/zhang-hong
[16] A. D. Mohapa a and K. Oh, “Sma pick: Wo kload P edic ion o
Se e less-enabled Scalable Da a Analy ics Sys ems,” in P oceedings
o he 24 h In e na ional Middlewa e Con e ence, se . Middlewa e ’23.
New Yo k, NY, USA: Associa ion o Compu ing Machine y, 2023, p.
29–42. [Online]. A ailable: h ps://doi.o g/10.1145/3590140.3592850
[17] C. Zhang, M. Yu, W. Wang, and F. Yan, “MA k: Exploi ing Cloud
Se ices o Cos -E ec i e, SLO-Awa e Machine Lea ning In e ence
Se ing,” in 2019 USENIX Annual Technical Con e ence, se . USENIX
ATC ’19. Ren on, WA: USENIX Associa ion, Jul. 2019, pp. 1049–1062.
[Online]. A ailable: h ps://dl.acm.o g/doi/10.5555/3358807.3358897
[18] M. Zaha ia, M. Chowdhu y, T. Das, A. Da e, J. Ma, M. McCauley, M. J.
F anklin, S. Shenke , and I. S oica, “Resilien Dis ibu ed Da ase s:
A Faul -Tole an Abs ac ion o In-Memo y Clus e Compu ing,”
in 9 h USENIX Symposium on Ne wo ked Sys ems Design and
Implemen a ion (NSDI), 2012. [Online]. A ailable: h ps://dl.acm.o g/
doi/10.5555/2228298.2228301
[19] P. Mo i z, R. Nishiha a, S. Wang, A. Tumano , R. Liaw, E. Liang,
M. Elibol, Z. Yang, W. Paul, M. I. Jo dan, and I. S oica, “Ray:
A Dis ibu ed F amewo k o Eme ging AI Applica ions,” in 13 h
USENIX Symposium on Ope a ing Sys ems Design and Implemen a ion,
se . OSDI ’18. Ca lsbad, CA: USENIX Associa ion, Oc . 2018, pp.
561–577. [Online]. A ailable: h ps://dl.acm.o g/doi/10.5555/3291168.
3291210
[20] Dask De elopmen Team, “Dask: Lib a y o Dynamic Task Schedul-
ing,” h ps://www.dask.o g/, 2025, [Online; accessed 2025-05-30].
[21] “EMR Se e less,” h ps://aws.amazon.com/es/em /se e less, [Online;
accessed 2025-05-30].
[22] “Coiled,” h ps://www.coiled.io, [Online; accessed 2025-05-30].
[23] Anyscale, “Scaling Ray Wo kloads wi h Anyscale,” h ps://www.
anyscale.com, 2025, accessed: 2025-03-05.
[24] Pandas De elopmen Team, “pandas: Py hon Da a Analysis Lib a y,”
2025, [Online; accessed 2025-05-30]. [Online]. A ailable: h ps:
//pandas.pyda a.o g/
[25] Pola s De elope s, “Pola s: Ligh ning- as Da aF ame lib a y o Rus
and Py hon,” h ps://gi hub.com/pola- s/pola s, 2025, [Online; accessed
2025-02-26].
[26] M. Zaha ia, M. Chowdhu y, T. Das, A. Da e, J. Ma, M. McCauly, M. J.
F anklin, S. Shenke , and I. S oica, “Resilien Dis ibu ed Da ase s: A
Faul -Tole an Abs ac ion o In-Memo y Clus e Compu ing,” in 9 h
USENIX Symposium on Ne wo ked Sys ems Design and Implemen a ion,
se . NSDI ’12. San Jose, CA: USENIX Associa ion, Ap . 2012,
pp. 15–28. [Online]. A ailable: h ps://dl.acm.o g/doi/10.5555/2228298.
2228301
[27] F. Rome o, G. I. Chaudh y, I. n. Goi i, P. Gopa, P. Ba um, N. J.
Yadwadka , R. Fonseca, C. Kozy akis, and R. Bianchini, “Faa$T:
A T anspa en Au o-Scaling Cache o Se e less Applica ions,” in
P oceedings o he ACM Symposium on Cloud Compu ing, se . SoCC
’21. New Yo k, NY, USA: Associa ion o Compu ing Machine y,
2021, p. 122–137. [Online]. A ailable: h ps://doi.o g/10.1145/3472883.
3486974
[28] D. M ondo, M. Bacou, K. Ngue chouang, L. Ngale, S. Pouge ,
J. Kouam, R. Lachaize, J. Hwang, T. Wood, D. Hagimon ,
N. De Palma, B. Ba chakui, and A. Tchana, “OFC: an oppo unis ic
caching sys em o FaaS pla o ms,” in P oceedings o he Six een h
Eu opean Con e ence on Compu e Sys ems, se . Eu oSys ’21. New
Yo k, NY, USA: Associa ion o Compu ing Machine y, 2021, p.
228–244. [Online]. A ailable: h ps://doi.o g/10.1145/3447786.3456239
[29] A. Klimo ic, Y. Wang, P. S uedi, A. T i edi, J. P e e le, and
C. Kozy akis, “Pocke : Elas ic Epheme al S o age o Se e less
Analy ics,” in 13 h USENIX Symposium on Ope a ing Sys ems
Design and Implemen a ion, se . OSDI ’18. Ca lsbad, CA: USENIX
Associa ion, Oc . 2018, pp. 427–444. [Online]. A ailable: h ps:
//dl.acm.o g/doi/10.5555/3291168.3291200
[30] Kube ne es Au ho s, “Kube ne es: P oduc ion-G ade Con aine O ches-
a ion,” h ps://kube ne es.io/, 2025, accessed: 2025-02-26.
[31] A. S. Founda ion, “PyA ow: Py hon lib a y o Apache A ow,” h ps:
//gi hub.com/apache/a ow, 2025, [Online; accessed 2025-02-26].
[32] H. Zhang, B. Cho, E. Sey e, A. Ching, and M. J. F eedman, “Ri le:
op imized shu le se ice o la ge-scale da a analy ics,” in P oceedings
o he Thi een h Eu oSys Con e ence, se . Eu oSys ’18. New Yo k,
NY, USA: Associa ion o Compu ing Machine y, 2018. [Online].
A ailable: h ps://doi.o g/10.1145/3190508.3190534
[33] Google, “gRPC: A high-pe o mance, open-sou ce uni e sal RPC ame-
wo k,” h ps://g pc.io/, 2025, [Online; accessed 2025-02-26].
[34] “AWS EC2,” 2025, accessed: 2025-02-28. [Online]. A ailable:
h ps://aws.amazon.com/ec2/
[35] “AWS Lambda,” 2025, accessed: 2025-02-28. [Online]. A ailable:
h ps://aws.amazon.com/lambda/
[36] “AWS S3,” 2025, accessed: 2025-02-28. [Online]. A ailable: h ps:
//aws.amazon.com/s3/
[37] O. O’Malley, “Te aBy e So on Apache Hadoop,” Yahoo!, Tech.
Rep., May 2008. [Online]. A ailable: h ps://so benchma k.o g/
YahooHadoop.pd
[38] D. Ba celona-Pons, A. A jona, P. Ga c´
ıa-L´
opez, E. Molina-Gim´
enez,
and S. Klymonchuk, “FaaS Is No Enough: Se e less Handling
o Bu s -Pa allel Jobs,” 2024. [Online]. A ailable: h ps://doi.o g/10.
48550/a Xi .2407.14331
[39] S. Fouladi, R. S. Wahby, B. Shackle , K. V. Balasub amaniam,
W. Zeng, R. Bhale ao, A. Si a aman, G. Po e , and K. Wins ein,
“Encoding, Fas and Slow: Low-La ency Video P ocessing Using
Thousands o Tiny Th eads,” in 14 h USENIX Symposium on
Ne wo ked Sys ems Design and Implemen a ion, se . NSDI ’17.
Bos on, MA: USENIX Associa ion, Ma . 2017, pp. 363–376. [Online].
A ailable: h p://dl.acm.o g/doi/10.5555/3154630.3154660
[40] S. Fouladi, F. Rome o, D. I e , Q. Li, S. Cha e jee, C. Kozy akis,
M. Zaha ia, and K. Wins ein, “F om Lap op o Lambda: Ou sou cing
E e yday Jobs o Thousands o T ansien Func ional Con aine s,” in
2019 USENIX Annual Technical Con e ence, se . USENIX ATC ’19.
Ren on, WA: USENIX Associa ion, Jul. 2019, pp. 475–488. [Online].
A ailable: h ps://dl.acm.o g/doi/10.5555/3358807.3358848
[41] A. A jona, G. Finol, and P. G. L´
opez, “T anspa en se e less
execu ion o Py hon mul ip ocessing applica ions,” Fu u e Gene a ion
Compu e Sys ems, ol. 140, pp. 436–449, 2023. [Online]. A ailable:
h ps://doi.o g/10.1016/j. u u e.2022.10.038
[42] “BigQue y,” h ps://cloud.google.com/bigque y, [Online; accessed 2025-
05-30].
[43] “Da ap oc Se e less,” h ps://cloud.google.com/da ap oc-se e less,
[Online; accessed 2025-05-30].
[44] “Modal,” h ps://www.modal.com, [Online; accessed 2025-05-30].
[45] J. Tagliabue, T. Ca aza-Ha e , and C. G eco, “Bauplan: Ze o-copy,
Scale-up FaaS o Da a Pipelines,” in P oceedings o he 10 h
In e na ional Wo kshop on Se e less Compu ing, se . WoSC 10. New
Yo k, NY, USA: Associa ion o Compu ing Machine y, 2024, p.
31–36. [Online]. A ailable: h ps://doi.o g/10.1145/3702634.3702955
[46] P. Ga c´
ıa-L´
opez, M. S´
anchez-A igas, S. Shillake , P. Pie zuch,
D. B ei gand, G. Ve nik, P. Su a, T. Ta an , and A. J. Fe e ,
“Se e Mix: T adeo s and Challenges o Se e less Da a Analy ics,”
2019. [Online]. A ailable: h ps://a xi .o g/abs/1907.11465
[47] G. T. Eizagui e, D. Ba celona-Pons, A. A jona, G. Ve nik, P. Ga c´
ıa-
L´
opez, and T. Alexand o , “Se e ul Func ions: Le e aging Se e s
in Complex Se e less Wo k lows (Indus y T ack),” in P oceedings o
he 25 h In e na ional Middlewa e Con e ence Indus ial T ack, se .
Middlewa e Indus ial T ack ’24. New Yo k, NY, USA: Associa ion
o Compu ing Machine y, 2024, p. 15–21. [Online]. A ailable:
h ps://doi.o g/10.1145/3700824.3701095
[48] M. S´
anchez-A igas, G. T. Eizagui e, G. Ve nik, L. S ua , and
P. Ga c´
ıa-L´
opez, “P imula: a P ac ical Shu le/So Ope a o o
Se e less Compu ing,” in P oceedings o he 21s In e na ional
Middlewa e Con e ence Indus ial T ack, se . Middlewa e Indus ial
T ack ’20. New Yo k, NY, USA: Associa ion o Compu ing
Machine y, 2020, p. 31–37. [Online]. A ailable: h ps://doi.o g/10.1145/
3429357.3430522
[49] S. Qi, L. Monis, Z. Zeng, I.-c. Wang, and K. K. Ramak ishnan,
“SPRIGHT: ex ac ing he se e om se e less compu ing! high-
pe o mance eBPF-based e en -d i en, sha ed-memo y p ocessing,” in
P oceedings o he ACM SIGCOMM 2022 Con e ence, se . SIGCOMM
’22. New Yo k, NY, USA: Associa ion o Compu ing Machine y,
2022, p. 780–794. [Online]. A ailable: h ps://doi.o g/10.1145/3544216.
3544259
[50] M. Waw zoniak, I. M¨
ulle , G. Alonso, and R. B uno, “Boxe : Da a
Analy ics on Ne wo k-enabled Se e less Pla o ms,” in P oceedings
o he 11 h Con e ence on Inno a i e Da a Sys ems Resea ch.
www.cid db.o g, 2021. [Online]. A ailable: h ps://doi.o g/10.3929/
e hz-b-000456492
[51] V. Shanka , K. K au h, K. Vod ahalli, Q. Pu, B. Rech , I. S oica,
J. Ragan-Kelley, E. Jonas, and S. Venka a aman, “Se e less linea
algeb a,” in P oceedings o he 11 h ACM Symposium on Cloud
Compu ing, se . SoCC ’20. New Yo k, NY, USA: Associa ion
o Compu ing Machine y, 2020, p. 281–295. [Online]. A ailable:
h ps://doi.o g/10.1145/3419111.3421287
[52] L. Toade , A. U a, A. Musaa i , and A. Iosup, “G aphless:
Towa d Se e less G aph P ocessing,” in P oceedings o he 18 h
In e na ional Symposium on Pa allel and Dis ibu ed Compu ing,
se . ISPDC ’19, 2019, pp. 66–73. [Online]. A ailable: h ps:
//doi.o g/10.1109/ISPDC.2019.00012
