Co esponding au ho : Mu husel am Chand amohan.
Copy igh © 2025 Au ho (s) e ain he copy igh o his a icle. This a icle is published unde he e ms o he C ea i e Commons A ibu ion License 4.0.
Disk p o iling and cos op imiza ion o la ency-sensi i e Cassand a DB (NoSQL DB)
Mu husel am Chand amohan *
Leading Comme cial Banking in USA.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2973-2978
Publica ion his o y: Recei ed on 03 Ap il 2025; e ised on 11 May 2025; accep ed on 13 May 2025
A icle DOI: h ps://doi.o g/10.30574/wja .2025.26.2.1852
Abs ac
This a icle explo es he c i ical ole o disk s o age selec ion o op imizing Cassand a da abase pe o mance in la ency-
sensi i e en i onmen s. Beginning wi h an examina ion o IOPS measu emen echniques, he discussion p og esses
h ough Cassand a's unique wo kload pa e ns, comp ehensi e disk p o iling me hodologies, igo ous benchma king
app oaches, and cos -e ec i e op imiza ion s a egies. By unde s anding he in e play be ween sequen ial w i es,
andom eads, and mixed I/O pa e ns du ing compac ion p ocesses, da abase adminis a o s can make in o med
decisions when selec ing app op ia e s o age solu ions. The a icle emphasizes he impo ance o e alua ing sus ained
pe o mance a he han elying on bu s capabili ies o endo speci ica ions, while highligh ing how ie ed s o age
app oaches can balance pe o mance equi emen s wi h budge cons ain s. Th ough sys ema ic e alua ion and
wo kload-speci ic op imiza ion, o ganiza ions can achie e op imal pe o mance o NoSQL deploymen s while
con olling in as uc u e cos s.
Keywo ds: NoSQL pe o mance; IOPS measu emen ; Disk p o iling; Tie ed s o age; La ency op imiza ion
1. In oduc ion
In he e a o eal- ime ansac ions and digi al mode niza ion, NoSQL da abases a e c i ical o deli e ing millisecond
esponse imes while managing massi e da a olumes. The choice o disk s o age plays a pi o al ole in ensu ing an
op imal, scalable, and cos -e ec i e da abase in as uc u e. Fac o s such as disk ype, capaci y, pe o mance, and cos
mus be ca e ully e alua ed o suppo high- h oughpu , low-la ency applica ions.
The mode n digi al landscape demands da abase sys ems capable o handling ex ensi e wo kloads wi h minimal
la ency. Cassand a DB, as desc ibed by A inash Lakshman and P ashan Malik, was designed speci ically o un on
hund eds o nodes ac oss mul iple da a cen e s wi h asynch onous mas e less eplica ion, p o iding highly a ailable
se ice wi h no single poin o ailu e [1]. This a chi ec u e makes Cassand a pa icula ly sensi i e o s o age
pe o mance cha ac e is ics, as i mus manage a w i e-op imized s o age engine wi h backg ound compac ion
p ocesses.
The implemen a ion o e icien s o age solu ions is exempli ied by s eaming pla o m’s Cassand a deploymen , which
s a ed wi h jus 12 nodes in 2013 bu g ew subs an ially o suppo c i ical s eaming se ices. By implemen ing ca e ul
disk p o iling and op imiza ion, s eaming pla o m's in as uc u e e ol ed o manage a Cassand a clus e wi h
hund eds o nodes s o ing me ada a in a 2.6TB da ase [2]. Thei expe ience demons a es how p ope s o age selec ion
di ec ly impac s he abili y o main ain pe o mance a scale, pa icula ly o w i e-hea y wo kloads in dis ibu ed
en i onmen s.
This a icle explo es s a egies o selec ing he mos e icien disk solu ions o NoSQL da abases, balancing
pe o mance equi emen s agains cos conside a ions while ensu ing seamless scalabili y o e ol ing digi al
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2973-2978
2974
ecosys ems. By implemen ing s uc u ed app oaches o s o age e alua ion and selec ion, o ganiza ions can achie e
op imal pe o mance o la ency-sensi i e wo kloads wi hou unnecessa y in as uc u e expendi u e.
2. IOPS Measu emen : The Founda ion o Disk Selec ion
Op imizing disk selec ion o Cassand a DB equi es a s uc u ed app oach, beginning wi h IOPS measu emen . Disk
IOPS signi ican ly impac la ency and concu ency, making i essen ial o measu e ac ual pe o mance using ools like
BCC- ools (caches a , biola ency, x sslowe , x sdis ), eBPF, and FIO a he han elying solely on endo speci ica ions.
Selec ing he igh ool ensu es accu a e pe o mance insigh s o in o med decision-making.
Accu a e IOPS measu emen has become mo e sophis ica ed wi h he eme gence o eBPF echnology, which p o ides
unp eceden ed isibili y in o s o age pe o mance. Acco ding o B endanG egg's comp ehensi e wo k on BPF
Pe o mance Tools, adi ional s o age benchma ks o en miss c i ical pe o mance cha ac e is ics ha a ec da abase
ope a ions. The book de ails how biola ency, one o he BCC ools, can measu e disk I/O la ency dis ibu ions wi h
mic osecond p ecision ac oss di e en ope a ion ypes, e ealing pe o mance cli s ha endo speci ica ions
equen ly obscu e [3]. This g anula isibili y is pa icula ly aluable o Cassand a deploymen s, whe e occasional
la ency spikes can d ama ically impac applica ion pe o mance e en when a e age IOPS appea adequa e.
When conduc ing s o age e alua ions o Cassand a, i 's c ucial o es wi h wo kloads ha accu a ely e lec p oduc ion
pa e ns. Ins aclus 's benchma king esea ch demons a es his impo ance h ough hei sys ema ic app oach o
measu ing node h oughpu o Apache Cassand a. Thei es ing e ealed ha he ype o s o age olume signi ican ly
impac ed pe o mance, wi h es s o AWS gp2 olumes showing pe o mance deg ada ion a a ound 3,000 IOPS pe
olume, while p o isioned IOPS olumes main ained consis en pe o mance up o hei con igu ed limi s [4]. The
benchma k es ing used he Yahoo! Cloud Se ing Benchma k (YCSB) wi h a mixed 50/50 ead/w i e wo kload,
demons a ing ha a single i3.2xla ge ins ance wi h NVMe s o age could achie e up o 38,680 ope a ions pe second,
as ly ou pe o ming equi alen gp2-based ins ances a jus 13,561 ope a ions pe second.
The benchma king me hodology mus inco po a e ealis ic da a sizes ha accoun o Cassand a's compac ion
o e head. Ins aclus 's indings highligh ha s o age pe o mance du ing compac ion can d op by 30-40% i he disk
selec ion doesn' accoun o his wo kload, a ec ing o e all clus e h oughpu [4]. Thei es ing amewo k
inco po a ed a ying da a sizes o ensu e ealis ic measu emen o compac ion impac , using da ase s ha g ew o
300GB pe node du ing ex ended es ing pe iods.
When selec ing measu emen ools o IOPS e alua ion, adminis a o s should combine low-le el ke nel acing
p o ided by BCC- ools wi h applica ion-speci ic benchma king amewo ks o cap u e bo h he aw s o age capabili ies
and hei impac on ac ual Cassand a pe o mance. This mul i-laye ed app oach p o ides he mos accu a e ounda ion
o s o age decision-making, ensu ing ha selec ed disk solu ions will deli e consis en pe o mance unde he
complex I/O pa e ns ha Cassand a gene a es in p oduc ion en i onmen s.
Table 1 Cassand a Pe o mance Compa ison Ac oss S o age Types [3, 4]
S o age Type
Ope a ions Pe Second (50/50
Read/W i e)
Pe o mance Du ing Compac ion
(% Reduc ion)
La ency Measu emen
P ecision
AWS NVMe
(i3.2xla ge)
38,680
30-40%
Mic osecond
AWS gp2
13,561
30-40%
AWS P o isioned
IOPS
No speci ied
30-40%
3. Unde s anding key cassand a wo kload pa e ns
Cassand a's dis ibu ed a chi ec u e c ea es unique disk access pa e ns ha mus be conside ed when selec ing
s o age solu ions. The da abase pe o ms sequen ial w i es o commi logs, andom eads o que ies, and bo h
sequen ial and andom I/O du ing compac ion p ocesses. These di e se pa e ns equi e s o age solu ions ha can
handle mixed wo kloads wi hou pe o mance deg ada ion.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2973-2978
2975
The dis inc i e I/O cha ac e is ics o Cassand a s em di ec ly om i s a chi ec u e as a dis ibu ed NoSQL da abase
designed o high a ailabili y and scalabili y. As explained by Ins aclus , Cassand a uses a dis ibu ed a chi ec u e wi h
a ing design ha pa i ions da a ac oss mul iple nodes, allowing i o dis ibu e wo kloads e enly o p e en
bo lenecks [5]. This a chi ec u al app oach di ec ly in luences disk access pa e ns. Cassand a's w i e pa h i s eco ds
ope a ions sequen ially in commi logs o du abili y, hen s o es da a in memo y s uc u es called mem ables, and
inally lushes hese o disk as immu able SSTables. This mul i-s age p ocess c ea es a complex I/O p o ile whe e
sequen ial w i es p edomina e in no mal ope a ions, bu andom I/O inc eases signi ican ly du ing eads and
main enance ope a ions.
The complexi y o hese wo kload pa e ns becomes pa icula ly e iden du ing scaling e en s and main enance
ope a ions. Acco ding o Walma Global Tech's analysis o Cassand a deploymen s, disk I/O pa e ns change
d ama ically du ing compac ion p ocesses [6]. Thei p oduc ion clus e s exhibi ed no mal ead la encies o 3-5ms, bu
hese inc eased o 15-20ms du ing hea y compac ion ac i i ies. The enginee ing eam obse ed ha compac ion
p ocesses could consume up o 25% o disk I/O capaci y on no mal wo kloads, bu his inc eased o o e 70% du ing
peak a ic pe iods, c ea ing po en ial o cascading pe o mance issues i no p ope ly managed.
This sensi i i y o disk pe o mance becomes especially c i ical when conside ing Cassand a's epai ope a ions.
Walma 's enginee s documen ed ha du ing epai p ocesses, disk I/O can inc ease by 3-4 imes compa ed o no mal
ope a ions, wi h associa ed inc eases in disk usage o up o 40% [6]. Thei analysis showed ha epai ope a ions on a
1TB node wi h s anda d SSDs ook app oxima ely 6 hou s o comple e, while he same ope a ions on NVMe s o age
comple ed in jus 2.5 hou s, demons a ing he signi ican impac s o age selec ion has on main enance ope a ions.
The in e play be ween hese di e en I/O pa e ns c ea es unique challenges o s o age selec ion. While Cassand a can
handle up o 10,000 ope a ions pe second pe node wi h p ope s o age con igu a ion, inapp op ia e disk selec ion
can educe his by 50-80% du ing compac ion o epai ac i i ies [6]. This pe o mance a ia ion unde sco es why
unde s anding Cassand a's complex wo kload pa e ns is essen ial o selec ing s o age solu ions ha main ain
consis en pe o mance ac oss all da abase ope a ions, including bo h o eg ound clien eques s and backg ound
main enance p ocesses.
4. Disk p o iling: ca ego izing s o age op ions
Once IOPS is measu ed, disk p o iling helps ca ego ize a ailable s o age op ions, including NVMe, SSD, and gene al-
pu pose o p o isioned IOPS disks. Some disks suppo cascading o highe IOPS, bu inco ec con igu a ions may lead
o pe o mance deg ada ion. E alua ing sus ained s. max IOPS is c ucial in de e mining he op imal s o age
con igu a ion ailo ed o wo kload demands.
Comp ehensi e disk p o iling equi es sys ema ic e alua ion o s o age echnologies agains Cassand a's wo kload
pa e ns. Recen ad ancemen s in cloud in as uc u e ha e expanded a ailable op ions, pa icula ly in he NVMe space.
Acco ding o AWS, he la es EC2 R8g ins ances powe ed by G a i on4 p ocesso s ea u e up o 4TB o NVMe SSD
s o age wi h signi ican ly enhanced I/O capabili ies [7]. These ins ances deli e up o 1,000,000 IOPS wi h 60% be e
p ice/pe o mance compa ed o p e ious gene a ion ins ances. This pe o mance ie ep esen s a subs an ial
imp o emen o Cassand a wo kloads, whe e he combina ion o compu e powe and s o age pe o mance di ec ly
impac s o e all da abase esponsi eness. The local NVMe s o age on hese ins ances p o ides an impo an ad an age
o Cassand a deploymen s, as he di ec -a ached na u e elimina es ne wo k o e head ha exis s wi h ne wo k-
a ached s o age op ions.
When p o iling disks o Cassand a, unde s anding he ela ionship be ween s o age pe o mance and da abase
ope a ions is essen ial. Da aS ax documen a ion emphasizes ha p ope s o age selec ion mus accoun o Cassand a's
unique compac ion p ocesses and ead pa e ns [8]. Thei guidelines no e ha SSTable compac ion can consume
signi ican I/O esou ces, wi h majo compac ions po en ially using 50% o mo e o disk h oughpu . This ac i i y
di ec ly impac s o eg ound ope a ions, making sus ained IOPS capabili ies c i ical o consis en pe o mance. The
documen a ion u he ecommends con igu ing concu en _compac o s based on a ailable I/O capaci y, sugges ing
one compac o pe physical disk o RAID olume o maximize h oughpu wi hou o e whelming s o age esou ces.
Fo p o isioned IOPS solu ions, he con igu a ion o concu en _ eads and concu en _w i es se ings mus align wi h
a ailable disk pe o mance. Da aS ax ecommends se ing concu en _ eads o 16 imes he numbe o d i es and
concu en _w i es o 8 imes he numbe o co es when using SSDs [8]. This guidance e lec s he asymme ic
pe o mance cha ac e is ics o mos s o age sys ems, whe e ead and w i e ope a ions ha e di e en pe o mance
p o iles ha mus be balanced agains Cassand a's wo kload pa e ns.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2973-2978
2976
When e alua ing bu s s. sus ained pe o mance, disk p o iling should inco po a e ex ended load es ing ha mi o s
p oduc ion pa e ns. While many s o age solu ions o e imp essi e bu s capabili ies o sho du a ions, Cassand a's
con inuous ope a ion means sus ained pe o mance is ypically mo e ele an . This is pa icula ly impo an when
conside ing gene al-pu pose SSD op ions ha may h o le pe o mance a e deple ing I/O c edi s, po en ially causing
unexpec ed la ency spikes du ing c i ical ope a ions. P ope disk p o iling ensu es ha selec ed s o age solu ions
main ain consis en pe o mance ac oss all da abase ope a ions, including bo h clien eques s and main enance
p ocesses.
5. Benchma king Me hodology o nosql Pe o mance
Implemen ing a comp ehensi e benchma king me hodology is essen ial o accu a ely assessing disk pe o mance o
Cassand a wo kloads. This should include es ing ead/w i e a ios ha ma ch p oduc ion pa e ns, simula ing bo h
peak and sus ained wo kloads, and measu ing no jus h oughpu bu also la ency dis ibu ion pe cen iles (p95, p99)
which a e c i ical o use expe ience in la ency-sensi i e applica ions.
E ec i e benchma king o Cassand a deploymen s equi es ca e ul conside a ion o wo kload cha ac e is ics ha
accu a ely e lec p oduc ion en i onmen s. Resea ch by Tilmann Rabl e al. demons a es ha s anda d benchma king
app oaches o en ail o cap u e he complexi y o NoSQL da abase pe o mance unde eal-wo ld condi ions [9]. Thei
compa a i e s udy e alua ed six NoSQL da abases, including Cassand a, using he Yahoo! Cloud Se ing Benchma k
(YCSB) wi h a ied wo kloads. The esea ch e ealed ha Cassand a showed signi ican ly di e en pe o mance
cha ac e is ics ac oss wo kload ypes, achie ing 7,000-9,000 ope a ions pe second o upda e-hea y wo kloads bu
only 2,000-3,000 ope a ions pe second o ead-hea y scena ios on iden ical ha dwa e. This pe o mance a iance
highligh s why benchma king me hodologies mus inco po a e wo kload pa e ns ha ma ch in ended p oduc ion
usage a he han using gene ic es pa e ns.
The con igu a ion o benchma king pa ame e s subs an ially impac s measu emen accu acy. Acco ding o C is ina
E angelis a's comp ehensi e analysis o Cassand a pe o mance benchma king, p ope me hodology mus inco po a e
bo h sho - e m and long- e m es ing o cap u e pe o mance cha ac e is ics accu a ely [10]. Thei esea ch
demons a ed ha Cassand a's h oughpu ypically dec eased by 15-20% a e unning o 24 hou s compa ed o he
i s hou o ope a ion due o compac ion ac i i ies and memo y p essu e. Thei indings showed ha sho benchma ks
consis en ly o e es ima ed pe o mance, wi h 10-minu e es s showing a e age la encies o 1.3ms compa ed o 2.8ms
obse ed in ex ended 24-hou es s using iden ical ha dwa e and wo kloads.
Table 2 Cassand a Pe o mance Me ics Ac oss Di e en Wo kload Types and Tes Du a ions [9, 10]
Wo kload
Type
Ope a ions Pe
Second
A e age
La ency (ms)
p95 La ency
(ms)
p99 La ency
(ms)
Pe o mance A e 24
Hou s (% Dec ease)
Upda e-
hea y
7,000-9,000
5-15
15-60
40-150
15-20%
Read-hea y
2,000-3,000
5-15
15-60
40-150
15-20%
Measu emen o la ency dis ibu ions a he han a e ages p o ides c i ical insigh s o la ency-sensi i e applica ions.
Tilmann Rabl e al. ound ha while a e age la encies o Cassand a ope a ions ypically anged om 5-15ms, he 95 h
pe cen ile alues we e o en 3-4 imes highe , and 99 h pe cen ile measu emen s could be 8-10 imes highe han
a e ages [9]. Thei es ing demons a ed ha s o age I/O capabili ies di ec ly impac ed hese la ency dis ibu ions, wi h
highe -pe o mance s o age op ions educing he gap be ween a e age and ail la encies. This ela ionship be ween
s o age pe o mance and la ency dis ibu ions makes comp ehensi e benchma king essen ial o accu a e disk
selec ion.
The benchma king me hodology should also inco po a e a ying le els o concu ency o iden i y op imal
con igu a ions. C is ina E angelis a's esea ch iden i ied ha Cassand a's pe o mance scaling is no linea wi h h ead
coun , showing op imal h oughpu wi h 32-64 clien h eads in hei es en i onmen , wi h pe o mance declining
when h ead coun s exceeded 128 [10]. Thei wo k demons a ed ha di e en s o age con igu a ions exhibi ed
di e en concu ency cha ac e is ics, wi h NVMe-based solu ions main aining consis en la ency a highe h ead
coun s compa ed o adi ional SSDs, which showed mo e signi ican la ency deg ada ion as concu ency inc eased.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2973-2978
2977
6. Balancing Cos and Pe o mance: Op imiza ion S a egies
Cos op imiza ion is essen ial o balance pe o mance and budge . High-IOPS disks may be o e kill and cos ly, while
cheape al e na i es migh ail o mee la ency equi emen s. Benchma king disk op ions agains pla o m needs helps
in selec ing he mos e icien , scalable, and cos -e ec i e solu ion o Cassand a DB. S a egies include ie ed s o age
app oaches, whe e ho da a esides on high-pe o mance media while colde da a mo es o mo e economical op ions.
Finding he op imal balance be ween pe o mance and cos equi es sys ema ic e alua ion o s o age op ions agains
ac ual wo kload equi emen s. Facebook's enginee ing eam demons a ed his p inciple in hei la ge-scale dis ibu ed
caching in as uc u e, whe e s o age pe o mance di ec ly impac s o e all sys em e iciency [11]. Thei analysis
showed ha while highe -pe o mance lash s o age educed a e age ead la ency om 10ms o 0.5ms compa ed o
adi ional s o age, he pe -gigaby e cos inc eased by app oxima ely 15x. This cos di e en ial equi ed ca e ul
wo kload analysis o jus i y he in es men . By implemen ing a selec i e app oach whe e only he mos pe o mance-
c i ical da a was placed on lash s o age, hey educed o e all s o age cos s by 30% while main aining pe o mance
SLAs. The eam measu ed ac ual IOPS equi emen s du ing peak hou s (12pm o 2pm) and ound ha many sys ems
we e signi ican ly o e -p o isioned, wi h wo kloads equi ing only 25-30% o he p o isioned IOPS capaci y du ing
ypical ope a ion.
The p ope benchma king me hodology is c ucial o making in o med cos -op imiza ion decisions. Coope e al.'s wo k
on YCSB (Yahoo! Cloud Se ing Benchma k) p o ides a amewo k o e alua ing di e en s o age op ions o
Cassand a wo kloads [12]. Thei esea ch documen ed ha ead-hea y wo kloads (Wo kload B wi h 95% eads)
showed only 15-20% pe o mance imp o emen when upg ading om s anda d o high-pe o mance s o age, while
w i e-hea y wo kloads (Wo kload A wi h 50% eads/w i es) demons a ed up o 3x h oughpu imp o emen wi h he
same upg ade. These wo kload-speci ic pe o mance di e ences di ec ly impac cos -op imiza ion s a egies, as
o ganiza ions can a ge expensi e s o age in es men s only whe e hey deli e subs an ial bene i s.
The ie ed s o age app oach le e ages hese pe o mance cha ac e is ics o op imize cos s. YCSB es ing showed ha
when da a access pa e ns ollow a Zip ian dis ibu ion (as is common in many applica ions), jus 20% o he da a
ypically ecei es 80% o access ope a ions [12]. This access pa e n lends i sel o ie ed s o age app oaches whe e
equen ly accessed da a is placed on high-pe o mance media while less equen ly accessed da a uses mo e
economical op ions. The esea ch demons a ed ha p ope implemen a ion o a wo- ie s o age app oach could
educe s o age cos s by 40-60% compa ed o uni o m high-pe o mance s o age, while main aining la ency wi hin 10-
15% o he all-p emium con igu a ion.
When implemen ing cos op imiza ion s a egies, o ganiza ions mus conside no jus ha dwa e cos s bu also
ope a ional ac o s. Facebook's app oach included e alua ing he ope a ional o e head o di e en s o age
con igu a ions, inding ha simple s o age sys ems wi h sligh ly highe ha dwa e cos s o en esul ed in lowe o al
cos o owne ship due o educed main enance equi emen s and highe eliabili y [11]. Thei ope a ional da a showed
ha mo e complex s o age con igu a ions inc eased ope a ional inciden s by 2-3x, c ea ing hidden cos s ha o se
appa en ha dwa e sa ings.
Table 3 S o age Op imiza ion S a egies: Pe o mance Gains and Cos Sa ings [11, 12]
S o age S a egy
La ency Reduc ion
Rela i e Cos Pe GB
Pe o mance Imp o emen
T adi ional S o age
Baseline
1x
Baseline
High-Pe o mance Flash
95% (10ms o 0.5ms)
15x
Va ies by wo kload
Tie ed S o age (Two-
ie )
0-10% penal y s. all-
p emium
Mixed
Wi hin 10-15% o all-p emium
7. Conclusion
E ec i e disk selec ion o Cassand a deploymen s equi es a holis ic app oach ha add esses bo h pe o mance
equi emen s and economic cons ain s. By implemen ing s uc u ed me hodologies o IOPS measu emen ,
unde s anding he unique access pa e ns o dis ibu ed NoSQL a chi ec u es, ca ego izing a ailable s o age op ions
h ough comp ehensi e p o iling, conduc ing igo ous benchma king wi h app op ia e wo kload simula ion, and
applying a ge ed op imiza ion s a egies, o ganiza ions can c ea e da abase in as uc u es ha deli e consis en low-
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2973-2978
2978
la ency esponses while main aining budge discipline. The ie ed s o age app oach o e s pa icula ly p omising
esul s, allowing sys ems o le e age high-pe o mance media o equen ly accessed da a while u ilizing economical
op ions o colde da a. As digi al ecosys ems con inue o e ol e and scale, his balanced app oach o s o age
in as uc u e becomes inc easingly i al o main aining compe i i e ad an ages in applica ions whe e millisecond
esponse imes di ec ly impac use expe ience and business ou comes.
consis en low-la ency esponses while main aining budge discipline. This me hodical app oach o disk selec ion
p o ides a ounda ion o scalable, esponsi e Cassand a deploymen s ha can adap o g owing wo kloads while
a oiding unnecessa y in as uc u e cos s.
Re e ences
[1] A inash Lakshman and P ashan Malik, "Cassand a: a decen alized s uc u ed s o age sys em," ACM SIGOPS
Ope a ing Sys ems Re iew, 2010. [Online]. A ailable: h ps://dl.acm.o g/doi/10.1145/1773912.1773922
[2] E an S ille , "Ne lix Buil a Scalable Anno a ion Se ice Using Cassand a, Elas icsea ch and Icebe g," In oq, 2023.
[Online]. A ailable: h ps://www.in oq.com/news/2023/02/ne lix-anno a ions-
cassand a/#:~: ex =Ne lix%20s a ed%20wi h%20a%2012,da a%20size%20o %202.6TB.
[3] B endanG egg, "BPF Pe o mance Tools," B endanG egg, 2019. [Online]. A ailable:
h ps://www.b endang egg.com/bp -pe o mance- ools-book.h ml
[4] Ins aclus , "Node h oughpu Benchma king o Apache Cassand a® on AWS, Azu e and GCP," Ins aclus , 2024.
[Online]. A ailable: h ps://www.ins aclus .com/blog/node- h oughpu -benchma king- o -apache-cassand a/
[5] Ins aclus , "Apache Cassand a: Fea u es, a chi ec u e, and how o ge s a ed," Ins aclus , 2022. [Online].
A ailable: h ps://www.ins aclus .com/educa ion/apache-cassand a/apache-cassand a- ea u es-
a chi ec u e-and-how- o-ge -
s a ed/#:~: ex =Cassand a%20uses%20a%20dis ibu ed%20a chi ec u e,wo kloads%20e enly%20 o%20p
e en %20bo lenecks.
[6] Sandeep Malik, "A oid pi alls in scaling you Cassand a clus e : lessons and emedies," Walma Global Tech,
2017. [Online]. A ailable: h ps://medium.com/walma global ech/a oid-pi alls-in-scaling-you -cassand a-
clus e -lessons-and- emedies-a71ca01 8c04
[7] Micah Wal e , "New Amazon EC2 G a i on4-based ins ances wi h NVMe SSD s o age," AWS, 2025. [Online].
A ailable: h ps://aws.amazon.com/blogs/aws/new-amazon-ec2-g a i on4-based-ins ances-wi h-n me-ssd-
s o age/
[8] Da aS ax, "Tuning he da abase schema," Da aS ax, 2025. [Online]. A ailable:
h ps://docs.da as ax.com/en/planning/oss/schema- uning.h ml
[9] Tilmann Rabl e al., "Sol ing Big Da a Challenges o En e p ise Applica ion Pe o mance Managemen ,"
P oceedings o he VLDB Endowmen , 2012. [Online]. A ailable:
h ps:// ldb.o g/p ldb/ ol5/p1724_ ilmann abl_ ldb2012.pd
[10] C is ina E angelis a, "Pe o mance Modelling o NoSQL DBMS," P oceedings o he 2020 OMI Semina s Resea ch
T ends in Da a Cen e Ope a ions, Selec ed Topics in Da a Cen e Ope a ions, and Resea ch T ends in he In e ne
o Things (PROMIS 2020), 2020. [Online]. A ailable: h ps://d-nb.in o/1239737092/34
[11] Rajesh Nish ala e al., "Scaling Memcache a Facebook," 10 h USENIX Symposium on Ne wo ked Sys ems Design
and Implemen a ion (NSDI ’13), 2013. [Online]. A ailable:
h ps://www.usenix.o g/sys em/ iles/con e ence/nsdi13/nsdi13- inal170_upda e.pd
[12] B ian F. Coope e al., "Benchma king Cloud Se ing Sys ems wi h YCSB," cou ses.cs, 2010. [Online]. A ailable:
h ps://cou ses.cs.duke.edu/ all13/compsci590.4/838-CloudPape s/ycsb.pd
