An extensible lightweight framework for distributed telemetry of microservices

Con en s lis s a ailable a ScienceDi ec
Sus ainable Compu ing: In o ma ics and Sys ems
jou nal homepage: www.else ie .com/loca e/suscom
An ex ensible ligh weigh amewo k o dis ibu ed eleme y o
mic ose ices
Manuel O e o a, José Ma ía Ga cía b, Pablo Fe nandez b,∗
aUni e sidad de Se illa, Spain
bSCORE Lab, Uni e sidad de Se illa, Se ille, Spain
ARTICLE INFO
Keywo ds:
Teleme y
OAS
API
Mic ose ices
ABSTRACT
Mic ose ice a chi ec u es ha e become he s anda d o de eloping scalable dis ibu ed sys ems ha o e
signi ican bene i s in managing he in eg a ion and e olu ion o complex applica ions. Howe e , hey ace
challenges in e ec i ely diagnosing and esol ing pe o mance and eliabili y issues. T adi ional cen alized
eleme y models and cloud-based moni o ing pla o ms o en equi e complex o cos ly con igu a ions and a e
no op imized o REST ul mic ose ices. In ac , al hough he OpenAPI Speci ica ion (OAS) has become a key
s anda d o desc ibing mic ose ice APIs, exis ing eleme y ools do no le e age his in o ma ion o enhance
se ice analysis and diagnos ics. This pape in oduces a ligh weigh and dis ibu ed app oach o eleme y ha
uses OAS-based API in o ma ion, o e ing an au oma ed, con igu a ion- ee sys em ha enables de elope s and
ope a ions eams o pe o m oo cause analysis mo e e icien ly. Mo eo e , we p opose a plugin sys em o
inco po a e in elligen beha io in o he eleme y sys em, such as an adap i e p oac i e ale mechanism when
esponse- ime anomalies a e de ec ed. By inco po a ing his ex ensibili y mechanism, he amewo k pa es
he way o add ess issues such as ene gy consump ion and pe o mance, allowing he sys em o dynamically
adjus i s moni o ing ac i i ies o op imize esou ce usage and minimize he ca bon oo p in o mic ose ice
deploymen and execu ion. This adap abili y educes ope a ional o e head and suppo s sus ainable compu ing
p ac ices. To alida e ou app oach, we p esen a p oo -o -concep in he o m o a eady- o-use package o
he NodeJS ecosys em, demons a ing ha his dis ibu ed eleme y model can ope a e wi h minimal impac
on sys em pe o mance and esou ce usage, p o ing i s e ec i eness o suppo mo e obus and sus ainable
IT sys ems.
1. In oduc ion
The ise o mic ose ice a chi ec u es has es ablished hem as he
s anda d o building scalable and con inuous sys ems [1]. This ap-
p oach o e s conside able ad an ages when i comes o managing he
in eg a ion and e olu ion o mul iple componen s. Howe e , as wi h
any sophis ica ed a chi ec u e, issues such as pe o mance de ia ions
and componen ailu es can esul in cascading impac s h oughou he
sys em. In ac , despi e he e ec i eness o mic ose ice a chi ec u es
in add essing a ious conce ns, hey also in oduce signi ican hu dles,
pa icula ly in de ec ing and esol ing issues wi hin he sys em. These
challenges ha e been ho oughly examined in exis ing esea ch [2],
∗Co esponding au ho .
E-mail add esses: [email p o ec ed] (M. O e o), [email p o ec ed] (J.M. Ga cía), [email p o ec ed] (P. Fe nandez).
1h ps://p ome heus.io.
2h ps://elas ic.co.
3h ps://da adoghq.com.
4h ps://new elic.com.
5h ps://kube ne es.io.
wi h solu ions such as he ci cui b eake pa e n [3] being p oposed
o mi iga e he isks o ailu e.
An essen ial equi emen o e ec i ely managing mic ose ice a -
chi ec u es is he abili y o collec and analyze sys em da a, which
plays a key ole in iden i ying he oo causes o a chi ec u al p ob-
lems. T adi ionally, bo h academia and indus y ha e leaned owa d
a cen alized model o collec diagnos ic da a, whe e sys em compo-
nen s ansmi hei in o ma ion o a cen al hub o analysis. Popula
ools such as P ome heus1and ELK2 ollow his app oach. Fu he -
mo e, pla o m-as-a-se ice (PaaS) o e ings, including Da adog3and
NewRelic,4p o ide cloud-based solu ions o isualizing and analyzing
diagnos ics. In mo e complex en i onmen s, deploymen on con aine
o ches a ion pla o ms like Kube ne es5allows he implemen a ion
h ps://doi.o g/10.1016/j.suscom.2025.101100
Recei ed 14 Sep embe 2024; Recei ed in e ised o m 28 Decembe 2024; Accep ed 10 Feb ua y 2025
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
A ailable online 26 Feb ua y 2025
2210-5379/© 2025 The Au ho s. Published by Else ie Inc. This is an open access a icle unde he CC BY-NC license ( h p://c ea i ecommons.o g/licenses/by-
nc/4.0/ ).
M. O e o e al.
o Se ice Mesh models, which cen alize da a collec ion ac oss a -
ious se ices and o e he ools needed o managing elas ici y and
esilience wi hin. Howe e , hese cen alized app oaches ypically e-
qui e complex con igu a ions and ongoing main enance o incu addi-
ional de elopmen and main enance cos s. In addi ion, he lea ning
cu e is s eep and speci ic o each applica ion. Fu he mo e, he ype
o me ics collec ed is o en low-le el and no in eg a ed in o a REST-
ul pa adigm o he applica ions, necessi a ing cus om con igu a ions
o diagnosing API-based a chi ec u es. These issues a e p ominen in
solu ions ha use a cloud con inuum app oach, whe e a dis ibu ed
eleme y amewo k is c ucial o enable seamless moni o ing o cloud
con inuum scena ios [4] and, o his ex en , some app oaches ad o-
ca e p o iding o ches a ion amewo ks wi h eleme y agen s al eady
in eg a ed [5,6].
In ecen yea s, wi h he ad en o he Dis ibu ed Compu ing
Con inuum (DCC) pe spec i e [7], a dis ibu ed eleme y model could
also be a key elemen due o he inhe en complexi y and dis ibu ed
na u e o hese sys ems, which span ac oss a ious compu ing laye s
such as cloud, og, edge, and IoT. T adi ional cen alized eleme y
solu ions a e o en inadequa e o such en i onmen s because hey do
no e icien ly handle he la ency, bandwid h, and da a so e eign y
issues ha a ise when da a mus a e se mul iple ne wo k segmen s
and adminis a i e domains.
In his a icle, we p esen a ligh weigh decen alized eleme y
app oach ha aligns wi h he DCC pa adigm by enabling da a o be
p ocessed and analyzed close o i s poin o gene a ion. This p ox-
imi y minimizes la ency, educes ansmission cos s, and enhances he
sys em’s esponsi eness o de ec and ec i y issues in eal- ime. Fu -
he mo e, ollowing p inciples om Dis ibu ed Compu ing esea ch [7,
8], ou p oposed model suppo s a mo e lexible app oach o moni o -
ing and managing mic ose ices heal h. This ensu es ha ou eleme y
sys em can adap o he dynamic na u e o DCC en i onmen s, whe e
esou ces and se ices may luc ua e, scale, o mig a e ac oss di e -
en laye s depending on demand, connec i i y, o o he ope a ional
conside a ions [9]. Such adap abili y is a key ac o in main aining
sys em pe o mance, as highligh ed in he discussion on adap able
business models in DCC sys ems [10]. Mo eo e , his need o measu e
dis ibu ed pe o mance becomes e en mo e challenging and c i ical in
c oss-o ganiza ional scena ios, whe e in e dependencies be ween se -
ices p o ided by di e en o ganiza ions necessi a e a comp ehensi e
QoS en o cemen o a global se ice le el ag eemen (SLA) [11,12].
An addi ional mo i a ion o ou app oach lies in he g owing
need o lexible, in elligen eleme y solu ions capable o op imizing
esou ce usage and ene gy consump ion. In pa icula , in DCC en i on-
men s, whe e mic ose ices migh ope a e in esou ce-cons ained o
dynamic nodes, a igid eleme y model is ine icien . The e o e, we
in oduce a plugin-based ex ension mechanism ha allows he eleme-
y sys em o adjus i s beha io acco ding o he con ex . Fo example,
he sys em could sa e ene gy by educing he equency o moni o ing
du ing pe iods o low ac i i y while scaling up i s moni o ing p ecision
du ing peak loads o p e en pe o mance issues. This ene gy-e icien
and adap i e app oach suppo s sus ainable compu ing p ac ices in
mode n mic ose ice a chi ec u es.
F om a echnological poin o iew, i is also impo an o highligh
he ole o he OpenAPI Speci ica ion (OAS)6as a c i ical s anda d o
desc ibing APIs in mic ose ice a chi ec u es, pa icula ly in complex
DCC en i onmen s. OAS enables an explici model o he in e ace o
a mic ose ice, de ining ope a ions and expec ed inpu s and ou pu s.
Despi e his, cu en eleme y app oaches o en do no ake ull ad-
an age o he de ailed in o ma ion p o ided by OAS o imp o e he
analysis and epo ing o mic ose ice pe o mance.
To add ess his gap, we p opose a ligh weigh dis ibu ed eleme-
y managemen model ha akes ull ad an age o he OAS o de-
elop ools ha allow de elope s and ope a ions eams o analyze
6h ps://gi hub.com/OAI/OpenAPI-Speci ica ion.
oo causes o pe o mance issues o se ice dis up ions wi h minimal
con igu a ion. Using he OAS-de ined API s uc u e, ou sys em can
au oma ically in e ope a ional pa ame e s and gene a e de ailed e-
po s ailo ed o he beha io o indi idual mic ose ices. This design
elimina es he need o ex ensi e manual se up, o e ing de elope s a
anspa en , ully ope a ional sys em om he s a . The shi om a
cen alized eleme y model o a mo e dis ibu ed, de elope - iendly
app oach enables a as e and mo e e icien analysis o se ice heal h
wi hou he bu den o addi ional con igu a ion.
Speci ically, we p o ide a i s p oo o concep consis ing o a
eady- o-use package (OAS Teleme y) o he NodeJS ecosys em, one
o he mos popula echnology s acks o implemen ing mic ose -
ices. This package demons a es he p ac icali y o ou dis ibu ed
eleme y model and shows ha i can ope a e wi h minimal o e head
while deli e ing signi ican imp o emen s in lexibili y and in elligen
moni o ing. Mo eo e , om an ex ensibili y s andpoin , we p opose
a plugin model ha enables dynamic load o mo e complex moni o -
ing in elligence. Speci ically, we p esen his model wi h an adap i e
anomaly de ec ion plugin ha iden i ies abno mal esponse imes in
mic ose ice eques s. This plugin wo ks by analyzing his o ical da a o
es ablish dynamic h esholds o expec ed esponse imes and adjus ing
hese h esholds based on eal- ime obse a ions. When he sys em
de ec s esponse imes ou side hese accep able anges, an ale is
gene a ed, allowing apid in e en ion. The plugin-based a chi ec u e
ensu es ex ensibili y, allowing de elope s o in eg a e addi ional plug-
ins in he u u e o u he op imize ene gy consump ion, moni o ing
p ecision, and sys em esilience.
The es o he a icle is s uc u ed as ollows. Sec ion 2discusses e-
la ed wo k in he a ea compa ed wi h ou p oposal, which is desc ibed
in Sec ion 3. We alida e ou solu ion in h ee di e en scena ios
o empi ically assess i s pe o mance impac in Sec ion 4. Finally,
Sec ion 5p o ide ou concluding ema ks and iden i y a ious a eas
o u he esea ch on his opic.
2. Rela ed wo k
Con ollabili y has been iden i ied as a key issue in he con ex
o Mic ose ice A chi ec u es [13] and consequen ly moni o ing and
eleme y ha e gained a en ion in bo h he academy and indus y [14].
Speci ically, in he con ex o mic ose ice eleme y, se e al wo ks can
be ound ha p opose app oaches o ex ac eleme y da a o analyze
so wa e me ics [15,16]. Howe e , hey do no p o ide a conc e e
model o speci ica ion o applica ion-le el eleme y. In u n, o he
con ibu ions ocus on moni o ing sys em a chi ec u e using eleme-
y da a, de ining hei own model [17–19] o in eg a ing widely
used app oaches such as P ome heus o ELK [20,21]. By using hese
amewo ks, all he eleme y da a a e cen alized, depending on an ad-
di ional back-end o analyzing he da a, such as comme cial pla o ms
like NewRelic o Da adog, o Open-Sou ce al e na i es like Jaegge
o SigNoz. In u n, ou app oach le e ages he dis ibu ed na u e o
mic ose ice a chi ec u es o allow dis ibu ed ins umen a ion and
analysis o his ype o so wa e sys ems, while elying on s anda d
speci ica ions and models such as OpenTeleme y and OAS. Conce n-
ing moni o ing, he e a e also app oaches ha use OpenTeleme y
o in oduce agen s in o a se ice mesh o ins umen he compo-
nen s [22]. Howe e , hey do no allow o a ine-g ain con igu a ion o
he eleme y collec ion in a dis ibu ed manne . Recen ly, in [23] he
au ho s p oposed an in e es ing low-o e head, anspa en acing and
pe o mance debugging o Node.js-based mic ose ices, howe e , his
app oach s ill elies on a cen alized a chi ec u e o ga he he aces
and pe o m he analysis.
One o he key bene i s o a dis ibu ed eleme y app oach is i s
abili y o enable eal- ime adap a ion o bo h moni o ing s a egies and
sys em con igu a ions. By decen alizing da a collec ion and p ocessing,
sys ems can dynamically adjus hei moni o ing s a egies, such as
al e ing he equency o da a collec ion o e en econ igu ing sys em
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
2
M. O e o e al.
componen s o be e align wi h cu en condi ions. This capabili y is
pa icula ly aluable o mee ing se ice le el objec i es (SLOs) in com-
plex en i onmen s. Fo ins ance, as discussed in [24], ac i e in e ence
models can be applied o adjus sys em con igu a ions in esponse o
eal- ime da a, ensu ing ha he sys em ope a es e icien ly and mee s
pe o mance goals e en unde luc ua ing condi ions. This dynamic
adap a ion is mos e ec i e when eleme y da a is dis ibu ed and kep
close o he sou ce, allowing o immedia e ac ions and adjus men s.
3. OAS eleme y amewo k
OAS Teleme y is a amewo k designed o enhance obse abili y
wi hin Node.js applica ions by collec ing eleme y da a ha in eg a e
in o he OAS-Tools ecosys em, which is dedica ed o p o iding ools
and u ili ies o managing applica ions based on OpenAPI Speci ica-
ion (OAS), he s anda d o desc ibing APIs wi hin a mic ose ice
a chi ec u e. In such a con ex , he main aim o OAS Teleme y is o
p o ide a suppo ing ool o de elope s looking o gain insigh s in o
he ope a ion and pe o mance o hei OAS-based APIs, enabling hem
o moni o , debug, and op imize hei applica ions e ec i ely.
As a key co ne s one o ou p oposal, we can ind OpenTeleme y,7
which p o ides a eleme y amewo k whose aim is o enable obse -
abili y o a so wa e sys em, allowing ope a o s o analyze he in e nal
s a e o he sys em by examining hei ex e nal beha io . In o de o
do so, he sys em mus be ins umen ed, so ha eleme y da a ( aces,
me ics, and logs) can be gene a ed and la e analyzed by an obse abil-
i y backend, such as P ome heus. To his ex en , OpenTeleme y ocuses
on collec ing and managing he eleme y da a, p o iding a speci ica-
ion and s anda d p o ocol o eleme y managemen ha can be used
ega dless o he conc e e implemen a ion o he sys ems’ componen s.
OpenTeleme y p o ides a s anda dized SDK ha can ins umen aces,
me ics, and logs, al hough in he con ex o ou p oposal, p ima ily
ocuses on ace spans, which p o ide c ucial de ails such as s a us
codes and esponse imes. He e is a minimal example:
{"a ibu es": {
"h p.u l": "h p://localhos :8080/docs",
"h p.me hod": "GET",
"h p.s a us_code": 200
},
"_du a ion": [ 0, 2041000 ],
"_spanCon ex ": {
" aceId": "d 3...",
"spanId": "9e0e..."
}
Wi hin he NodeJS ecosys em, he p ima y aim o he OpenTeleme-
y SDK is o s eamline he debugging, p o iling, and pe o mance
moni o ing p ocesses o applica ions. In he con ex o mic ose ices,
de elope s ypically ely on amewo ks o sca old and de elop es -
ul APIs; in pa icula , one o he mos used amewo ks is Exp ess8
ha p o ides a ligh weigh web amewo k o Node.js ha simpli ies
HTTP eques handling and allows e icien ou ing and middlewa e
suppo , allowing de elope s o ocus on building obus web ap-
plica ions and APIs wi h ease. Speci ically, he Exp ess F amewo k
adop s a middlewa e-cen ic a chi ec u e ha s uc u es each eques
o he mic ose ice (o API) as a sequence o unc ion calls (i.e. he
middlewa es); his sequence acili a es he modula assembly o he
eques -handling pipeline, whe e each middlewa e unc ion can p ocess
he eques o esponse, o delega e o he nex unc ion in he s ack.
In such a con ex , he ole o middlewa e in Exp ess is o in e cep
he eques – esponse cycle, allowing ope a ions such as modi ying
7h ps://open eleme y.io/.
8h ps://exp essjs.com/.
Fig. 1. A chi ec u e diag am.
eques and esponse objec s, e mina ing he cycle, o passing con ol
o subsequen middlewa e. These unc ions enable c i ical capabili-
ies such as logging eques s, au hen ica ing use s, managing sessions,
and alida ing inpu . Middlewa e unc ions a e execu ed sequen ially,
p o iding a mechanism o decomposing he applica ion in o smalle ,
eusable componen s ha pe o m speci ic asks independen ly while
main aining he low o eques p ocessing.
Using his idea, he OAS Teleme y amewo k p o ides an Exp ess
middlewa e o ga he ing eleme y da a and managing he expo
p ocess. I is compa ible wi h ESM and CommonJS implemen a ions
and can be seamlessly in eg a ed in o exis ing Exp ess applica ions
wi h minimal se up equi ed. By simply including wo lines o code,
de elope s can enable eleme y unc ionali y wi hin hei applica ions:
impo oasTeleme y om ’@oas- ools/oas- eleme y’;
... // Ini ialize Exp ess app and load openapispec
app.use(oasTeleme y(openapispec))
... // Res o mic ose ice code
3.1. In e nal a chi ec u e
In he con ex o a mic ose ice, he OAS Teleme y package unc-
ions as middlewa e ha si s be ween he applica ion logic and he
incoming/ou going HTTP eques s. This middlewa e allows he appli-
ca ion o collec eleme y da a wi hou modi ying he o iginal code,
equi ing only he impo o he package and he addi ion o he
middlewa e.
As shown in Fig. 1, The OAS Teleme y middlewa e has 4 main
componen s: The con olle ini ializes he eleme y elemen s, hen he
managemen componen allows he de elope o s a /s op/ ese he
eleme y p ocessing. The pe sis ence componen allows da a e ching
om he DB, and he UI componen ende s he in o ma ion in a iew
designed o de elope s. Speci ically, we can ou line he componen s
as ollows:
•Con olle : This componen manages he in eg a ion o he Open-
Teleme y SDK by ini ializing he ace p o ide and con igu ing
he expo e and he span p ocesso . The span p ocesso handles
he p ocessing o spans and hei expo a ion o he con igu ed
expo e . The expo e is an ins ance o he InMemo yDBExpo e ,
a cus om implemen a ion o he OpenTeleme y amewo k.
•Managemen : This componen manages he span expo s. I p o-
ides he endpoin s o s op and s a he eleme y da a collec ion
dynamically a / eleme y/s op and / eleme y/s a
endpoin s, espec i ely. Despi e he ac ha eleme y ins umen-
a ion, once ini ialized, canno be s opped, he InMemo yDBEx-
po e enables he sys em o s op he inse ion o spans in o he
da abase. This helps mi iga e memo y consump ion. In addi ion,
a/ eleme y/ ese endpoin is p o ided, which clea s he
da abase om he spans.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
3
M. O e o e al.
Fig. 2. Global A chi ec u e compa a i e diag am be ween OAS- eleme y and adi-
ional eleme y solu ions.
•Pe sis ence: The pe sis ence componen is esponsible o man-
aging he NeDB da abase wi hin he InMemo yDBExpo e com-
ponen . This componen acili a es he s o age and e ie al o
eleme y da a gene a ed by he sys em. Speci ically, i p o ides
wo endpoin s: / eleme y/lis and / eleme y/ ind
(accessible ia he HTTP POST me hod).
The / eleme y/lis endpoin allows use s o e ie e a
lis o all eleme y span da a s o ed in he da abase. On he
o he hand, he / eleme y/ ind endpoin enables use s o
pe o m a ge ed sea ches o eleme y da a by submi ing NeDB
que ies in he eques body. This unc ionali y empowe s use s o
ex ac speci ic subse s o eleme y da a based on hei unique
equi emen s, enhancing he lexibili y and u ili y o he sys em’s
da a managemen capabili ies.
•UI: This componen gene a es a ligh weigh HTML page based
on an OAS ile pa h (see Figs. 3and 4) ha a e desc ibed in
Sec ion 3.4.
3.2. Global a chi ec u e compa a i e
In his sec ion, we explo e how OAS-Teleme y in eg a es wi hin a
mic ose ices in as uc u e and compa e i wi h adi ional eleme y
solu ions, such as Jaege and simila pla o ms. This compa ison will
help highligh he a chi ec u al di e ences in he way eleme y da a
is cap u ed, s o ed, and accessed ac oss bo h app oaches.
This is illus a ed in Fig. 2, which p esen s he wo in eg a ion
pa e ns side by side: on he le , we show OAS-Teleme y embedded
wi hin each mic ose ice, wi h local in-memo y s o age, and on he
igh a adi ional eleme y se up ha elies on a cen alized eleme y
pla o m. No e ha he Teleme y pla o m is a simpli ied iew o
all ex e nal se ices such as agen s, collec o s, and back-ends, whe e
eleme y da a a e agg ega ed, s o ed, and ul ima ely sen o a cen al
poin o analysis.
In a adi ional eleme y se up, such as he one depic ed on he
igh side o he igu e, he sys em is ypically designed a ound a cen-
alized a chi ec u e. This model elies on he deploymen o mul iple
componen s ha wo k oge he o collec , p ocess, and s o e eleme y
da a. In he igu e, hese componen s a e simpli ied and labeled as he
‘‘Teleme y Pla o m’’.
A he co e o his se up is he ins umen a ion wi hin each o he
mic ose ices, in eg a ed alongside he applica ion code. This ins u-
men a ion, ypically based on OpenTeleme y SDK, gene a es eleme y
da a, which can include aces, me ics, and logs ha p o ide insigh s
in o sys em pe o mance and beha io . Once gene a ed, hese da a a e
sen o he cen alized eleme y pla o m.
One challenge wi h adi ional se ups like Jaege is he need o
a ne wo ked in as uc u e whe e all mic ose ices mus connec o
a cen al collec ion poin . This o en esul s in highe ene gy con-
sump ion and ne wo k o e head, as eleme y da a mus a el ac oss
he ne wo k o each he cen alized back-end. Mo eo e , his is a
simpli ied iew; a adi ional eleme y p ocess ypically in ol es o he
se ices like collec o agen s ha agg ega e, il e , and o wa d eleme-
y da a; S o age backends, ha s o e he eleme y da a; and da a
isualiza ion se ices, which a e necessa y o p o ide use s wi h an
in e ace o que y, analyze, and isualize he eleme y da a. Each o
hese componen s (i.e. collec o agen s, s o age backends, and da a
isualiza ion se ices) can consume signi ican esou ces.
In con as , OAS-Teleme y in oduces a undamen ally di e en
app oach by embedding he eleme y amewo k di ec ly wi hin each
mic ose ice. As shown on he le side o he igu e, OAS-Teleme y
ope a es wi hin he same p ocess as he mic ose ice, s o ing da a
locally in in-memo y s o age (i.e., NeDB, a ligh weigh da abase based
on he NodeJS ecosys em). The key ad an age o his app oach is ha
eleme y da a a e s o ed locally o each mic ose ice, elimina ing he
need o sepa a e componen s, which could inc ease esou ce usage.
One o he key ad an ages o his amewo k is i s aul ole ance.
The in as uc u e does no need o be connec ed o a cen alized da a
collec ion sys em con inuously wi h he consequen isk o a single
poin o ailu e. Ins ead, all nodes unc ion independen ly (in line wi h
he mic ose ice pa adigm). While he ope a o has access o each
mic ose ice, hey can iew he eleme y da a di ec ly. Addi ionally,
hanks o he eleme y endpoin p o ided by OAS-Teleme y, he e is
he possibili y o accessing eleme y da a om o he in e connec ed
mic ose ice nodes, e en i hese nodes a e no ex e nally connec ed.
This design o e s u he bene i s. By s o ing da a locally in mem-
o y, he amewo k educes ansmission cos s, as he e is no need
o ne wo k communica ion o ansmi eleme y da a o an ex e nal
se ice.
3.3. Se up
OAS Teleme y amewo k is a ailable in he global npm egis y
and can be easily ins alled in he p ojec wi h he s anda d package
manage npm in NodeJS:
npm ins all @oas- ools/oas- eleme y
Once ins alled, he only necessa y addi ion o he mic ose ice code
is a single line. Fo CommonJS ype modules:
le oasTeleme y = equi e(’@oas- ools/oas- eleme y’);
o he ollowing equi alen line o ES6 Modules:
impo oasTeleme y om ’@oas- ools/oas- eleme y’;
3.4. Teleme y UI
The eleme y use in e ace (UI) is accessed wi hin he Exp ess
applica ion by na iga ing o he / eleme y endpoin , p o iding a
con enien in e ace o iewing eleme y da a and moni o ing appli-
ca ion pe o mance (see Figs. 3and 4). Speci ically, he UI p esen s a
able ha displays eleme y da a in a simple o ma , p o iding key
de ails such as he numbe o spans pe pa h and he a e age du a ion.
A he op o he page, use s can access a panel wi h h ee main
unc ions: oggling eleme y da a collec ion ( o educe memo y usage),
ese ing eleme y da a, and s opping able upda es o inspec ion.
Addi ionally, o each endpoin , use s can manually e esh he da a
by clicking a bu on, which e ie es he la es da a om he da abase
and upda es he able. Fo deepe analysis, use s can na iga e o he
‘‘De ails’’ sec ion by clicking on an endpoin name, whe e a lis o all
spans is displayed o u he inspec ion.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
4
M. O e o e al.
Fig. 3. UI o OAS eleme y.
Fig. 4. UI o aces de ail.
3.5. Ex ensibili y
The ex ensibili y o he p oposed eleme y amewo k is a c i ical
aspec ha enables i o dynamically adap o he e ol ing needs
o dis ibu ed mic ose ice en i onmen s. Th ough a plugin sys em,
de elope s can ex end i s co e unc ionali y wi hou modi ying he
unde lying sys em, enhancing i s lexibili y and capabili y o pe o m
a wide ange o asks. In his sec ion, we desc ibe he ex ensibili y a -
chi ec u e, ocusing on he plugin amewo k, and explo e i s po en ial
applica ions o ad anced eleme y scena ios.
One o he key ea u es o he p oposed amewo k is he abili y
o dynamically load and con igu e plugins h ough simple HTTP e-
ques s o he eleme y API. This allows o eal- ime adap a ion and
cus omiza ion wi hou he need o es a o econ igu e he en i e
eleme y s ack. As illus a ed in Fig. 5, a ypical scena io un olds as
ollows:
1. A plugin is eques ed and loaded ia an HTTP eques o he
eleme y API.
2. The eleme y middlewa e dynamically inco po a es he plugin
in o i s p ocessing pipeline, allowing he plugin o in e cep and
analyze all aces as hey low h ough he sys em.
3. The example plugin ( o anomaly de ec ion9) collec s esponse
ime aces om a ious mic ose ice endpoin s. I de ines a
beha io al window o accep able esponse imes based on his o -
ical da a and au oma ically aises ale s in a Teleg am channel
i i de ec s an abno mal esponse ime. In his plugin example,
a simple ou lie mechanism is de eloped, bu mo e complex
9h ps://bi .ly/o -anomaly-ale -plugin.
echniques o anomaly de ec ion [25] could be in eg a ed in he
same ashion.
This example showcases he ligh weigh and con igu able na u e o
he amewo k. By allowing ex e nal plugins o de ine and implemen
cus om beha io s, he sys em can p o ide mo e ailo ed and p oac i e
insigh s in o mic ose ice pe o mance.
To ensu e consis ency and in e ope abili y, each plug-in wi hin he
eleme y amewo k adhe es o a s anda dized s uc u e. A plugin is
encapsula ed in a single Ja aSc ip ile ha implemen s an in e ace
wi h wo key me hods: (i) he load me hod, which is in oked when
he plugin is i s loaded, ini ializes he plugin and con igu es any
pa ame e s needed o pe o m i s unc ion. Fo example, he plugin may
de ine h eshold pa ame e s o accep able esponse imes o speci y
how ale s should be gene a ed and sen ; and (ii) newT ace, which
is called each ime a new ace is p ocessed by he eleme y sys em.
The plugin can analyze he ace, apply any necessa y compu a ions,
and ake app op ia e ac ions (such as logging da a, igge ing ale s, o
adjus ing moni o ing h esholds).
The plugin s uc u e enables a di ec in eg a ion in o he eleme-
y amewo k while main aining lexibili y o de elope s o imple-
men cus om logic and beha io s. This design also ensu es ha he
plugins can be eused in mul iple mic ose ices and adap ed as he
equi emen s o he eleme y e ol e.
The plugin sys em enables se e al ad anced eleme y ex ensions
ha go beyond s anda d moni o ing and ale ing. In he ollowing,
we discuss h ee po en ial key di ec ions in which he ex ensibili y
mechanism can be used o enhance sys em analysis and pe o mance
op imiza ion.
•Dynamic Me ic Compu a ion: In dynamic mic ose ice en i-
onmen s, whe e sys em beha io can change apidly, he abili y
o compu e mo e sophis ica ed me ics on demand is essen ial.
Fo example, when pe o mance de ia es om he no m, he
eleme y sys em can ac i a e plugins ha use ad anced ech-
niques such as polling o s a is ical analysis o ga he mo e de-
ailed me ics o e ime. This allows o a mo e g anula un-
de s anding o pe o mance bo lenecks, especially du ing asyn-
ch onous o high-la ency ope a ions, and imp o es he p ecision
o da a collec ion du ing c i ical pe iods.
•Cascading Ale s and P eemp i e Measu es: The plugin ame-
wo k suppo s he abili y o issue cascading ale s ac oss mul-
iple mic ose ices when ea ly signs o sys em deg ada ion a e
de ec ed. Fo example, i a plugin iden i ies a slowdown in a
c i ical ups eam se ice, i could igge ale s o downs eam
se ices, p omp ing hem o ini ia e allback mechanisms o ci -
cui b eake s o p e en a po en ial sys em-wide ailu e. This
p eemp i e ale ing s a egy enhances sys em esilience and helps
mi iga e he impac o ailu es be o e hey sp ead ac oss he en i e
in as uc u e.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
5

M. O e o e al.
Fig. 5. Teleme y plugin showcase.
•Con igu able T ace Feeds: Ano he a ea o u u e de elopmen
is he dynamic con igu a ion o ace eeds o cen alized eposi o-
ies such as OpenTeleme y collec o s. By inco po a ing eal- ime
analysis in o he eleme y amewo k, plugins could selec i ely
o wa d only he mos ele an aces o hese eposi o ies o
deepe analysis. This ea u e would educe he load on s o -
age and p ocessing esou ces while ensu ing ha c i ical eleme-
y da a is a ailable when needed o pe o mance uning o
oubleshoo ing. Plugins could also adap hei beha io based
on eal- ime condi ions, u he op imizing esou ce usage and
suppo ing e icien da a collec ion s a egies.
Th ough hese applica ions, he ex ensibili y capabili ies o he
p oposed amewo k open up new possibili ies o in elligen , adap i e
eleme y in dis ibu ed mic ose ice en i onmen s.
4. Valida ion
Gi en he c i ical ole o pe o mance in eleme y, we conduc ed a
s udy o e alua e he impac o he OAS Teleme y package on Exp ess
mic ose ice pe o mance. We measu ed memo y usage and eques
p ocessing ime in bo h a sample API mic ose ice and a p oduc ion
en i onmen , wi h and wi hou he eleme y middlewa e enabled. In
addi ion, we compa ed ou package wi h an open sou ce indus ial
eleme y solu ion wi h espec o he same me ics.
All es iles, sample applica ions, and esul s a e a ailable in he
es olde o he OAS Teleme y eposi o y.10 The es iles include
10 h ps://gi hub.com/oas- ools/oas- eleme y.
a sample Exp ess mic ose ice (ks-api) and a es sc ip ( es .js) ha
measu es he pe o mance impac o he OAS Teleme y package.
Fu he mo e, due o he complexi y o alida ing in a p oduc ion
en i onmen , whe e eques s need o be spaced ou o a oid se e
sa u a ion, and conside ing ha he se e may no accep eques s o
se e al seconds due o con igu a ion asks, he es s we e conduc ed in
an ad hoc es ing package. Speci ically, hese alida ion es s we e pe -
o med in he con ex o he p oduc ion en i onmen Bluejay11 which
consis s o a mic ose ice a chi ec u e ha p o ides a de elopmen
eam p ac ice audi pla o m [26]. Among he di e en componen s o
he Bluejay a chi ec u e, alida ion es s we e de eloped in i s mos
c i ical mic ose ice called Regis y.
4.1. Me hodology
The ollowing subsec ions desc ibe he me hodology employed o
alida e he pe o mance and e ec i eness o OAS Teleme y ac oss
h ee dis inc scena ios. These include es ing he pe o mance impac
o OAS Teleme y on a syn he ic mic ose ice applica ion, e alua ing
i s beha io wi hin a eal-wo ld p oduc ion en i onmen , and compa -
ing he pe o mance o ou eleme y solu ion agains Jaege , a widely
adop ed eleme y pla o m c ea ed by Ube Technologies.
Fo each alida ion scena io, we compa e he pe o mance o he in-
s umen ed e sion o he applica ion, which in eg a es OAS Teleme y,
wi h he o iginal unins umen ed e sion. This compa ison is pe -
o med h ough a se ies o es cases (TC) designed o assess he impac
on a ious pe o mance me ics, pa icula ly ocusing on esponse ime,
11 h ps://docs.bluejay.go e ni y.io/.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
6
M. O e o e al.
memo y usage, and CPU usage. The es s a e designed o encompass
di e en o de s o magni ude in e ms o esponse imes when ex-
ecu ing he p oblems, e lec ing he di e si y o use cases and he
ange o la encies ypically encoun e ed in mic ose ice a chi ec u es.
By e alua ing pe o mance ac oss hese a ied scena ios, we aim o
demons a e he obus ness o ou amewo k and i s abili y o handle
a wide spec um o wo kload in ensi ies.
4.1.1. Syn he ic mic ose ice eleme y pe o mance alida ion
To measu e he pe o mance impac , we use a modi ied e sion
o KS-API, an Exp ess.js app ha simula es a Knapsack p oblem sol -
ing API. The applica ion includes a simple API wi h wo endpoin s:
/api/ 1/s ess and /api/ 1/p oblems. The /api/ 1/
s ess endpoin gene a es and sol es a knapsack p oblem wi h a
speci ied numbe o i ems and a maximum weigh .
In o de o measu e he memo y usage, he index.js ile o he ks-api
applica ion was modi ied o use he s anda d NodeJS module ‘‘ 8’’ ha
p o ides heapS a s logging; speci ically, i p o ides in o ma ion abou
he memo y usage o he Node.js applica ion, including he o al heap
size, used heap size, and heap size limi . In addi ion, a new e sion o
he index.js ile was c ea ed o include he OAS Teleme y middlewa e
(including all code in he o iginal index.js ile plus he OAS Teleme-
y middlewa e); his duali y allows he po en ial execu ion o wo
ins ances o he mic ose ice: one wi h eleme y (TLM) and ano he
wi hou eleme y (NoTLM). Mo eo e , o es di e en scena ios, we
a y he API load in di e en o de s o magni ude in esponse ime
o obse e i s impac on pe o mance wi hin he KS-API. Wi h hese
a iabili y poin s, we ha e he ollowing wo con igu a ion dimensions:
(1) Teleme y s. No Teleme y wi h he ollowing op ions:
•TLM: This con igu a ion in ol es using he OAS Teleme y mid-
dlewa e in he KS-API calling indexTeleme y.js.
•NoTLM: This con igu a ion excludes he OAS Teleme y middle-
wa e calling index.js.
(2) API Loads wi hin he KS-API wi h he ollowing op ions. The e
will be 3 di e en o de s o magni ude in he p oblem size sen o
ks-api:
•ONES (ones o milliseconds): This load in ol es sending he
p oblem sizes o he /api/ 1/s ess endpoin , esul ing in
esponse imes anging om 1 o 10 ms. P e ious es s ha e de-
e mined he p oblem size o his ca ego y as 13. Consequen ly,
calling he /api/ 1/s ess/13/13 endpoin , ep esen ing a
p oblem size o 13, will esul in a esponse ime be ween 1 and
10 ms on he es ing machine.
•TENS ( ens o milliseconds): He e, p oblem sizes sen o
/api/ 1/s ess yield esponse imes be ween 10 and 100 ms.
P e ious es s ha e de e mined he p oblem size o his ca ego y
as 6000 (/api/ 1/s ess/6000/6000).
•HUNDREDS (hund eds o milliseconds): P oblem sizes sen o
/api/ 1/s ess lead o esponse imes anging om 100 o
1000 ms. P e ious es s ha e de e mined he p oblem size o his
ca ego y as 80000 (/api/ 1/s ess/80000/80000).
In addi ion, o he pa ame e s a e kep cons an du ing he es ing
p ocess o ensu e ha esul s can be compa ed accu a ely. These pa-
ame e s include se ing concu en Use s o 1 o ocus solely on one
eques a a ime, main aining a ixed coun o 200 o eques s o
ensu e wo kload consis ency, and egula ing he pace o eques gene -
a ion wi h a s eady eques Delay o 100 ms. The eason o using only
one concu en use is oo ed in NodeJS single- h eaded a chi ec u e.
Since Node.js p ocesses eques s in a single h ead, in oducing mo e
han one concu en use would esul in eques s being queued. This
queueing could in oduce a iabili y in esponse ime and pe o mance
measu emen s, hus obscu ing he impac o changes o he applica ion
o en i onmen unde es .
Based on hese dimensions o a iabili y, we measu e he impac on
pe o mance unde di e en condi ions and du a ions by de ining he
ollowing speci ic es cases:
•TC-01 — Sho du a ion wi h di e en loads: This es aims o
measu e he memo y usage and p ocessing ime o eques s wi h
and wi hou he OAS Teleme y middlewa e enabled. I includes:
–15 i e a ions wi h eleme y enabled (TLM) (5 pe p oblem
size o de o magni ude: ONES, TENS, HUNDREDS).
–15 i e a ions wi h eleme y disabled (NO_TLM) (5 pe p ob-
lem size o de o magni ude: ONES, TENS, HUNDREDS).
•TC-02 — In e mi en eleme y: This es measu es he memo y
usage and p ocessing ime o eques s when he OAS Teleme y
middlewa e is unning, hen s opped, and inally s a ed again. I
in ol es:
–15 i e a ions (5 o each o de o magni ude: ONES, TENS,
HUNDREDS). Fo each i e a ion, he ollowing phases a e
execu ed:
1. Phase I: 66 eques s wi h eleme y enabled (TLM).
2. Reques o s op eleme y (spans a e no longe ex-
po ed).
3. Phase II: 66 eques s wi h eleme y s opped.
4. Reques o s a eleme y again.
5. Phase III: 66 eques s wi h eleme y s a ed again
(TLM).
•TC-03 — Long unning wi h di e en loads: This es e alua es
he memo y usage and p ocessing ime o eques s wi h and
wi hou OAS Teleme y middlewa e enabled o a longe du a ion
o 30 min. I includes:
–15 i e a ions o eleme y enabled (TLM) o 30 min (5 o
each o de o magni ude: ONES, TENS, HUNDREDS).
–15 i e a ions o eleme y disabled (NO_TLM) o 30 min (5
o each o de o magni ude: ONES, TENS, HUNDREDS).
4.1.2. Realis ic mic ose ice eleme y pe o mance alida ion
To measu e he impac o pe o mance in a p oduc ion en i onmen ,
we use he egis y mic ose ice, a c i ical componen o he Blue-
jay a chi ec u e [27] esponsible o he compu a ion and s o age o
Team P ac ice Ag eemen s (TPA). The egis y mic ose ice in e -
ac s wi h he collec o mic ose ice, which ga he s ele an pieces
o e idence om Gi Hub and sends hem back o he egis y so
i can e alua e whe he each eam membe is ollowing he p ac ices
desc ibed in he TPA.
A TPA consis s o me ics and gua an ees. In he con ex o Bluejay,
i measu es de elope s’ adhe ence o good p ac ices. Fo example, a
de elopmen eam p ac ice (o TP) may analyze he numbe o pull
eques s e iewed by a eam membe , wi h he co esponding gua an ee
ha he use mus e iew a leas ou pull eques s pe mon h. See12
o examples o TPAs.
Fo he pu pose o es ing, we c ea ed a speci ic e sion o he
mic ose ice which can be pulled using he ollowing command:
docke pull go e ni y/ egis y:en - eleme y
This e sion includes a eleme y package ha can be ac i a ed o
deac i a ed based on he alue o he en i onmen a iable
REGISTRY_TELEMETRY. When he a iable is se o ue, eleme y
12 h ps://gi hub.com/go e ni y/zoo/ ee/main/bluejay/ pa.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
7
M. O e o e al.
is enabled, allowing us o moni o and collec ele an pe o mance
da a. In con as , se ing i o alse disables eleme y, acili a ing
he baseline pe o mance es wi hou he addi ional o e head o
moni o ing.
In o de o imp o e con ol o e execu ion imes and eques man-
agemen , an ad hoc module,13 was de eloped o es ing. This module
pe o ms asks simila o hose pe o med o es ing he ks-api
applica ion bu o e s be e con ol o e execu ion imes and he a e
a which eques s a e sen .
The package includes a dedica ed module ha manages he execu-
ion o eques s using he apipecke ool. Addi ionally, he memo y
and CPU usage o he egis y mic ose ice a e measu ed by a
sepa a e module ha in e aces wi h Docke ’s S a s API.
The combina ion o hese modules allows o p ecise pe o mance
moni o ing in p oduc ion en i onmen s while ensu ing eques s a e
spaced app op ia ely o a oid o e loading he se e .
Using he modules desc ibed abo e, we can e ec i ely moni o how
he sys em beha es while calcula ing TPA compliance. The complexi y
o his calcula ion depends on se e al ac o s:
1. The numbe o gua an ees wi hin he TPA.
2. The numbe o me ics pe gua an ee.
3. The numbe o use s included in he ag eemen ( ypically 1 o 5
use s).
4. The window ype: Gua an ees can be calcula ed o a speci ic
poin based on da a om he p e ious hou , day, week, o
mon h.
5. The pe iod size: The ange (s a - o-end) o e which all da a
poin s a e calcula ed, de e mined by he window ype (hou ly,
daily, weekly, o mon hly).
To ensu e accu a e compa isons ac oss all es s, we main ain ce ain
cons an s: The TPA will always in ol e one use , wi h no a ia ions in
he gua an ees, me ics, o window ype.
To e alua e he impac on pe o mance, we will pe o m es s
ha a y he pe iod size, di ec ly a ec ing he ime equi ed o TPA
calcula ions. Speci ically, we will es pe iod du a ions in hou s (1, 24,
and 48). These a ia ions will allow us o obse e pe o mance impac s
in he Bluejay en i onmen . F om hese poin s o a iabili y, we de ine
wo con igu a ion dimensions:
(1) Teleme y in Applica ion s. No Teleme y, wi h he ollow-
ing op ions:
•Teleme y Enabled (TLM): This con igu a ion enables eleme y
wi hin he egis y mic ose ice o moni o he pe o mance
o TPA calcula ions.
•Teleme y Disabled (NoTLM): This con igu a ion disables
eleme y, measu ing baseline pe o mance wi hou any moni o -
ing o e head.
(2) TPA Pe iod Du a ions, wi h he ollowing op ions: We will es
h ee di e en pe iod du a ions, co esponding o di e en o de s o
magni ude in esponse ime:
•SMALL (1 h): This load calcula es TPA compliance o a 1-h
pe iod. P e ious es s show esponse imes anging om 1 o 3 s.
•MEDIUM (24 h): The pe iod du a ion is se o 1 day, wi h
expec ed esponse imes be ween 4 and 6 s.
•LARGE (48 h): The pe iod du a ion ex ends o 2 days, gene ally
esul ing in esponse imes be ween 6 and 10 s.
13 h ps://gi hub.com/oas- ools/oas- eleme y- es e .
The egis y canno handle pa allel eques s, and he
collec o se ice s ops accep ing eques s e e y minu e be ween
seconds hh:mm:50 and hh:mm+1:05, wi h a ma gin o e o . The e-
o e, all es eques s will be sen sequen ially wi hin he a ailable 45 s
(s a ing a hh:mm:05). The delay be ween eques s will be calcula ed
as:
delay =⌊45000
es ima ed-max- esponse- ime⌋
Fo each o de o magni ude:
•SMALL (1 h): Assuming a max esponse ime o 3 s, he delay
would be loo (45/3) = loo (15.0), allowing 15 eques s
wi hin he minu e/i e a ion.
•MEDIUM (24 h): Wi h a max esponse ime o 6 s, he delay
would be loo (45/6) = loo (7.5), allowing 7 eques s.
•LARGE (48 h): Assuming a max esponse ime o 10 s, he delay
would be loo (45/10) = loo (4.5), allowing 4 eques s.
These es s will help de e mine he sys em’s pe o mance unde
di e en load condi ions wi hin he Bluejay a chi ec u e. Based on
hese con igu a ion dimensions, we de ine he ollowing speci ic es
cases, simila o he ks-api ones:
•TC-01 — Sho du a ion wi h di e en loads: This es mea-
su es esponse imes, Docke CPU, and memo y usage pe cen age
o a ian s wi h eleme y enabled (TLM) and eleme y disabled
(NoTLM) o each o de o magni ude. I in ol es:
–9 i e a ions wi h eleme y enabled (TLM) (3 pe o de o
magni ude: SMALL, MEDIUM, LARGE).
–9 i e a ions wi h eleme y disabled (NoTLM) (3 pe o de
o magni ude: SMALL, MEDIUM, LARGE).
•TC-02 — In e mi en eleme y: This es measu es Docke CPU
and memo y usage when eleme y is i s unning, hen s opped,
and inally es a ed. I in ol es:
–9 i e a ions o 3 min (3 o each pe iod du a ion: SMALL,
MEDIUM, LARGE). Fo each i e a ion, he ollowing phases
a e execu ed:
1. Phase I: 1 min sending eques s wi h eleme y en-
abled (TLM).
2. Reques o s op eleme y (spans a e no longe ex-
po ed).
3. Phase II: 1 min sending eques s wi h eleme y
s opped.
4. Reques o s a eleme y again.
5. Phase III: 1 min sending eques s wi h eleme y
es a ed (TLM).
•TC-03 — Long unning wi h di e en loads: This es e alua es
docke CPU and memo y usage o e a longe du a ion wi h (TLM)
and wi hou (NoTLM) eleme y enabled. I includes:
–3 i e a ions o 28–30 min eleme y enabled (TLM) (1 o
each pe iod du a ion: SMALL, MEDIUM, LARGE).
–3 i e a ions o 28–30 min eleme y disabled (NoTLM) (1 o
each pe iod du a ion: SMALL, MEDIUM, LARGE).
4.1.3. Jaege pe o mance compa a i e
In addi ion o he alida ion cases desc ibed abo e, we include an
addi ional expe imen o compa e he pe o mance o ou solu ion wi h
a widely used eleme y pla o m in he indus y, namely Jaege .14
14 h ps://www.jaege acing.io/.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
8
M. O e o e al.
Jaege is an open sou ce p ojec unde he Cloud Na i e Compu ing
Founda ion ha p o ides a eleme y pla o m o dis ibu ed sys ems
wi h a special ocus on mic ose ice a chi ec u es. I also uses he
Open Teleme y s anda d unde he hood, hough aces a e cen ally
collec ed and analyzed. Thus, we design a alida ion case o compa e
he pe o mance o ou o ally dis ibu ed eleme y app oach wi h an
exis ing eleme y pla o m whe e some componen s a e cen alized.
In o de o ca y ou his compa ison, we eplica ed he TC-01
es case by adding ano he 15 i e a ions whe e we included he
Jaege ins umen a ion in he ks-api ins ead o OAS-Teleme y, calling
indexJaege .js. In addi ion, Jaege needs a collec o componen
o manage he expo ed eleme y.
The Jaege ‘‘all-in-one’’ sys em p o ides a con enien , single-
con aine deploymen ha includes se e al key componen s necessa y
o acing and moni o ing: he aces collec o , which ecei es ace
da a om he expo e and w i es i in o he s o age backend, and he
que y se ice, which enables he que ying and isualiza ion o ace
da a. To be ai and ensu e consis ency wi h ou OAS-Teleme y es s,
he Jaege sys em u ilized in-memo y s o age ins ead o an ex e nal
s o age back-end, mi o ing he a chi ec u e o he OAS-Teleme y
implemen a ion.
To acili a e hese es s, we c ea ed a lexible Docke image wi hin
he es olde o he OAS-Teleme y Gi Hub eposi o y. This image
includes mul iple iles essen ial o compa ing pe o mance ac oss di -
e en eleme y implemen a ions. Each ile se es a speci ic pu pose in
he es ing amewo k:
1. indexSelec o .js: A dynamic implemen a ion selec o con-
olled by an en i onmen a iable. This ile enables seamless
swi ching be ween di e en eleme y implemen a ions du ing
di e en es execu ions, ensu ing ha all es s a e conduc ed
unde consis en condi ions.
2. index.js: The o iginal implemen a ion o he p e iously
es ed ks-api, used as a baseline o compa ison.
3. indexTeleme y.js: An enhanced e sion o ks-api in-
s umen ed wi h he OAS-Teleme y expo e , impo ed di ec ly
om he npm package. This implemen a ion uses in-memo y
s o age o ace da a, p o iding a ligh weigh eleme y solu-
ion.
4. indexJaege .js: A e sion o he ks-api in eg a ed wi h
he Jaege OpenTeleme y expo e . This implemen a ion ex-
po s he aces o he Jaege ‘‘all-in-one’’ se ice.
This se up allows o de ailed compa isons o esponse imes, e-
sou ce usage, and o e all sys em pe o mance ac oss he o iginal im-
plemen a ion, he OAS-Teleme y-enhanced e sion, and he Jaege -
ins umen ed e sion, all wi hin a con olled and ep oducible en i on-
men .
Wi h espec o ene gy e iciency, we es ima e he educ ions in
ene gy consump ion based on Jaege ’s es CPU usage esul s. By
analyzing bo h he CPU consump ion o TC-01 and he CPU speci i-
ca ions, pa icula ly he The mal Design Powe (TDP), we calcula ed
he expec ed ene gy sa ings in e ms o CPU powe inpu . In such a
con ex , i is impo an o highligh ha his analysis ep esen s an
ini ial baseline o ene gy sa ings ha lea es ou impo an elemen s
such as I/O ope a ions on s o age de ices o Ne wo k Usage, which
could poin o e en bigge sa ings.
4.1.4. Scena io es ing p ocess
Fo each o he p e ious h ee scena ios, we ca y ou he same
es ing p ocess desc ibed in he ollowing.
Se e ini ializa ion. Ini ially, he se e (TLM o NO_TLM depending on
he con igu a ion) is ini ia ed o c ea e he en i onmen o conduc ing
he es s. Then a single eques is made o e i y he se e connec ion.
This s ep ensu es ha he sys em is eady o p ocess eques s. In si ua-
ions ha equi e complemen a y eleme y se ices, such as he Jaege
compa ison, addi ional componen s — like he Jaege ‘‘all-in-one’’
se ice — a e also ini ialized.
Fig. 6. TC-01: Rela i e change in esponse ime.
Tes ins ance execu ion. Du ing each es ins ance o a Tes Case, spe-
ci ic p ocedu es a e unde aken o cap u e and analyze ele an da a
poin s:
•Measu emen o Node heap s a is ics: The i s s ep in ol es
measu ing Node.js heap s a is ics (heapS a s) o assess he sys-
em’s memo y usage.
•Measu emen o Docke CPU and Memo y: Fo Docke con ain-
e s, such as he Bluejay egis y mic ose ice, he Docke S a s API
is used o eco d CPU and memo y usage o he con aine .
•Execu ion o Reques s: Following he heapS a s measu emen ,
he es ins ance u ilizes he apipecke ool15 o gene a e a p ede-
e mined numbe o eques s di ec ed owa ds he URL whe e he
ks-api applica ion is ac i ely unning.
•Reassessmen o Node heap s a is ics: A e comple ing he e-
ques cycle, he es ins ance measu es he Node.js heap s a is ics
again o ack any changes in memo y usage.
•Da a logging: The collec ed da a, including memo y be o e, e-
sponse s a is ics, memo y a e , CPU usage, and Docke memo y
usage, a e logged in o a CSV ile o u he analysis and e iew.
Se e e mina ion. Once he es ins ance comple es i s execu ion, he
se e is e mina ed o comple e he i e a ion. This ensu es a clean s a e
o subsequen es s and main ains he in eg i y o he es ing en i on-
men . In si ua ions ha equi e complemen a y eleme y se ices, such
as he Jaege compa ison, addi ional componen s (such as he Jaege
‘‘all-in-one’’ se ice) a e also e mina ed.
4.2. Valida ion esul s
In his subsec ion, we de ail he analysis o he impac o eleme y
on mic ose ice pe o mance, ocusing on esponse imes, CPU usage,
and memo y usage ac oss he di e en es cases (TC-01, TC-02, TC-03)
desc ibed abo e.
4.2.1. Syn he ic mic ose ice eleme y pe o mance esul s
In his scena io, we e alua e he pe o mance o he eleme y
amewo k o e a syn he ic mic ose ice ( he KS-API) in he h ee
di e en es cases.
TC-01 — Sho du a ion wi h di e en loads
15 h ps://gi hub.com/pa mon/apipecke .
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
9
M. O e o e al.
Fig. 15. TC-01; Syn he ic en i onmen : Response imes and Memo y Consump ion. (Fo in e p e a ion o he e e ences o colo in his igu e legend, he eade is e e ed o
he web e sion o his a icle.)
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
16

M. O e o e al.
wo solu ions is expec ed, as bo h packages u ilize OpenTeleme y o
ace expo . O e all, hese indings demons a e ha ou in-memo y
expo e implemen a ion in oduces no signi ican o e head.
Memo y usage.In Fig. 15(b), memo y consump ion is analyzed ac oss
h ee p oblem di icul y le els. This igu e p esen s h ee dis inc
g aphs, each co esponding o a di e en p oblem di icul y. The g aphs
compa e memo y consump ion in megaby es (MB) agains he imes-
amp in seconds. I is impo an o no e ha he 𝑌-axis in hese g aphs
ep esen s memo y usage in megaby es a he han pe cen ages, as seen
in he p e ious es esul s. This dis inc ion a ises because i would no
be app op ia e o combine memo y pe cen ages, which a e calcula ed
pe con aine . The e o e, he ac ual memo y usage in megaby es is
depic ed o p o ide a ai compa ison.
The igu e con ains h ee key lines o each eleme y e sion: a
black dashed line o he o iginal KS-API applica ion (baseline), a g een
do -dashed line o he KS-API ins umen ed wi h Jaege eleme y,
and a ed solid line o he KS-API ins umen ed wi h OAS- eleme y.
No ably, he e is a signi ican dis inc ion ega ding he Jaege ins u-
men a ion. Unlike OAS- eleme y, which s o es eleme y da a in an
in-memo y da abase, Jaege necessi a es an ex e nal se ice o col-
lec he da a (all-in-one con aine ), depic ed by he dashed g ay line
wi h a ligh g ay backg ound. The g een line, ep esen ing he o al
memo y consump ion o he Jaege ecosys em, is he agg ega e o he
memo y usage om bo h he all-in-one con aine (g ay backg ound)
and he KS-API ins umen ed wi h Jaege (g een backg ound). This
dis inc ion unde sco es he addi ional esou ce consump ion associa ed
wi h Jaege ’s ex e nal se ice equi emen .
In he nex pa ag aphs, he esul s a e analyzed o he h ee p ob-
lem sizes: small, medium, and la ge. The compa ison includes he
o iginal KS-API applica ion e e ed as ks-api(No TLM) o baseline, he
KS-API ins umen ed wi h Jaege eleme y e e ed as ks-api(JGR), he
KS-API ins umen ed wi h OAS- eleme y e e ed as ks-api(OAS-TLM),
he Jaege ex e nal se ice e e ed as all-in-one(JGR), and he o al
memo y consump ion o he Jaege ecosys em e e ed as o al(JGR).
Fo small-scale p oblem ins ances, he memo y consump ion o he
ks-api (OAS-TLM) con igu a ion a e ages 73.00 MB, ep esen ing a
69.93% inc ease ela i e o he baseline o 42.96 MB. In compa ison, i
we only conside ins umen a ion mechanisms, he ks-api (JGR) con ig-
u a ion exhibi s a sligh ly lowe a e age memo y usage o 65.28 MB,
which co esponds o a 51.95% inc ease o e he baseline. Howe e ,
he ex e nal Jaege se ice (all-in-one (JGR)) con ibu es an a e age
memo y consump ion o 79.61 MB. When he memo y usage o bo h
he ks-api (JGR) and he ex e nal Jaege se ice (all-in-one (JGR))
is agg ega ed, he o al memo y consump ion ises o 144.89 MB,
e lec ing a subs an ial 237.25% inc ease by Jaege se ices compa ed
o he baseline.
Fo medium-scale p oblem ins ances, he ks-api (OAS-TLM) con ig-
u a ion demons a es an a e age memo y usage o 92.45 MB, which
cons i u es a 47.98% inc ease o e he baseline alue o 62.47 MB.
The ks-api (JGR) con igu a ion consumes 77.96 MB on a e age, ep-
esen ing a 24.79% inc ease compa ed o he baseline. The ex e nal
Jaege se ice (all-in-one (JGR)) u he con ibu es an a e age mem-
o y consump ion o 73.96 MB. When he memo y consump ion o bo h
he ks-api (JGR) and he ex e nal se ice (all-in-one (JGR)) is summed,
he o al memo y usage amoun s o 151.92 MB, which indica es a
signi ican 143.17% inc ease ela i e o he baseline.
Fo la ge-scale p oblem ins ances, he ks-api (OAS-TLM) con ig-
u a ion exhibi s an a e age memo y consump ion o 117.73 MB, a
20.82% inc ease o e he baseline o 97.44 MB. The ks-api (JGR) con-
igu a ion equi es sligh ly mo e memo y, wi h an a e age o 125.57
MB, e lec ing a 28.87% inc ease ela i e o he baseline. Addi ionally,
he ex e nal Jaege se ice (all-in-one (JGR)) consumes 78.52 MB on
a e age. The combined memo y consump ion o bo h he ks-api (JGR)
and he ex e nal Jaege se ice (all-in-one (JGR)) esul s in a o al
Table 11
CPU usage pe cen and ela i e change o baseline in Jaege compa ison es s.
O de Mic ose ice Mean CPU usage (%) Rela i e change o baseline (%)
L
All-in-one 0.019 –
ks-api(JGR) 6.311 4.403
To al(JGR) 6.330 4.722
ks-api(baseline) 6.045 0
ks-api(oas- lm) 6.296 4.157
M
All-in-one 0.019 –
ks-api(JGR) 0.522 68.350
To al(JGR) 0.541 74.517
ks-api(baseline) 0.310 0
ks-api(oas- lm) 0.411 32.750
S
All-in-one 0.019 –
ks-api(JGR) 0.221 195.176
To al(JGR) 0.239 220.293
ks-api(baseline) 0.075 0
ks-api(oas- lm) 0.127 70.211
memo y usage o 204.09 MB, which co esponds o a no able 109.45%
inc ease when compa ed o he baseline.
Memo y consump ion da a e eals ha he inco po a ion o an
ex e nal se ice o eleme y, such as Jaege , signi ican ly con ibu es
o inc eased memo y consump ion. In con as , he ks-api (OAS-TLM)
con igu a ion exhibi s a lowe memo y o e head by u ilizing an in-
memo y da abase o eleme y s o age. This dis inc ion is o pa icula
impo ance in scena ios whe e memo y esou ces a e cons ained, as
he addi ional memo y consump ion associa ed wi h ex e nal eleme y
se ices can subs an ially impac o e all sys em pe o mance.
CPU usage.Table 11 p esen s CPU usage da a ac oss di e en p oblem
sizes: la ge (L), medium (M), and small (S). The able compa es he CPU
usage (in %) and he ela i e change om he baseline o a ious con-
igu a ions, including he Jaege all-in-one ex e nal se ice (all-in-one),
he Jaege ins umen a ion o ks-api (ks-api(JGR)), he o al Jaege
se ices (ins umen a ion + all-in-one), he baseline wi h no eleme y
(ks-api(baseline)), and OAS-TLM (ks-api(oas- lm)) ins umen a ion.
The obse a ions indica e ha he CPU usage o he ‘‘all-in-one’’
Jaege se ice emains consis en ly low, as expec ed, since i is no
hea ily s essed. Howe e , his se ice is essen ial o cap u ing he
o al ene gy consump ion o he Jaege ecosys em.
When compa ing he CPU usage be ween Jaege ins umen a ion
(ks-api(JGR)) and OAS-TLM (ks-api(oas- lm)), a clea dis inc ion
eme ges in e ms o esou ce usage ac oss he di e en p oblem sizes. In
he la ge size (L), Jaege ins umen a ion esul s in 6.311% CPU usage,
a 4.4% inc ease om he baseline, while OAS-TLM shows sligh ly
lowe usage a 6.296%, wi h a 4.16% inc ease. This shows ha bo h
con igu a ions ha e a simila impac a he la ge scale, wi h Jaege
ins umen a ion ma ginally highe . In he medium size (M), he gap
widens: Jaege ’s CPU usage inc eases o 0.522%, a 68.4% ise om
he baseline, while OAS-TLM is sligh ly lowe a 0.411%, e lec ing
a 32.8% inc ease. This di e ence becomes mo e p onounced in he
small p oblem size (S), whe e Jaege ins umen a ion esul s in 0.221%
CPU usage, a 195% inc ease om he baseline, compa ed o 0.127%
o OAS-TLM, which ep esen s a 70.2% ise. O e all, Jaege ins u-
men a ion ends o use mo e CPU esou ces han OAS-TLM ac oss all
sizes, wi h he dispa i y becoming mo e signi ican as he p oblem size
dec eases. These esul s highligh he pe o mance o e head in oduced
by bo h eleme y amewo ks, wi h OAS-TLM o e ing a mo e CPU-
e icien solu ion. The implica ions o ene gy consump ion will be
explo ed in he discussion sec ion.
Ene gy consump ion.The compa ison o eleme y solu ions e ealed
an ene gy sa ings po en ial when using ou OAS- eleme y amewo k
compa ed o Jaege . In o de o do his compa ison we es ima e he
ene gy consump ion based on CPU usage esul s using he The mal
Design Powe (TDP) ha is p o ided by manu ac u e s and ep esen s
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
17
M. O e o e al.
he amoun o hea a p ocesso mus dissipa e unde maximum load,
ypically measu ed in wa s. Al hough TDP is no di ec ly equi alen o
inpu powe , i is closely ela ed, as nea ly all he ene gy consumed by
a p ocesso is ul ima ely con e ed in o hea .
The analysis p esen ed is based on he expe imen al esul s ob ained
using a machine wi h an In el Co e i5-12600KF p ocesso , ea u ing 16
logical p ocesso s (10 physical co es: 4 op imized o ene gy e iciency
and 6 o high pe o mance). Speci ically, he p ocesso ’s The mal
Design Powe (TDP) alues, p o ided by he manu ac u e , a e as
ollows: he Base TDP is 125 W, which we will e e o as he B eeze
scena io, ep esen ing ypical ene gy consump ion unde s anda d load.
The Tu bo TDP is 150 W, e e ed o as he In e no scena io, e lec ing
maximum ene gy consump ion unde peak pe o mance.
The CPU usage esul s p esen ed in Table 11 a e gi en as pe cen -
ages. Howe e , o ou calcula ions, we will use he CPU usage a io 𝑢𝑖
( a he han he pe cen age). To con e he pe cen age alues in o he
a io, we use he ollowing o mula:
𝑢𝑖=
𝑢pe cen
𝑖
100
Nex , o calcula e he ene gy sa ings in he CPU when using ou
amewo k compa ed o Jaege , we i s need o de e mine he addi-
ional CPU usage in oduced by each ins umen a ion me hod. To do
his, we sub ac he o iginal (unins umen ed) CPU usage om he
ins umen ed CPU usage o each me hod.
The o mula o his is as ollows:
𝛥𝑢𝑖=𝑢ins um
𝑖−𝑢o ig
whe e:
•𝛥𝑢𝑖 ep esen s he addi ional CPU usage caused by he eleme y
ins umen a ion (such as OAS-TLM o Jaege ) in he mic ose ice.
•𝑢ins um
𝑖is he CPU usage a io o he ins umen ed applica ion.
•𝑢o ig is he CPU usage a io o he o iginal, unins umen ed
applica ion.
This calcula ion allows us o quan i y he impac o each ins u-
men a ion on CPU usage, enabling us o assess he po en ial ene gy
sa ings when swi ching om one amewo k o ano he . To es ima e
he addi ional powe consump ion in oduced by he ins umen a ion,
we mul iply he addi ional CPU usage a io by he The mal Design
Powe (TDP) o he p ocesso . In ou case, we conside wo p ocesso
scena ios: B eeze (TDP =125 W) and In e no (TDP =150 W), ep-
esen ing he minimum and maximum powe consump ion scena ios,
espec i ely.
𝑃add,𝑖 =𝛥𝑢𝑖×TDP𝑠
whe e 𝑃add,𝑖 is he es ima ed powe added by he ins umen a ion
o mic ose ice 𝑖, and TDP𝑠is he The mal Design Powe (TDP) o
he p ocesso in he ‘‘B eeze’’ (minimum) o ‘‘In e no’’ (maximum)
scena io.
To calcula e he ene gy consump ion o e he cou se o one hou ,
we mul iply he addi ional powe consump ion by he ime du a ion
(in hou s). This gi es he ene gy consumed in wa -hou s (Wh), which
is a sui able uni o his s udy since we a e ocusing on CPU- ela ed
ene gy consump ion. Al hough elec ical ene gy is commonly measu ed
in kilowa -hou s (kWh) in la ge -scale powe sys ems, wa -hou s (Wh)
a e mo e app op ia e he e, as hey di ec ly e lec he inc ease in
ene gy usage due o CPU o e heads. The o mula o his calcula ion
is:
𝐸add,𝑖 =𝑃add,𝑖 × 1h
whe e 𝐸add,𝑖 is he es ima ed ex a ene gy consumed by he ins umen-
a ion in one hou o mic ose ice ins umen a ion 𝑖.
The esul s o applying he abo e o mulas a e p esen ed in
Table 12. The da a show ha , o la ge p oblem sizes (o de s o magni-
ude), bo h amewo ks add mo e powe consump ion (P+). Howe e ,
Table 12
Inc eased CPU and ene gy consump ion due o eleme y ins umen a ions in ks-api.
O de Ins um. B eeze (125 W) In e no (150 W)
P+(W) E+(Wh) P+(W) E+(Wh)
LJaege 0.333 0.333 0.399 0.399
oas- lm 0.314 0.314 0.377 0.377
MJaege 0.265 0.265 0.318 0.318
oas- lm 0.127 0.127 0.152 0.152
SJaege 0.182 0.182 0.219 0.219
oas- lm 0.066 0.066 0.079 0.079
Table 13
To al CPU powe and ene gy consumed by Jaege all-in-one se ice.
O de B eeze (125 W) In e no (150 W)
P (W) E (Wh) P (W) E (Wh)
L 0.024 0.024 0.029 0.029
M 0.024 0.024 0.029 0.029
S 0.023 0.023 0.028 0.028
Jaege consis en ly uses mo e powe , wi h he la ges di e ences
obse ed in smalle p oblem sizes. Speci ically, Jaege ’s addi ional
powe consump ion anges om 0.182 W o 0.219 W, while OAS-
TLM consumes be ween 0.066 W and 0.079 W. This indica es ha ,
while Jaege in oduces a highe o e head, he ex a CPU ene gy
consump ion (E+) di e ence becomes mo e p onounced in smalle
p oblem sizes, whe e he o e heads o OAS-TLM a e ela i ely lowe .
To quan i y he ene gy sa ings when using OAS-TLM ins ead o
Jaege , we i s calcula e he di e ence in ene gy consump ion be ween
he wo eleme y solu ions. This di e ence is gi en by he ollowing
o mula:
𝐸di =𝐸i=jaege −𝐸i=oas- lm
whe e 𝐸di ep esen s he ene gy sa ed (in CPU) by using he OAS-
TLM eleme y solu ion ins ead o Jaege . 𝐸OAS-TLM and 𝐸Jaege e e
o he ex a ene gy consumed by he OAS-TLM and Jaege eleme y
solu ions, espec i ely.
I is impo an o conside ha Jaege ’s ecosys em equi es an
addi ional mic ose ice, he ‘‘all-in-one’’ se ice. This se ice plays a
c i ical ole by se ing as bo h a collec o and a que y engine o
eleme y da a, and i is esponsible o s o ing he collec ed da a.
As a esul , his addi ional mic ose ice in oduces ex a CPU ene gy
consump ion, which is no p esen in he OAS-TLM se up. This added
ene gy consump ion mus be ac o ed in o ou o e all ene gy sa ings
calcula ion when compa ing he wo eleme y solu ions.
The o al powe (P) and ene gy (E) consumed by he ‘‘all-in-one’’
se ice in one hou a e p esen ed in Table 13. The ene gy consumed
by his se ice is calcula ed using he same o mula as he ene gy
consumed by he indi idual mic ose ices. Speci ically, he ene gy
consump ion o he ‘‘all-in-one’’ se ice is gi en by:
𝑃all-in-one =𝑢all-in-one ×TDP𝑠(1)
𝐸all-in-one =𝑃all-in-one × 1h (2)
whe e 𝑢all-in-one is he CPU usage a io o he ‘‘all-in-one’’ se ice, and
TDP𝑠is he The mal Design Powe o he p ocesso . This addi ional en-
e gy consump ion is coun ed only once in he calcula ion, as i applies
o he en i e Jaege se up, ega dless o he numbe o mic ose ices
being ins umen ed.
To be e unde s and he impac o ene gy sa ings in eal-wo ld
scena ios, we calcula e he ene gy sa ings based on he numbe o
mic ose ices in he applica ion; in ac , he magni ude o his numbe
can g ea ly a y om dozens o housands o mic ose ice in ce ain
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
18
M. O e o e al.
Table 14
Addi ional CPU ene gy consump ion sa ed when ins umen ed wi h OAS-Teleme y
ins ead o Jaege .
O de Mic o-se ices B eeze (125 W) In e no (150 W)
E sa ed (Wh) E sa ed (Wh)
L
1 0.043 0.051
20 0.395 0.474
300 5.591 6.710
1000 18.582 22.298
M
1 0.162 0.194
20 2.781 3.337
300 41.380 49.656
1000 137.877 165.453
S
1 0.140 0.168
20 2.359 2.831
300 35.053 42.064
1000 116.789 140.147
scena ios (e.g. Ne lix had o e 700 mic ose ices back in 201516). Con-
sequen ly, in ou analysis, we compu e he ene gy sa ings o di e en
numbe s o mic ose ices: 1, 20, 300, and 1000. This enables us o
cap u e how he ene gy sa ings scale as he numbe o mic ose ices
inc eases.
The o al ene gy sa ings ac oss all mic ose ices is hen calcula ed
using he ollowing o mula:
𝐸sa ed,𝑛 =𝐸di ×𝑛+𝐸all-in-one
whe e 𝐸sa ed,𝑛 is he o al CPU ene gy sa ed by OAS-TLM in wa -
hou s o 𝑛mic ose ices, 𝐸di is he ene gy di e ence be ween Jaege
and OAS-TLM, 𝑛is he numbe o mic ose ices, and 𝐸all-in-one is he
ene gy consumed by he Jaege ‘‘all-in-one’’ se ice. By calcula ing
he ene gy sa ings ac oss di e en mic ose ice scales, we can mo e
e ec i ely compa e he ene gy e iciency o he wo eleme y solu ions
in la ge-scale, eal-wo ld applica ions.
The esul s in Table 14 demons a e he ene gy sa ings achie ed
by he OAS- eleme y amewo k compa ed o Jaege ac oss di e en
wo kloads and p ocesso con igu a ions. The da a highligh s he po en-
ial o signi ican ene gy educ ion, wi h a ia ions based on bo h he
numbe o mic ose ices and he he mal design powe (TDP) o he
p ocesso .
4.3. Discussion
The alida ion esul s p o ide a comp ehensi e analysis o he im-
pac o eleme y on mic ose ice pe o mance. By examining a ious
es cases (TC-01, TC-02, and TC-03) in di e en en i onmen s (node.js
in a Windows pc, docke in a Linux se e , and docke in a windows
pc) unde di e en load condi ions (small, medium, and la ge), we
ha e been able o assess he e ec s on esponse imes, CPU usage, and
memo y consump ion. This di e se se o alida ions ensu es ha ou
indings a e obus and applicable o a wide ange o scena ios. In he
ollowing discussion, we will in e p e he esul s, pu hem in con ex ,
and explain hei signi icance. We will highligh key ea u es such as
ene gy e iciency, eliabili y, and he ade-o s o using a dis ibu ed
eleme y app oach.
The syn he ic en i onmen es s demons a ed ha ou dis ibu ed
eleme y amewo k has minimal impac on esponse imes, CPU
usage, and memo y consump ion. The esul s con i med ha ou ap-
p oach is bo h ligh weigh and e icien , wi h esponse imes emaining
consis en unde a ying load condi ions. The eleme y sys em success-
ully collec ed da a wi hou signi ican ly a ec ing he pe o mance o
he ins umen ed mic ose ices. Addi ionally, he amewo k’s abili y
16 h ps://www.you ube.com/wa ch? =-mL3zT1iIKw.
o s a and s op eleme y collec ion u he highligh s i s esou ce-
e icien design, minimizing sys em esou ce usage when eleme y is
no needed.
The capabili y o dynamically adjus moni o ing collec ion p o-
ides a signi ican ad an age in esou ce managemen . This lexibili y
opens up oppo uni ies o in eg a ing AI-d i en plugins ha could
au oma ically adjus da a collec ion me hods, simila o he manual
adjus men s we cu en ly suppo . Fu he mo e, he in eg a ion o AI
could ex end o modi ying sys em con igu a ions — such as igge ing
ci cui b eake s o au oma ically scaling Kube ne es esou ces — o
u he op imize pe o mance.
In ou cu en sys em, plugins a e al eady unning wi hin he in-
s umen ed mic ose ice, in eg a ed wi h he plugin sys em. This se up
allows he plugins o ha e di ec access o he da a being collec ed,
enabling eal- ime analysis and ac ions. Wi h he in eg a ion o AI-
d i en plugins, hese plugins could modi y con igu a ions and da a
collec ion me hods dynamically. This ensu es a apid esponse o any
pe o mance issues, educing la ency and enhancing he sys em’s o e -
all esponsi eness. By enabling di ec in e ac ion wi h he sys em’s da a
and con igu a ions, he AI plugins could de ec and ec i y issues in eal
ime, u he op imizing pe o mance wi hou manual in e en ion.
To u he alida e he e iciency o ou amewo k, we ins u-
men ed a eal p oduc ion en i onmen wi h ou eleme y package.
The esul s showed ha ou app oach is capable o moni o ing mi-
c ose ices in a eal-wo ld se ing wi hou in oducing signi ican o e -
head. The eleme y sys em was able o collec da a e ec i ely, p o-
iding aluable insigh s in o he pe o mance o he ins umen ed
mic ose ices.
In TC-02, he esul s (see Figs. 12(b),12(d),12( )) indica ed ha no
signi ican changes we e obse ed in memo y usage when eleme y
was s opped. As men ioned in he esul s sec ion, his is likely due
o he low numbe o eques s, which helps p e en se e o e load,
as he egis y manages la ge, cos ly da a. Howe e , in TC-02 o he
syn he ic alida ion, he ks-api ecei ed eques s e e y 200 ms and a
no iceable di e ence in memo y usage was obse ed (see Fig. 8(b))
when eleme y collec ion was hal ed.
The compa ison wi h Jaege highligh ed he ad an ages o ou dis-
ibu ed eleme y app oach. Ou amewo k consumes ewe esou ces
han cen alized solu ions like Jaege , especially unde smalle loads,
while main aining simila pe o mance in e ms o esponse imes.
This demons a es he ene gy e iciency o ou app oach, making i
well-sui ed o dynamic and dis ibu ed mic ose ice a chi ec u es.
The esul s show ha ou imp o ed pe o mance is pa ly a ibu ed
o elimina ing he eliance on ex e nal se ices such as Jaege . S o ing
da a in-memo y educes he need o con inuous communica ion wi h a
cen alized sys em, he eby minimizing la ency and ne wo k o e head.
This app oach no only cu s ansmission cos s bu also boos s o e all
sys em e iciency. While he ad an ages a e clea , a mo e p ecise
measu emen o his impac would be aluable in u u e e alua ions.
One o he ade-o s o ou app oach is he po en ial loss o da a
when a node goes down o when he e is a la ge olume o eleme y
da a being gene a ed. To add ess his challenge, we a e de eloping
he dynamicExpo e — a module ha allows ou sys em o unc ion
like s anda d OpenTeleme y ins umen a ion, wi h he abili y o expo
da a o cen alized eleme y se ices. This expo e will enable he
sys em o decide whe he o s o e da a in memo y o , when necessa y,
‘dump’ i back o a cen alized se ice. This hyb id app oach p o ides
he ad an age o o e ing eleme y da a close o he sou ce, while
s ill main aining he unc ionali y o adi ional eleme y sys ems, hus
o e ing g ea e lexibili y o amewo k use s.
Ano he ade-o o conside is he esou ce consump ion associa ed
wi h unning mul iple nodes o eleme y da a collec ion. Howe e ,
eleme y ins umen a ion is always in eg a ed in o exis ing mic ose -
ices, and he cos incu ed by he eleme y sys em is p ima ily he di -
e ence in esou ce usage be ween ins umen ed and non-ins umen ed
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
19
M. O e o e al.
mic ose ices. Fu he mo e, adi ional cen alized sys ems o en e-
qui e addi ional in e media ies, such as agen s o ga eways, alongside
he backend s o age, making ou sys em po en ially mo e e icien .
We plan o alida e his e iciency in u u e es s o compa e he
pe o mance and esou ce consump ion o ou dis ibu ed model wi h
ha o adi ional cen alized sys ems.
Wi h espec o ene gy consump ion, he esul s show how OAS-
eleme y amewo k consis en ly deli e s ene gy sa ings ac oss small,
medium, and la ge wo kloads when compa ed wi h a widely used
al e na i e (Jaege ). Al hough he absolu e ene gy sa ings pe mi-
c ose ice a e smalle o la ge p oblem sizes, he amewo k achie es
impo an educ ions in ene gy consump ion o small o medium-
sized wo kloads. As he numbe o mic ose ices inc eases, he o e all
ene gy sa ings g ow subs an ially, u he emphasizing he scalabili y
o he amewo k. As expec ed, he ene gy sa ings a e mo e p o-
nounced unde highe TDP condi ions (In e no) compa ed o he lowe
TDP (B eeze), highligh ing he c i ical ole o ha dwa e e iciency in
op imizing o e all ene gy consump ion.
These indings unde sco e he ad an ages o an in-memo y eleme-
y s o age solu ion, pa icula ly in scena ios whe e ene gy e iciency
is c i ical. In addi ion o no equi ing an ‘‘all-in-one’’ se ice, ou
amewo k a oids he o e head o deploying addi ional componen s
such as collec o agen s, ga eways, and backend se ices. This con-
ibu es o a mo e s eamlined a chi ec u e and u he educes ene gy
consump ion. I is impo an o no e ha ou analysis ocuses solely on
CPU usage, and does no accoun o he addi ional ene gy consump ion
equi ed by he deploymen o o he se ices needed in o he eleme y
amewo ks, which also in luence he o al ene gy consumed when
using eleme y amewo ks.
While ou analysis p o ides an ini ial ounda ion o ene gy e i-
ciency, i is essen ial o alida e hese indings wi h eal-wo ld ha d-
wa e measu emen s. Fu u e wo k will in ol e implemen ing a es bed
o measu e ac ual powe consump ion du ing he execu ion o eleme-
y asks. Such an expe imen al se up will p o ide a mo e p ecise
e alua ion o ene gy e iciency and enable us o e ine ou amewo k
u he .
5. Conclusions
This a icle p esen s a no el, ligh weigh dis ibu ed eleme y
amewo k speci ically ailo ed o mic ose ice a chi ec u es, le e -
aging he OpenAPI Speci ica ion (OAS). Unlike adi ional cen alized
eleme y sys ems, which o en ace challenges o complexi y and igid-
i y in dis ibu ed en i onmen s, ou app oach u ilizes decen alized
eleme y da a s o age. This imp o es sys em esilience in he e en o
ne wo k ailu es and educes he bo lenecks ypically associa ed wi h
cen alized sys ems.
The e ec i eness o ou app oach was alida ed in h ee dis inc
en i onmen s and applica ion scena ios. Expe imen al esul s show ha
ou dis ibu ed eleme y amewo k pe o ms wi h minimal o e head
compa ed o unins umen ed applica ions and o he moni o ing ools
(i.e. Jaege , a widely used open sou ce al e na i e). This highligh s i s
po en ial o enhance obse abili y while main aining sys em e iciency
and demons a ing ad an ages in e ms o ene gy sa ings, pa icula ly
o dis ibu ed and dynamic mic ose ice a chi ec u es.
Fu u e wo k will ocus on expanding he amewo k’s capabili ies by
enabling dis ibu ed access o eleme y da a ac oss mul iple nodes in
he use in e ace. To complemen his, new plugins will be in oduced
o no only collec da a bu also enable he amewo k o adap in
eal- ime using a i icial in elligence, imp o ing anomaly de ec ion
and op imizing sys em pe o mance. Addi ionally, we will alida e he
amewo k in new en i onmen s and compa e i wi h di e en mon-
i o ing ools such as OpenZipkin, while measu ing addi ional me ics
such as ne wo k usage and i s ela ed ene gy consump ion.
CRediT au ho ship con ibu ion s a emen
Manuel O e o: W i ing – e iew & edi ing, W i ing – o iginal
d a , So wa e, Fo mal analysis, Da a cu a ion, Concep ualiza ion.
José Ma ía Ga cía: W i ing – e iew & edi ing, Concep ualiza ion.
Pablo Fe nandez: W i ing – e iew & edi ing, W i ing – o iginal d a ,
Supe ision, Funding acquisi ion, Concep ualiza ion.
Decla a ion o compe ing in e es
The au ho s decla e he ollowing inancial in e es s/pe sonal ela-
ionships which may be conside ed as po en ial compe ing in e es s:
Pablo Fe nandez epo s inancial suppo was p o ided by Spain Min-
is y o Science and Inno a ion. Pablo Fe nandez epo s inancial
suppo was p o ided by Go e nmen o Andalusia. I he e a e o he
au ho s, hey decla e ha hey ha e no known compe ing inancial in-
e es s o pe sonal ela ionships ha could ha e appea ed o in luence
he wo k epo ed in his pape .
Acknowledgmen s
This publica ion is pa o he R&D p ojec PID2021-126227NB-C22,
unded by MICIU/AEI/10.13039/501100011033/ERDF/EU, Spain and
o p ojec s TED2021-131023B-C21 and PDC2022-133521-I00 unded
by MICIU/AEI/10.13039/501100011033/Eu opean Union Nex Gene -
a ionEU/PRTR, Spain.
Da a a ailabili y
Da a will be made a ailable on eques .
Re e ences
[1] T. Salah, M. Jamal Zeme ly, C.Y. Yeun, M. Al-Qu ay i, Y. Al-Hammadi, The
e olu ion o dis ibu ed sys ems owa ds mic ose ices a chi ec u e, in: 2016
11 h In e na ional Con e ence o In e ne Technology and Secu ed T ansac ions,
ICITST, IEEE, 2016, h p://dx.doi.o g/10.1109/ICITST.2016.7856721.
[2] S. Hassan, R. Bahsoon, Mic ose ices and hei design ade-o s: A sel -adap i e
oadmap, in: 2016 IEEE In e na ional Con e ence on Se ices Compu ing, SCC,
2016, pp. 813–818, h p://dx.doi.o g/10.1109/SCC.2016.113.
[3] F. Mon esi, J. Webe , Ci cui b eake s, disco e y, and API ga eways in
mic ose ices, 2016, a Xi :1609.05830.
[4] Y. Ve ginadis, A e iew o moni o ing p obes o cloud compu ing con inuum,
in: L. Ba olli (Ed.), Ad anced In o ma ion Ne wo king and Applica ions, Sp inge
In e na ional Publishing, Cham, 2023, pp. 631–643, h p://dx.doi.o g/10.1007/
978-3-031-28694-0_59.
[5] C. Misa, R. Du ai ajan, R. Rejaie, W. Willinge , Re isi ing ne wo k eleme y in
COIN: A case o un ime p og ammabili y, IEEE Ne w. 35 (5) (2021) 14–20,
h p://dx.doi.o g/10.1109/MNET.201.2100064.
[6] A. Fe ikoglou, A. Kokkinis, D. Danopoulos, I. O ou zoglou, A. Nanos, S.
Ka anas asis, M. Sipos, J.F. Gho bi, J.J. Vegas Olmos, D. Masou os, K. Siozios,
The SERRANO pla o m: S epping owa ds seamless applica ion de elopmen
& deploymen in he he e ogeneous edge-cloud con inuum, in: 2023 Design,
Au oma ion & Tes in Eu ope Con e ence & Exhibi ion, 2023, pp. 1–4, h p:
//dx.doi.o g/10.23919/DATE56975.2023.10137039.
[7] S. Dus da , V. Casamayo Pujol, P.K. Don a, On dis ibu ed compu ing con inuum
sys ems, IEEE T ans. Knowl. Da a Eng. 35 (4) (2022) 4092–4105, h p://dx.doi.
o g/10.1109/TKDE.2022.3142856.
[8] V. Casamayo Pujol, P.K. Don a, A. Mo iche a, I. Mu u i, S. Dus da , Edge
in elligence— esea ch oppo uni ies o dis ibu ed compu ing con inuum sys-
ems, IEEE In e ne Compu . 27 (4) (2023) 53–74, h p://dx.doi.o g/10.1109/
MIC.2023.3284693.
[9] S. Laso, J. Be ocal, P. Fe nandez, J.M. Ga cía, J. Ga cia-Alonso, J.M. Mu illo, A.
Ruiz-Co és, S. Dus da , Elas ic da a analy ics o he cloud- o- hings con inuum,
IEEE In e ne Compu . 26 (6) (2022) 42–49, h p://dx.doi.o g/10.1109/MIC.
2021.3138153.
[10] V. Casamayo Pujol, A. Mo iche a, I. Mu u i, P. Kuma Don a, S. Dus da ,
Fundamen al esea ch challenges o dis ibu ed compu ing con inuum sys ems,
In o ma ion 14 (3) (2023) 198, h p://dx.doi.o g/10.3390/in o14030198.
[11] R. Engel, P. Fe nandez, A. Ruiz-Co es, A. Megahed, J. Ojeda-Pe ez, SLA-awa e
ope a ional e iciency in AI-enabled se ice chains: challenges ahead, In . Sys .
E- Bus. Manag. 20 (1) (2022) 199–221, h p://dx.doi.o g/10.1007/s10257-022-
00551-w.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
20
M. O e o e al.
[12] S. Laso, J. Be ocal, P. Fe nández, A. Ruiz-Co és, J.M. Mu illo, Pe ses: A
amewo k o he con inuous e alua ion o he QoS o dis ibu ed mobile
applica ions, Pe asi e Mob. Compu . 84 (2022) 101627, h p://dx.doi.o g/10.
1016/j.pmcj.2022.101627.
[13] P. Jamshidi, C. Pahl, N. Mendonça, J. Lewis, S. Tilko , Mic ose ices: The
jou ney so a and challenges ahead, IEEE So w. 35 (2018) 24–35, h p://dx.
doi.o g/10.1109/MS.2018.2141039.
[14] M. Waseem, P. Liang, M. Shahin, A.D. Salle, G. M’a quez, Design, moni o ing,
and es ing o mic ose ices sys ems: The p ac i ione s’ pe spec i e, J. Sys .
So w. 182 (2021) 111061, h p://dx.doi.o g/10.1016/j.jss.2021.111061.
[15] J. Me z, I. Nunes, Tig is: a DSL and amewo k o moni o ing so wa e sys ems
a un ime, 2021, CoRR abs/2103.15986,a Xi :2103.15986. URL h ps://a xi .
o g/abs/2103.15986.
[16] A. an Hoo n, J. Walle , W. Hasselb ing, Kieke : A amewo k o applica ion
pe o mance moni o ing and dynamic so wa e analysis, in: P oceedings o he
3 d ACM/SPEC In e na ional Con e ence on Pe o mance Enginee ing, ICPE ’12,
Associa ion o Compu ing Machine y, New Yo k, NY, USA, 2012, pp. 247–248,
h p://dx.doi.o g/10.1145/2188286.2188326.
[17] J. Enes, R.R. Expósi o, J. Tou iño, BDWa chdog: Real- ime moni o ing and
p o iling o big da a applica ions and amewo ks, Fu u e Gene . Compu . Sys .
87 (2018) 420–437, h p://dx.doi.o g/10.1016/j. u u e.2017.12.068.
[18] P. Faymon ille, B. Finkbeine , M. Schledjewski, M. Schwenge , M. S enge ,
L. Ten up, H. To ah, S eamLAB: S eam-based moni o ing o cybe -physical
sys ems, in: I. Dillig, S. Tasi an (Eds.), Compu e Aided Ve i ica ion, Sp inge
In e na ional Publishing, Cham, 2019, pp. 421–431, h p://dx.doi.o g/10.1007/
978-3-030-25540-4_24.
[19] A. Noo , D. Jha, K. Mi a, P. Jaya aman, A. Souza, R. Ranjan, S. Dus da , A
amewo k o moni o ing mic ose ice-o ien ed cloud applica ions in he e oge-
neous i ualiza ion en i onmen s, in: 2019 IEEE 12 h In e na ional Con e ence
on Cloud Compu ing, CLOUD, 2019, pp. 156–163, h p://dx.doi.o g/10.1109/
CLOUD.2019.00035.
[20] A. Tundo, M. Mobilio, M. O ù, O. Riganelli, M. Guzmàn, L. Ma iani, VARYS:
an agnos ic model-d i en moni o ing-as-a-se ice amewo k o he cloud, in:
P oceedings o he 2019 27 h ACM Join Mee ing on Eu opean So wa e Engi-
nee ing Con e ence and Symposium on he Founda ions o So wa e Enginee ing,
in: ESEC/FSE 2019, Associa ion o Compu ing Machine y, New Yo k, NY, USA,
2019, pp. 1085–1089, h p://dx.doi.o g/10.1145/3338906.3341185.
[21] A. Ullah, H. Dagde i en, R.C. A iya u, J. DesLau ie s, T. Kiss, J. Bowden,
MiCADO-Edge: Towa ds an applica ion-le el o ches a o o he cloud- o-edge
compu ing con inuum, J. G id Compu . 19 (4) (2021) 47, h p://dx.doi.o g/10.
1007/s10723-021-09589-5.
[22] A. Al Ma u , A. Bakh in, T. Ce ny, D. Taibi, Using mic ose ice eleme y
da a o sys em dynamic analysis, in: 2022 IEEE In e na ional Con e ence on
Se ice-O ien ed Sys em Enginee ing, SOSE, 2022, pp. 29–38, h p://dx.doi.o g/
10.1109/SOSE55356.2022.00010.
[23] H.M. Kabamba, M. Khouzam, M. Dagenais, Vnode: Low-o e head anspa en
acing o node.js-based mic ose ice a chi ec u es, 2023, h p://dx.doi.o g/10.
48550/a Xi .2311.11095, A Xi a Xi :2311.11095.
[24] B. Sedlak, V.C. Pujol, P.K. Don a, S. Dus da , Equilib ium in he compu ing
con inuum h ough ac i e in e ence, Fu u e Gene . Compu . Sys . 160 (2024)
92–108, h p://dx.doi.o g/10.1016/j. u u e.2024.05.056.
[25] L. Meng, F. Ji, Y. Sun, T. Wang, De ec ing anomalies in mic ose ices wi h
execu ion ace compa ison, Fu u e Gene . Compu . Sys . 116 (2021) 291–301,
h p://dx.doi.o g/10.1016/j. u u e.2020.10.040.
[26] R. F esno-A anda, J. Ojeda-Pe ez, P. Fe nandez, A. Ruiz-Co és, Go e ni y.
An ag eemen -based se ice go e nance amewo k, So w. Impac s 19 (2024)
100629, h p://dx.doi.o g/10.1016/j.simpa.2024.100629.
[27] C. Ga cía, A. Gue e o, J. Zei so , S. Ko lakun a, P. Fe nandez, A. Fox, A.
Ruiz-Co és, Bluejay: A c oss- ooling audi amewo k o agile so wa e eams,
in: 2021 IEEE/ACM 43 d In e na ional Con e ence on So wa e Enginee ing:
So wa e Enginee ing Educa ion and T aining (ICSE-SEET), 2021, pp. 283–288,
h p://dx.doi.o g/10.1109/ICSE-SEET52601.2021.00038.
Sus ainable Compu ing: In o ma ics and Sys ems 46 (2025) 101100
21