Lura. fassara: A cikin wannan labarin, Banzai Cloud ya ba da misalin yadda za a iya amfani da kayan aikin sa na al'ada don yin Kafka sauƙi don amfani a cikin Kubernetes. Umurnai masu zuwa suna kwatanta yadda zaku iya tantance mafi kyawun girman kayan aikin ku kuma saita Kafka da kanta don cimma abubuwan da ake buƙata.
Apache Kafka dandamali ne mai rarrabawa don ƙirƙirar amintaccen, daidaitawa da babban aiki na ainihin-lokacin yawo. Za a iya ƙara ƙarfin ƙarfinsa ta amfani da Kubernetes. Don wannan mun ci gaba da kayan aiki da ake kira . Suna ba ku damar gudanar da Kafka akan Kubernetes kuma kuyi amfani da fasalulluka daban-daban, kamar ingantaccen daidaita tsarin dillali, ma'auni na tushen awo tare da sake daidaitawa, wayar da kan jama'a, "laushi" (mai godiya) mirgine sabuntawa, da sauransu.
Gwada Supertubes a cikin tarin ku:
curl https://getsupertubes.sh | sh и supertubes install -a --no-democluster --kubeconfig <path-to-eks-cluster-kubeconfig-file>Ko tuntuɓar juna . Hakanan zaka iya karanta game da wasu iyawar Kafka, aikin da aka sarrafa ta ta amfani da Supertubes da Kafka afareta. Mun riga mun rubuta game da su a kan blog:
- ;
- ;
- ;
- ;
- ;
- ;
- .
Lokacin da kuka yanke shawarar tura gungu na Kafka akan Kubernetes, wataƙila za ku fuskanci ƙalubalen ƙayyadaddun mafi girman girman kayan aikin da ke ƙasa da buƙatar daidaita tsarin Kafka don biyan buƙatun kayan aiki. Matsakaicin aikin kowane dillali an ƙaddara ta hanyar aiwatar da abubuwan abubuwan more rayuwa, kamar ƙwaƙwalwar ajiya, processor, saurin diski, bandwidth na cibiyar sadarwa, da sauransu.
Mahimmanci, saitin dillali ya kamata ya zama irin wannan don ana amfani da duk abubuwan abubuwan more rayuwa zuwa iyakar ƙarfin su. Koyaya, a rayuwa ta ainihi wannan saitin yana da rikitarwa sosai. Zai fi dacewa masu amfani za su saita dillalai don haɓaka amfani da abubuwa ɗaya ko biyu (faifai, ƙwaƙwalwar ajiya, ko processor). Gabaɗaya magana, dillali yana nuna matsakaicin aiki lokacin da tsarin sa ya ba da damar yin amfani da mafi ƙarancin sassa zuwa iyakarsa. Ta wannan hanyar za mu iya samun ra'ayi mara kyau na nauyin da dillali ɗaya zai iya ɗauka.
A ka'ida, za mu iya ƙididdige adadin dillalai da ake buƙata don ɗaukar nauyin da aka ba su. Koyaya, a aikace akwai zaɓuɓɓukan daidaitawa da yawa a matakai daban-daban wanda yana da matukar wahala (idan ba zai yiwu ba) don kimanta yuwuwar aikin na musamman. A takaice dai, yana da matukar wahala a tsara tsari bisa wasu ayyukan da aka bayar.
Ga masu amfani da Supertubes, yawanci muna ɗaukar hanya mai zuwa: muna farawa da wasu ƙayyadaddun tsari (kayan aiki + saitin), sannan auna aikin sa, daidaita saitunan dillalai kuma mu sake maimaita tsarin. Wannan yana faruwa har sai an yi amfani da mafi ƙarancin kayan aikin.
Ta wannan hanyar, muna samun ƙarin ra'ayi game da yawancin dillalai da gungun ke buƙatar ɗaukar wani nauyi (yawan dillalai kuma ya dogara da wasu dalilai, kamar ƙaramin adadin kwafin saƙon don tabbatar da juriya, adadin ɓarna. shugabanni, da sauransu). Bugu da kari, muna samun haske kan abin da kayan aikin ababen more rayuwa ke buƙatar sikeli a tsaye.
Wannan labarin zai yi magana game da matakan da muke ɗauka don samun mafi kyawun abubuwan da aka gyara a cikin saitunan farko da kuma auna kayan aiki na gungu na Kafka. Tsarin juriya mai ƙarfi yana buƙatar aƙalla dillalai guda uku masu gudana (min.insync.replicas=3), wanda aka rarraba a wurare daban-daban na samun dama. Don daidaitawa, aunawa da saka idanu akan abubuwan more rayuwa na Kubernetes, muna amfani da dandalin sarrafa kwantena namu don gajimare masu gauraya - . Yana goyan bayan kan-ƙarfe (ƙarfe bare, VMware) da nau'ikan gajimare guda biyar (Alibaba, AWS, Azure, Google, Oracle), da duk wani haɗin kansu.
Tunani a kan Kafka cluster kayayyakin more rayuwa da kuma tsari
Ga misalan da ke ƙasa, mun zaɓi AWS a matsayin mai ba da girgije da EKS a matsayin rarraba Kubernetes. Ana iya aiwatar da irin wannan tsari ta amfani da shi - Rarraba Kubernetes daga Banzai Cloud, wanda CNCF ta tabbatar.
faifai
Amazon yayi daban-daban . A jigon gp2 и io1 akwai faifan SSD, duk da haka, don tabbatar da babban kayan aiki gp2 yana cinye tara kuɗi (I/O credits), don haka mun fi son nau'in io1, wanda ke ba da daidaiton babban kayan aiki.
Nau'in misali
Ayyukan Kafka ya dogara sosai akan cache shafi na tsarin aiki, don haka muna buƙatar lokuta tare da isasshen ƙwaƙwalwar ajiya ga dillalai (JVM) da cache shafi. Misali c5.2 ku - kyakkyawan farawa, tunda yana da 16 GB na ƙwaƙwalwar ajiya da . Rashin hasara shi ne cewa yana da ikon samar da mafi girman aiki na kasa da mintuna 30 kowane awa 24. Idan nauyin aikin ku yana buƙatar mafi girman aiki na tsawon lokaci, ƙila kuna so kuyi la'akari da wasu nau'ikan misali. Abin da muka yi ke nan, muka tsaya c5.4 ku. Yana ba da mafi girman kayan aiki a ciki 593,75 Mb/s. Matsakaicin kayan aiki na ƙarar EBS io1 sama da misalin c5.4 ku, don haka mafi jinkirin kashi na abubuwan more rayuwa yana yiwuwa ya zama kayan aikin I/O na wannan nau'in misalin (wanda gwajin nauyin mu yakamata ya tabbatar).
Network
Dole ne kayan aikin cibiyar sadarwa ya zama babba idan aka kwatanta da aikin misalin VM da faifai, in ba haka ba hanyar sadarwar ta zama cikas. A cikin yanayinmu, hanyar sadarwa ta hanyar sadarwa c5.4 ku yana goyan bayan gudu har zuwa 10 Gb/s, wanda ya fi girma fiye da kayan aikin I/O na misalin VM.
Tukar Dillali
Ya kamata a tura dillalai (tsara a cikin Kubernetes) zuwa ƙofofin sadaukarwa don guje wa yin gasa tare da wasu matakai don CPU, ƙwaƙwalwar ajiya, cibiyar sadarwa, da albarkatun diski.
Sigar Java
Zaɓin ma'ana shine Java 11 saboda ya dace da Docker a ma'anar cewa JVM daidai yake ƙayyade masu sarrafawa da ƙwaƙwalwar ajiya da ke cikin akwati da dillali ke gudana. Sanin cewa iyakokin CPU suna da mahimmanci, JVM a ciki kuma a bayyane yana saita adadin zaren GC da zaren JIT. Mun yi amfani da hoton Kafka banzaicloud/kafka:2.13-2.4.0, wanda ya haɗa da sigar Kafka 2.4.0 (Scala 2.13) akan Java 11.
Idan kuna son ƙarin koyo game da Java/JVM akan Kubernetes, duba abubuwanmu masu zuwa:
- ;
- .
Saitunan ƙwaƙwalwar ajiya
Akwai maɓalli biyu masu mahimmanci don daidaita ƙwaƙwalwar dillali: saituna don JVM da na Kubernetes pod. Iyakar žwažwalwar ajiya da aka saita don kwafsa dole ne ya zama mafi girma fiye da matsakaicin girman tudu domin JVM ya sami wuri don metaspace Java wanda ke zaune a cikin ƙwaƙwalwar ajiyarsa da kuma ma'ajin shafin tsarin aiki wanda Kafka ke amfani da shi sosai. A cikin gwaje-gwajenmu mun ƙaddamar da dillalan Kafka tare da sigogi -Xmx4G -Xms2G, kuma iyakar ƙwaƙwalwar ajiya don kwafsa ya kasance 10 Gi. Lura cewa ana iya samun saitunan ƙwaƙwalwar ajiya don JVM ta atomatik ta amfani da su -XX:MaxRAMPercentage и -X:MinRAMPercentage, dangane da iyakar ƙwaƙwalwar ajiya don kwafsa.
Saitunan mai sarrafa dillali
Gabaɗaya magana, zaku iya haɓaka aiki ta hanyar haɓaka daidaito ta hanyar ƙara adadin zaren da Kafka ke amfani da shi. Yawancin na'urori masu sarrafawa don Kafka, mafi kyau. A cikin gwajin mu, mun fara da iyaka na masu sarrafawa 6 kuma a hankali (ta hanyar maimaitawa) sun ɗaga lambar su zuwa 15. Bugu da ƙari, mun saita. num.network.threads=12 a cikin saitunan dillali don ƙara yawan zaren da ke karɓar bayanai daga cibiyar sadarwa da aika shi. Nan da nan gano cewa dillalai masu bi ba za su iya karɓar kwafi da sauri ba, sun ɗaga num.replica.fetchers zuwa 4 don ƙara saurin da dillalan dillalai ke maimaita saƙonni daga shugabanni.
Load Generation Tool
Ya kamata ku tabbatar da cewa janareta mai ɗaukar nauyi da aka zaɓa bai ƙare aiki ba kafin gungu na Kafka (wanda ake ƙima) ya kai matsakaicin nauyinsa. A takaice dai, wajibi ne a gudanar da kima na farko na iyawar kayan aikin samar da kaya, sannan kuma zabar nau'ikan misali don shi tare da isassun na'urori masu sarrafawa da ƙwaƙwalwar ajiya. A wannan yanayin, kayan aikinmu zai samar da ƙarin kaya fiye da gungu na Kafka zai iya ɗauka. Bayan gwaje-gwaje da yawa, mun daidaita akan kwafi uku c5.4 ku, kowanne daga cikinsu yana da janareta mai aiki.
Benchmarking
Ma'aunin aiki tsari ne na maimaitawa wanda ya haɗa da matakai masu zuwa:
- kafa abubuwan more rayuwa (gungu na EKS, gungu na Kafka, kayan aikin haɓaka kaya, da Prometheus da Grafana);
- samar da kaya na wani ɗan lokaci don tace ɓangarorin bazuwar a cikin alamun aikin da aka tattara;
- daidaita kayan aikin dillali da daidaitawa bisa ga alamun aikin da aka lura;
- maimaita tsari har sai an cimma matakin da ake buƙata na kayan aikin cluster na Kafka. A lokaci guda, dole ne a kasance a kai a kai a sake sakewa kuma ya nuna ƙananan bambance-bambance a cikin kayan aiki.
Sashe na gaba yana bayyana matakan da aka yi yayin aiwatar da aikin tantance gungu na gwaji.
Kayan aiki
An yi amfani da waɗannan kayan aikin don hanzarta ƙaddamar da saitin tushe, samar da kaya, da auna aiki:
- don shirya gungu na EKS daga Amazon c (don tattara Kafka da ma'aunin kayan more rayuwa) da (don ganin waɗannan ma'auni). Mun yi amfani hadedde в sabis ɗin da ke ba da sa ido na tarayya, tattara bayanai na tsakiya, duban rashin lahani, dawo da bala'i, matakan tsaro na masana'antu da ƙari mai yawa.
- - kayan aiki don gwada lodin tarin Kafka.
- Dashboards na Grafana don ganin ma'aunin Kafka da abubuwan more rayuwa: , .
- Supertubes CLI don hanya mafi sauƙi don saita gungu na Kafka akan Kubernetes. Zookeeper, Kafka ma'aikacin, Manzo da sauran abubuwa da yawa an shigar kuma an tsara su yadda ya kamata don gudanar da gungun Kafka da aka shirya akan Kubernetes.
- Don sanyawa Farashin CLI yi amfani da umarnin da aka bayar .
Rahoton da aka ƙayyade na EKS
Shirya gungu na EKS tare da ƙofofin ma'aikata c5.4 ku a cikin wurare daban-daban na samuwa ga kwasfa tare da dillalan Kafka, da kuma ƙofofin sadaukarwa don janareta na kaya da kayan aikin sa ido.
banzai cluster create -f https://raw.githubusercontent.com/banzaicloud/kafka-operator/master/docs/benchmarks/infrastructure/cluster_eks_202001.jsonDa zarar tarin EKS ya tashi yana aiki, kunna haɗewar sa - za ta tura Prometheus da Grafana cikin tari.
Kafka tsarin sassan
Shigar da abubuwan tsarin Kafka (Zookeeper, kafka-operator) a cikin EKS ta amfani da supertubes CLI:
supertubes install -a --no-democluster --kubeconfig <path-to-eks-cluster-kubeconfig-file>Tarin Kafka
Ta hanyar tsoho, EKS yana amfani da nau'in nau'in juzu'in EBS gp2, don haka kuna buƙatar ƙirƙirar nau'in ajiya daban dangane da kundin io1 ga Kafka cluster:
kubectl create -f - <<EOF
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
name: fast-ssd
provisioner: kubernetes.io/aws-ebs
parameters:
type: io1
iopsPerGB: "50"
fsType: ext4
volumeBindingMode: WaitForFirstConsumer
EOF
Saita siga don dillalai min.insync.replicas=3 da tura kwas ɗin dillali a kan nodes a wurare daban-daban guda uku:
supertubes cluster create -n kafka --kubeconfig <path-to-eks-cluster-kubeconfig-file> -f https://raw.githubusercontent.com/banzaicloud/kafka-operator/master/docs/benchmarks/infrastructure/kafka_202001_3brokers.yaml --wait --timeout 600Maudu'i
Mun gudanar da nauyin janareta guda uku a layi daya. Kowannen su yana rubuta maudu’insa, wato muna bukatar batutuwa guda uku gaba daya:
supertubes cluster topic create -n kafka --kubeconfig <path-to-eks-cluster-kubeconfig-file> -f -<<EOF
apiVersion: kafka.banzaicloud.io/v1alpha1
kind: KafkaTopic
metadata:
name: perftest1
spec:
name: perftest1
partitions: 12
replicationFactor: 3
retention.ms: '28800000'
cleanup.policy: delete
EOF
supertubes cluster topic create -n kafka --kubeconfig <path-to-eks-cluster-kubeconfig-file> -f -<<EOF
apiVersion: kafka.banzaicloud.io/v1alpha1
kind: KafkaTopic
metadata:
name: perftest2
spec:
name: perftest2
partitions: 12
replicationFactor: 3
retention.ms: '28800000'
cleanup.policy: delete
EOF
supertubes cluster topic create -n kafka --kubeconfig <path-to-eks-cluster-kubeconfig-file> -f -<<EOF
apiVersion: kafka.banzaicloud.io/v1alpha1
kind: KafkaTopic
metadata:
name: perftest3
spec:
name: perftest3
partitions: 12
replicationFactor: 3
retention.ms: '28800000'
cleanup.policy: delete
EOFGa kowane maudu'i, ma'aunin maimaitawa shine 3-mafi ƙarancin ƙimar da aka ba da shawarar don tsarin samarwa sosai.
Load Generation Tool
Mun kaddamar da kwafi uku na injin samar da kaya (kowannensu ya rubuta a cikin wani batu daban). Don kwas ɗin janareta masu ɗaukar nauyi, kuna buƙatar saita kusancin kumburi don a tsara su kawai akan nodes ɗin da aka ware musu:
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
labels:
app: loadtest
name: perf-load1
namespace: kafka
spec:
progressDeadlineSeconds: 600
replicas: 1
revisionHistoryLimit: 10
selector:
matchLabels:
app: loadtest
strategy:
rollingUpdate:
maxSurge: 25%
maxUnavailable: 25%
type: RollingUpdate
template:
metadata:
creationTimestamp: null
labels:
app: loadtest
spec:
affinity:
nodeAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- key: nodepool.banzaicloud.io/name
operator: In
values:
- loadgen
containers:
- args:
- -brokers=kafka-0:29092,kafka-1:29092,kafka-2:29092,kafka-3:29092
- -topic=perftest1
- -required-acks=all
- -message-size=512
- -workers=20
image: banzaicloud/perfload:0.1.0-blog
imagePullPolicy: Always
name: sangrenel
resources:
limits:
cpu: 2
memory: 1Gi
requests:
cpu: 2
memory: 1Gi
terminationMessagePath: /dev/termination-log
terminationMessagePolicy: File
dnsPolicy: ClusterFirst
restartPolicy: Always
schedulerName: default-scheduler
securityContext: {}
terminationGracePeriodSeconds: 30Wasu abubuwan lura:
- Na'urar samar da kaya yana samar da sakonnin bytes 512 tsawon kuma yana buga su zuwa Kafka a cikin batches 500.
- Amfani da hujja
-required-acks=allAna ɗaukar littafin ya yi nasara lokacin da aka karɓi duk kwafin saƙon da aka haɗa tare da tabbatar da dillalan Kafka. Wannan yana nufin cewa a cikin ma'auni ba kawai saurin shugabanni ke karɓar saƙonni ba, har ma mabiyansu suna maimaita saƙonni. Manufar wannan gwajin ba don kimanta saurin karatun mabukaci ba (masu amfani) kwanan nan an karɓi saƙon da har yanzu suke cikin ma'ajin shafi na OS, da kwatancensa da saurin karanta saƙonnin da aka adana akan faifai. - The load Generator yana gudanar da ma'aikata 20 a layi daya (
-workers=20). Kowane ma'aikaci ya ƙunshi furodusa 5 waɗanda ke raba haɗin ma'aikaci zuwa gungu na Kafka. Sakamakon haka, kowane janareta yana da furodusa 100, kuma dukkansu suna aika saƙonni zuwa gungu na Kafka.
Kula da lafiyar tari
A yayin gwajin lodi na gungu na Kafka, mun kuma lura da lafiyar sa don tabbatar da cewa babu sake kunna kwafsa, babu kwafi da ba a daidaitawa ba, da matsakaicin kayan aiki tare da ƙaramin canji:
- Mai samar da kaya yana rubuta daidaitattun ƙididdiga game da adadin saƙonnin da aka buga da ƙimar kuskure. Yawan kuskure ya kamata ya kasance iri ɗaya
0,00%. - , wanda kafka-operator ya tura, yana ba da dashboard inda za mu iya sa ido kan yanayin gungu. Don duba wannan panel yi:
supertubes cluster cruisecontrol show -n kafka --kubeconfig <path-to-eks-cluster-kubeconfig-file> - Babban darajar ISR (lambar "in-sync" kwafi) raguwa da fadada suna daidai da 0.
Sakamakon aunawa
Dillalai 3, girman saƙo - 512 bytes
Tare da rarrabuwa a ko'ina cikin dillalai uku, mun sami damar yin aiki ~ 500 Mb/s (kimanin saƙonni dubu 990 a sakan daya):
Amfanin ƙwaƙwalwar ajiyar na'urar kama-da-wane ta JVM bai wuce 2 GB ba:
Fitar da diski ya kai matsakaicin adadin I/O a duk lokuta uku da dillalan ke gudana a kai:
Daga bayanan akan amfani da ƙwaƙwalwar ajiya ta nodes, yana biye da tsarin buffering da caching ya ɗauki ~ 10-15 GB:
Dillalai 3, girman saƙo - 100 bytes
Yayin da girman saƙon ke raguwa, abin da ake fitarwa yana raguwa da kusan 15-20%: lokacin da aka kashe sarrafa kowane saƙo yana rinjayar sa. Bugu da kari, nauyin mai sarrafa ya kusan ninki biyu.
Tun da dillali nodes har yanzu suna da muryoyin da ba a yi amfani da su ba, ana iya inganta aikin ta hanyar canza saitin Kafka. Wannan ba abu ne mai sauƙi ba, don haka don ƙara yawan kayan aiki yana da kyau a yi aiki tare da manyan saƙonni.
Dillalai 4, girman saƙo - 512 bytes
Kuna iya haɓaka aikin gungu na Kafka cikin sauƙi ta hanyar ƙara sabbin dillalai da kiyaye ma'auni na ɓangarori (wannan yana tabbatar da cewa an rarraba kaya daidai tsakanin dillalai). A cikin yanayinmu, bayan ƙara dillali, kayan aikin tari ya ƙaru zuwa ~ 580 Mb/s (~ saƙon miliyan 1,1 a sakan daya). Ci gaban ya juya ya zama ƙasa da yadda ake tsammani: wannan an bayyana shi ne ta hanyar rashin daidaituwa na ɓangarori (ba duk dillalai suna aiki a kololuwar ƙarfin su ba).
Amfanin ƙwaƙwalwar ajiya na injin JVM ya kasance ƙasa da 2 GB:
Ayyukan dillalai tare da tuƙi ya shafi rashin daidaituwa na ɓangarori:
binciken
Za a iya faɗaɗa tsarin maimaitawa da aka gabatar a sama don rufe ƙarin hadaddun yanayin yanayin da ya shafi ɗaruruwan masu siye, rarrabuwa, ɗaukakawar birgima, sake kunna kwafsa, da sauransu. Duk wannan yana ba mu damar yin la'akari da iyakokin iyakoki na gungu na Kafka a cikin yanayi daban-daban, gano matsalolin da ke cikin aikinsa da kuma nemo hanyoyin magance su.
Mun tsara Supertubes don sauri da sauƙi tura gungu, daidaita shi, ƙara / cire dillalai da batutuwa, amsa faɗakarwa, da tabbatar da Kafka gabaɗaya yana aiki daidai akan Kubernetes. Manufarmu ita ce ta taimaka muku mayar da hankali kan babban aiki ("ƙirƙira" da "cinye" saƙonnin Kafka), kuma ku bar duk wani aiki mai wuyar gaske ga Supertubes da mai aiki na Kafka.
Idan kuna sha'awar fasahar Banzai Cloud da ayyukan Buɗaɗɗen Madogara, yi rajista ga kamfani a , ko .
PS daga mai fassara
Karanta kuma a kan shafinmu:
- «";
- «";
- «".
source: www.habr.com