Menu • About Cookpad Global • Search-v2 and ML APIs • Gaps: ideal and reality • Designing fault-tolerant microservices with SRE and circuit breaker centric architecture SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 3
Out of scope • Monolith vs SOA vs Microservices • So2ware design and development in Cloud Na• Container orchestrators: Why ECS? Why EKS(k8s)? • Explanabudget) SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 4
Cookpad Global by numbers • 1 monolith + 7 microservices in produc5on • 300+ spot instances for ECS clusters • 400+ deployments per ECS task defini5on per day • 20 deployements to produc5on per day 7
See more details on Speaker Deck ... 1,2 2 Cookpad TechConf 2019, Challenges for Global Service from a Perspec>ve of SRE ~ 2nd season ~ 1 Cookpad TechConf 2018, Challenges for Global Service from a Perspec>ve of SRE SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 9
Search-v2 and ML APIs • Search-v2: people can meet their favorite recipes for cooking • (e.g) Personalized search, visual search, [email protected] • ML APIs: Other APIs can provide machine learning integrated features • (e.g) Image enhancement, image to recipe SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 16
As-Is Developers use restricted technology stack To-Be Search/ML team can use mainstream technology stack for their fields SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 41
As-Is Produc'on is down due to outages of new microservices To-Be No produc)on outages due to low service level microservices SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 48
Does team have enough capacity for on-call? "Assuming that there are always two people on-call (primary and secondary, with different du:es), the minimum number of engineers needed for on-call duty from a single-site team is eight: assuming week-long shi?s, each engineer is on-call (primary or secondary) for one week every month." 4 "For produc7on on-call responsibili7es, I’ve found that two-7er 24/7 support requires eight engineers. As teams holding their own pagers have become increasingly mainstream, this has become an important sizing constraint, and I try to ensure that every engineering team’s steady state is eight people" 5 5 Larson, Will. An Elegant Puzzle: Systems of Engineering Management, 2.1 Sizing teams (p.33) 4 Google - Site Reliability Engineering Chapter 11 - Being On-Call SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 51
As-Is People have to be responsible for on-call rota0ons for new mircorservices To-Be New search/ml team must be free from on-call pressures for their new microservices SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 52
As-Is Many teams need tough nego)a)ons to release ML related features To-Be ML team can release experimental features with light process in produc0on SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 55
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Produc6on outages due to new microservices No [email protected] outages due to low service level microservices People have to be responsible for on- call rota6ons for new mircorservices New search/ml team must be free from on-call pressures for their new microservices Many teams need tough nego6a6ons to release ML related features ML team can release experimental features with light process in [email protected] SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 56
Design Docs • Reach consensus against scopes and expecta2ons 6 • In Cookpad, only SRE team knows en2re system designs 7 7 Google, The Site Reliability Workbook, Chapter 7 - Simplicity 6 Google, Site Reliability Engineering, Chapter 31 - CommunicaSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 59
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document Produc=on outages due to new microservices No produc=on outages due to low service level microservices Design document People have to be responsible for on-call rota=ons for new mircorservices New search/ml team must be free from on-call pressures for their new microservices Design document Many teams need tough nego=a=ons to release ML related features ML team can release experimental features with light process in produc=on Design document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 62
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document + ? Produc8on outages due to new microservices No produc8on outages due to low service level microservices Design document People have to be responsible for on- call rota8ons for new mircorservices New search/ml team must be free from on-call pressures for their new microservices Design document Many teams need tough nego8a8ons to release ML related features ML team can release experimental features with light process in produc8on Design document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 63
Dedicated AWS account • Use AWS Organiza0ons to issue new AWS account • Design network by SRE • Build VPC peering between new and old VPCs SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 69
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document Delega1on and resource isola1on Produc=on outages due to new microservices No produc=on outages due to low service level microservices Design document People have to be responsible for on-call rota=ons for new mircorservices New search/ml team must be free from on-call pressures for their new microservices Design document Many teams need tough nego=a=ons to release ML related features ML team can release experimental features with light process in produc=on Design document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 73
Don't accept excep,ons • We only have 3 SREs (in 2019) • Follow the boundary we define in the design document • Don't share servers managed by SRE team • Use SaaS to accelerate minimum product development cycles 8 • e.g: CI • e.g: Observability 8 Prac'cal Monitoring: Effec've Strategies for the Real World, Chapter 2.3 PaAern #3: Buy, Not Build SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 74
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No produc2on outages due to low service level microservices Design document + ? People have to be responsible for on-call rotamircorservices New search/ml team must be free from on-call pressures for their new microservices Design document Many teams need tough negofeatures ML team can release experimental features with light process in producDesign document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 75
Why Circuit Breaker? • Fail fast strategy to prevent cascading failures • Limits external service and network impacts • Don’t waste capacity calling a broken service • External service is slow • External service is down • Network is unstable SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 80
• Closed • Traffic flows normaly • Health is assessed every 100ms based on a 10s rolling average • Open / Tripped • Fail fast - return 503 error • Stays in this state for 10s • Recovering / Half Open • Ramp up traffic over 10s • Check health every 100ms -> if fail go back to Open state • Return to Closed if health is OK aJer 10s SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 84
Circuit Breaker - Implica1ons • We can introduce experimental and new services with less risk to other parts of the applica8on • Slow responses ~= Outage! • Fallback strategies become more important • Add values to use SLOs for communica8on tools about service levels SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 85
Circuit breaker proxy side-car container • Use a L7 reverse proxy with circuit breaking middleware • Each microservice has it's own independently configured circuit breaker • Run as a sidecar container SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 87
Availability class • We customize produc0on readiness check as availablity class (a.k.a produc0on readiness review 9) 9 Google - Site Reliability Engineering, Chapter 32 - The Evolving SRE Engagement Model SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 96
Implemen'ng alerts on SLO There are several strategies to implement alerts on SLO 10 • Target Error Rate ≥ SLO Threshold • Increased Alert Window • Incremen• Alert on Burn Rate • Mul• Mul10 Google - The Site Reliability Workbook, Chapter 5: AlerSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 105
Implemen'ng alerts on SLO There are several strategies to implement alerts on SLO 10 • Target Error Rate ≥ SLO Threshold • Increased Alert Window • Incremen• Alert on Burn Rate • Mul• Mul$window, Mul$-Burn-Rate Alerts 10 Google - The Site Reliability Workbook, Chapter 5: AlerSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 106
Burn rates and +me to complete budget exhaus+on 10 Burn rate Error rate for 99.9% SLO Time to exhaus8on 1 0.1% 30 days 2 0.2% 15 days 10 1% 3 days 1000 100% 43minutes 10 Google - The Site Reliability Workbook, Chapter 5: AlerSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 108
Burn rates and +me to complete budget exhaus+on 10 10 Google - The Site Reliability Workbook, Chapter 5: AlerSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 109
Mul$window, Mul$-Burn-Rate Alerts 10 • This approach provides good precision alerts and reduce the number of false posi7ves • Make the short window 1/12 the dura7on of the long window as the star7ng point Severity No*fica*on Long window Short window Burn rate Error budget consumed Cri$cal Pager 1 hour 5 minutes 14.4 2% Cri$cal Pager 6 hour 30 minutes 6 5% Warning Chat, $cket 3 days 6 hours 1 10% 10 Google - The Site Reliability Workbook, Chapter 5: AlerSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 110
Implemen'ng Prometheus configs in Jsonnet • Jsonnet11 is a data templa0ng language • Simple extension of JSON • Eliminate duplica0on with object-orienta0on 11 google/jsonnet: Jsonnet - The data templa5ng language SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 116
Cau$on! • Jsonnet is super powerful language to elimiate redundancy • Too DRYed-configura• We have to control the power and make configuraSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 120
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No producservice level microservices Design document SLO Circuit breaker People have to be responsible for on- call rotaNew search/ml team must be free from on-call pressures for their new microservices Design document Many teams need tough negoto release ML related features ML team can release experimental features with light process in producDesign document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 123
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No producservice level microservices Design document SLO Circuit breaker People have to be responsible for on- call rotaNew search/ml team must be free from on-call pressures for their new microservices Design document + ? Many teams need tough negoto release ML related features ML team can release experimental features with light process in producDesign document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 124
Fallback to search-v1 when circuit breaker is open • Proxy par*al requests to search-v2 in feature toggle • Strict circuit breaking threshold (No SLO or extreamely low SLO) and fail fast when upstream is unstable • Rescue all errors in feature toggle • Fallback all requests to search-v1 when circuit breaker returns 503s SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 127
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No producservice level microservices Design document SLO Circuit breaker People have to be responsible for on- call rotaNew search/ml team must be free from on-call pressures for their new microservices Design document SLO Circuit breaker + Fallback Many teams need tough negoto release ML related features ML team can release experimental features with light process in producDesign document SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 133
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No producservice level microservices Design document SLO Circuit breaker People have to be responsible for on- call rotaNew search/ml team must be free from on-call pressures for their new microservices Design document SLO Circuit breaker + Fallback Many teams need tough negoto release ML related features ML team can release experimental features with light process in producDesign document + ? SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 134
Prod endpoint + feature toggle + path-based rou5ng • Specify shared single ML API endpoint in feature toggle • Strict circuit breaking threshold (No SLO) and fail fast when upstream is unstable • Rescue all errors in feature toggle and dismiss • Change desDnaDon for each ML API based on request path SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 138
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No producservice level microservices Design document SLO Circuit breaker People have to be responsible for on- call rotaNew search/ml team must be free from on-call pressures for their new microservices Design document SLO Circuit breaker + Fallback Many teams need tough negorelease ML related features ML team can release experimental features with light process in producDesign document SLO (No SLO) Circuit breaker Feature toggle + Path-based rouAng SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 142
Goals for the SRE team As-Is To-Be Approach Developers use restricted technology stack Search/ML team can use mainstream technology stack for their fields Design document DelegaProducmicroservices No producservice level microservices Design document SLO Circuit breaker People have to be responsible for on- call rotaNew search/ml team must be free from on-call pressures for their new microservices Design document SLO Circuit breaker + Fallback Many teams need tough negorelease ML related features ML team can release experimental features with light process in producDesign document SLO (No SLO) Circuit breaker Feature toggle + Path-based rouSRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 143
SRE exper)se and circuit breaker • Protect microservices from unreliable microservice • Enforce contracts(alignment) among teams • Provide on-call free environment for new team • Enable developers to release experimental features • Reduce unproduc=ve communica=on among teams SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 146
What is the best on-call rota0on? • It really depends on your team members • Someone loves weekly rota6on • Someone loves daily rota6on • Someone loves on-call on weekends • Don't create organiza6on-wide rota6on rule 13 13 Well designed policy about on-call compensa6on is necessary to achieve this SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 148
On-call rota+on in SRE team • Hybrid strategy to use advantages of 3me-zone differences • JP(UTC+9) & UK(UTC+0) business hour shiF • Daily off-hours rota3on SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 151
How can we introduce SRE in organiza3on? If you tackle to introduce the SRE methodology and culture with bo9om-up approaches, • Start from a small thing • Find your buddy from product develop teams who are happy to support your ideas • Provide incen;ve to your product developers • SREs are responsible for primary on-call if your services achieve your SLO standard (e.g: 99.99 % avaiability) for a month • Find win-win strategy for developers and SREs • Don't throw SRE sales pitch • Don't play "SRE is one of the Google best prac;ces" cards • We should seriously provide benefits to organiza;on with SRE methodologies (Why do we need SLO? What benefits do we have?) SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 154
Achievements • Improvement of produc0on stability • Apply SRE technique to real service • Release of machine learning integrated search in produc0on 15 • Release of machine learning oriented infrastruture 15 Vector scoring for term embeddings in Elas5csearch - Speaker Deck SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 155
What's next? • Promote SRE culture with ba4le-tested methodologies • Providing JWT auth endpoint for ML and other microservices • Machine learning researchers want to provide services that will be consumed by beta builds of mobile applicaCons • Monolith doesn't need frequent code changes for ML experiences • Monolith doesn't have to proxy anything (this sounds worry SRE NEXT 2020 (Jan 25, 2020) / Takayuki Watanabe 156
takanabe
nenonaninu
taka2noda
takagimeow
niryuu
tkhresk
yoshikikomoriya
startree
kaga
kawaguti
sat
moritamasami
newrelic2023
eileencodes
iamctodd
pauljervisheath
bluesmoon
mza
chrislema
philhawksworth
caitiem20
robhawkes
thekraken
chriscoyier
chrisshort