Ever since AI/ML workloads got here into play, firms have been exploring new information container orchestration platforms and slowly transferring away from Kubernetes. The most recent on the checklist is Juspay, a number one funds platform for retailers in India that additionally powers platforms like Namma Yatri.
Juspay’s Hyperswitch, an open-source cost change written in Rust, relied closely on Kafka for pushing occasions. Nonetheless, because the group talked about in its newest weblog publish on GitHub, the choice to transition from Kafka on Kubernetes, or K8s, to Amazon’s EC2 was pushed by the necessity to optimise efficiency, scale back prices, and simplify operations.
“After months of firefighting, we determined to maneuver from Kubernetes to EC2, a transition that improved efficiency, simplified operations, and reduce prices by 28%,” Neeraj Kumar, program supervisor at Juspay, stated within the weblog, highlighting the large price distinction.
After switching to EC2, the fee dropped to $130 from $180 a month per occasion, translating right into a 28% discount.
Whereas Kubernetes initially supplied a stable basis for container orchestration, managing Kafka at scale proved more difficult than Juspay anticipated. Rising infrastructure prices, inefficiencies in useful resource allocation, and auto-scaling points led to a essential reassessment of their infrastructure technique.
What Had been the Challenges?
Juspay’s determination emigrate Kafka from Kubernetes to EC2 ignited a vigorous dialogue on Reddit. Engineers and designers weighed in on the trade-offs of managing stateful workloads in Kubernetes. Whereas it feels like Juspay had deserted Kubernetes fully, in actuality, it solely shifted away from Kafka.
This highlights a broader pattern the place firms usually realise that working databases, queues, or brokers on Kubernetes introduces pointless complexity. One of many main challenges Juspay confronted with Kubernetes was useful resource allocation inefficiencies. Kubernetes dynamically managed assets, however in observe, this resulted in surprising waste.
As an example, after they allotted 2 CPU cores and 8GB of RAM, the precise provisioned assets have been barely decrease—1.8 CPU cores and seven.5GB RAM. Whereas this discrepancy might sound minor, at scale, it contributed to important price overruns. As Kumar put it, “Think about paying for a full tank of gas, however your automobile solely will get 90% of it. Over time, these lacking litres add up.”
That is just like what Christian Weichel, co-founder and CTO of Gitpod, and Alejandro de Brito Fontes, workers engineer on the firm, stated three months in the past. Kubernetes initially appeared like the plain alternative for Gitpod’s distant, standardised, and automatic improvement environments, however scaling up was a difficulty for them as effectively.
Additionally Learn: Why Companies are Quitting Kubernetes
In January 2024, Gitpod started creating Flex, which was launched in October. That is an ongoing pattern the place firms have began utilizing in-house merchandise for the duty. Constructed on Kubernetes-inspired rules like declarative APIs and management idea, Flex simplifies structure, prioritises zero-trust safety, and addresses the particular wants of improvement environments.
Ben Houston, founder and CTO of ThreeKit, a web-based visible commerce platform, illustrated one other of Kubernetes’s challenges in his latest weblog. Houston defined why he shifted from Kubernetes to Google Cloud Run. The first cause for his drifting away from Kubernetes was its complexity and excessive price, which outweighed its advantages for managing infrastructure at scale.
For Houston, Kubernetes required in depth provisioning, upkeep, and administration, resulting in important DevOps overhead. Moreover, its gradual autoscaling usually resulted in over-provisioning and paying for unused assets.
Auto-scaling additionally posed difficulties. Kubernetes’ scaling mechanisms are designed for stateless functions, however Kafka is stateful. As an alternative of seamlessly scaling up when assets ran low, Kubernetes would restart Kafka nodes, resulting in delays in message processing and elevated latency throughout scaling occasions.
Managing these stateful workloads turned an ongoing operational burden, additional complicating Kafka’s stability.
The Shift to EC2
Discussions concerning the challenges of Kubernetes have been ongoing for a while. In a Hacker Information thread about its viability, builders from totally different firms cited a number of explanation why Kubernetes may be cumbersome.
Initially, Juspay relied on Strimzi for Kafka cluster administration, however this answer launched its personal set of points. New Kafka nodes usually did not combine seamlessly, requiring handbook intervention for each scaling occasion. “Managing our Kafka clusters felt like taking part in whack-a-mole—each time we solved one difficulty, one other would pop up,” they famous.
Confronted with these challenges, Juspay opted emigrate Kafka from Kubernetes to EC2, a transfer that allowed for higher management over useful resource allocation, auto-scaling, and cluster administration. As an alternative of counting on third-party instruments, they constructed an in-house Kafka Controller tailor-made to their wants.
This shift enabled the seamless integration of latest Kafka nodes, automated scaling based mostly on real-time workload evaluation, and considerably improved cluster administration with minimal handbook intervention.
Not like Kubernetes, EC2 allowed for exact useful resource allocation. With Kubernetes, provisioning was usually imprecise, resulting in over-provisioning prices. On EC2, they may allocate precisely the CPU and reminiscence wanted, avoiding pointless expenditures. Beforehand, their Kubernetes-based setup price $180 per occasion per thirty days.
Juspay highlighted that whereas Kubernetes is great for stateless functions, stateful workloads like Kafka can introduce pointless complexity. Customized options, equivalent to their in-house Kafka Controller, supplied higher management, automation, and reliability.
The publish Why Juspay Stop Kubernetes appeared first on Analytics India Journal.
