Kubernetes & Autoscaling
Designing how workloads scale on EKS: Karpenter for nodes, KEDA for Kafka consumers, HPA and VPA for the rest, and fixed capacity where the load is steady. I also own the version upgrades, AMIs and launch templates.
At Polaris Smart Metering I build and look after the AWS and OCI platforms that state electricity utilities across India use to collect and manage smart-meter data.
I'm a DevOps engineer, and I lead a small infrastructure team. Most of my work is in the small decisions that keep a platform healthy: how a cluster scales, where a database should run, which alert is worth waking someone up for, and what a change will cost before it ships. I specialise in Kubernetes autoscaling, Kafka and MQTT, multi-account AWS and FinOps, and I pick up whatever else a project needs on AWS, OCI or open source. I care most about systems that are easy to operate and cheap to run.
how a meter reading moves through the platform
Where I spend most of my time, across AWS and OCI.
Designing how workloads scale on EKS: Karpenter for nodes, KEDA for Kafka consumers, HPA and VPA for the rest, and fixed capacity where the load is steady. I also own the version upgrades, AMIs and launch templates.
Running the messaging layer for IoT traffic: EMQX clusters for device connections and Kafka for everything after. I've moved Kafka from EC2 to MSK and from MSK to Strimzi, and I plan moves like these so they can be rolled back.
Treating cost as part of the design. I go through the AWS bill every month, right-size, buy reserved capacity, set retention policies, and switch off what isn't being used.
Setting up multi-account AWS so it holds up as projects are added: an OU layout, SCP guardrails, SSO instead of IAM users, cost tags, VPCs per tier, and site-to-site VPN into on-prem data centres. Some projects also run on OCI.
Deciding where data should live and how it comes back after a failure: Aurora and self-managed PostgreSQL, MySQL for older apps, DR with pgBackRest, and tuning for write-heavy workloads.
Building alerting people can act on: alarms from every account routed to one place, paging tuned so it means something, infra on-call, and a postmortem after each incident.
Work I've designed or led, and what came out of it.
Planned and ran the move of our Kafka clusters from Amazon MSK to Strimzi on our own Kubernetes clusters, with gzip as the cluster-wide compression since most producers already used it.
During the cutover, a few older applications that relied on lz4 started to fall behind. Instead of rolling back, I built a small transcoder so they could keep consuming, and the migration went ahead.
Cluster-wide changes now start with a compatibility check across every producer and consumer.
Designed the move of the head-end services from static node groups to Karpenter, with KEDA scaling Kafka consumers on lag, HPA for request traffic and Goldilocks-guided VPA for sizing. I kept the meter-data workloads on fixed capacity on purpose, because their load is steady and moving them around only adds risk.
Led the first round of a cost cleanup across our AWS accounts: retention policies on 500+ ECR repositories, 167 unused pipelines, build projects and registries removed after backing each one up, an idle node group switched off, and reserved instances for Aurora.
Moved an old application running on Python 2.7 and MySQL 8 onto smaller, cheaper infrastructure, with no change for the people who use it.
Reworked a pipeline that rebuilt and redeployed the whole platform on every commit. It now maps changed paths to services, builds only those, and deploys across accounts with shared IAM roles.
Upgraded every cluster from 1.33 to 1.35 before extended-support pricing started, working through custom AMIs, launch template versions that didn't apply as expected, and the IMDS hop limit.
Designed disaster recovery for self-managed Postgres on EC2 with pgBackRest physical and logical backups instead of streaming replicas, so recovery stays predictable without paying for a second database. The S3 path is locked so backups can't be deleted.
Designed the OU layout, SCP guardrails, tag policy and SSO permission sets, and moved the team off IAM users. I kept cost tags report-only on purpose, because enforcing them would have blocked node launches in the middle of a scale-up.
Built a capacity model for a 10-million-meter rollout from 14 days of production data, instead of multiplying today's numbers. It showed that pods and Kafka partitions scale out fine and the database writer doesn't, because its load is almost all writes.
Built and host this site on Cloudflare's free plan. Every script is pinned and hashed under a strict CSP, the background is rendered once and served as video, and a daily GitHub workflow checks the live site from outside.
View on GitHub ↗Issues I root-caused on call, and the fixes that followed.
Services across one cluster kept dropping their Postgres connections with "unexpected EOF", which pointed at the network. I traced it to the database host instead: strict memory overcommit with no swap left Postgres about 140 MB to start new connections. Fixed with one kernel setting and no restart.
A consumer group kept rebalancing and lag reached 20 million messages. VPA had cut the decoder pods from 500m to about 53m CPU, so batches ran past the poll timeout and the consumers were dropped from the group. I pinned CPU at 200m and moved VPA to recommendation-only mode.
After a full outage no broker would rejoin, because the boot script stopped any node that couldn't find a peer. I bootstrapped one node manually, joined the rest to it and restored the configuration. The investigation also showed the same check had been producing empty 187-byte backups.
Three years at Polaris Smart Metering. I joined as an intern and now lead the infrastructure team.
Each one links to its Credly record.
Longer notes from production work: what broke, why it broke, and how it was fixed.
If you're working on something similar, or hiring for it, email is the quickest way to reach me.
Sushant Nagil