I'm a backend engineer who builds Kubernetes controllers and APIs in Go. I write and review API changes for kubernetes-sigs/node-readiness-controller (Kubernetes SIG Node), and I'm the Lead Mentor for its Headlamp plugin in the Linux Foundation Mentorship (LFX). At Kotak Mahindra Bank, I own production on-call for a Node.js backend.
I settle a design before I write code: I put the constraint in numbers, look at how upstream projects solved the same problem, and pick what fits. Much of my upstream time goes into reviews.
node-readiness-controller: the Kubernetes SIG Node controller that keeps a node tainted until its readiness conditions are met.
- Added
defaultStatusto the NodeReadinessRule API (#283, shipped in v0.4.1), so a rule can taint a node only when a detector like Node Problem Detector reports a problem, and made the API reject it on bootstrap-only rules (#291). - Designed
evaluationSummary, per-rule node counts at up to 5,000 nodes: node events map to their rules and merge over a 2-second window, the Cluster API/EndpointSlice pattern, because the etcd status write is the bottleneck, not the recount (#502, in review). - Led the review of the
conditionPolicy(allOf/anyOf) API (#315), keeping taint decisions computed from the rule spec, never read back from status. Reviewed the new NodeReadinessEvaluation API and controller (#345).
Headlamp plugin for node-readiness-controller (Lead Mentor, Linux Foundation Mentorship): wrote the spec, reviewed all 40+ applications to pick our mentee, and led the plugin from spec to its first public release in 8 weeks.
OpenELB (Linux Foundation Mentorship mentee, 2022): built the Go REST API behind its web console inside the controller manager (#290), and fixed BGP peers with addPaths but no address family never coming up (#311).
Subscription Management Service: a Go backend for recurring subscriptions, built so a subscription only moves through valid lifecycle states and every request and background task can be traced end to end.
- Lifecycle: HTTP controllers only get an API-facing service interface, so the compiler keeps them from calling renewal or expiry. Every transition checks the current state first.
- Background pipeline: a scheduler enqueues reminder, renewal and expiry tasks on Asynq (Redis) with deduplication and retries. Workers re-check state before acting, so a retried task can't renew twice, and W3C trace context in the task headers keeps scheduler and worker in one trace.
- Fail-open rate limiting: when Redis is down, requests still go through. The error is recorded on the request's span, and an atomic
CompareAndSwapguard logs it at most once a minute. - Tests against real MongoDB: Testcontainers, seeded with decoy documents that a missing filter clause or the wrong sort order would return instead of the target.
Go & Kubernetes: controller-runtime, kubebuilder, CRDs, CEL
- node-readiness-controller: taking
evaluationSummary(#502) through review to a release, and NodeReadinessEvaluation support in the Headlamp plugin. - Task chaining and dead-letter queues in Asynq: running side effects like emails as their own tasks, apart from billing.
- Shutdown order: reverse-order shutdown through a dependency graph (
uber-go/fx), so connections drain cleanly on SIGTERM.