2017 / 616 pages
Designing Data-Intensive Applications
Martin Kleppmann
Why it's here
Replication, partitioning and consistency, explained from first principles. The trade-offs behind read replicas and an outbox.
Senior .NET engineer, Kathmanduopen to senior and lead roles
Ejan
Shrestha
I build the .NET backends that insurance and payments in Nepal run on, and keep them running in production. Right now that's a 50-project insurance platform and the portals its customers buy and claim through.
Or write to sthaejan00@gmail.com
Download the portfolio as a book(PDF)A simplified model of the platform: three replicas per service with autoscaling, Postgres read replicas, Redis, and RabbitMQ with an outbox and a saga. Tap a pod to kill it, or replay a transaction.
- .NET projects
- 50
- services rebuilt on .NET 8
- 9
- HttpClient call sites fixed
- 397
- portal API endpoints
- 298
- C#
- ASP.NET Core
- EF Core
- Dapper
- Hangfire
- gRPC
- SignalR
- Refit
- IdentityServer
- YARP
- PostgreSQL
- Views and dblink
- Partitioning
- SQL Server
- Redis
- MinIO
- Query tuning
- Migrations
- Docker
- Kubernetes
- Nginx
- Traefik
- Linux
- Let's Encrypt
- Serilog
- Seq
- Grafana Loki
- Correlation IDs
- Next.js
- React
- TypeScript
- Tailwind
- Strapi
Projects
The project book.
Eleven things I've built or run, from the insurance platform I look after at work to my own products.
Chapter 01 of 11A 50-project insurance platform
Deep dive4Swipe to turn the page
Deep dives.
Six problems from the projects above, written the way I'd brief a new engineer. Open one to see what broke, what I changed and what it bought us.
397Insurance platformEnding socket exhaustion across the platformnew HttpClient() call sites retired
- Problem
- Under load, outbound calls failed with address-already-in-use errors, and the underwriting pod's database connections broke with them.
- What I did
- Traced it to
new HttpClient()created on every call and left in TIME_WAIT. Replaced 397 call sites in 123 files with IHttpClientFactory named and typed clients, and gave every client the correlation-ID handler. - Result
- Occurrences went from 414 to 17, all of them intentional or commented out, and the error class left production logs.
1 IDInsurance platformOne trace from gateway to databasefollows a request across every service
- Problem
- Every service carried its own copy of logging code, and the .NET Core 2.2 services logged almost nothing useful.
- What I did
- Moved Serilog, Seq and correlation-ID code into one shared library, built a mirror of it for the 2.2 services with global exception middleware, and put the partner API and IdentityServer on it.
- Result
- Every response carries an X-Correlation-ID, one ID follows a request end to end in Seq, and errors return a safe body instead of internals.
0 → 13Insurance platformFrom end-of-life .NET to .NET 8.NET 8 projects in the platform
- Problem
- When I joined in-house, 36 of the platform's 37 projects ran on .NET Core 2.2, 3.1 or 5, all out of support.
- What I did
- Built new .NET 8 versions of nine services, among them accounting, claims, reinsurance, CRM and underwriting, moved Hangfire to .NET 8, and started the partner API's move to .NET 10.
- Result
- The platform went from zero to thirteen .NET 8 projects, and accounting, claims and reinsurance features now ship on the new services.
5 minDigital walletPaying a remittance out only onceclaim lock on every remittance payout
- Problem
- Two agents could look up and pay out the same cash-pickup remittance at the same time, and the money would leave twice.
- What I did
- Gave every remittance a five-minute claim: the first agent to track it locks it and others are turned away, a background job releases stale locks, and back-office staff can unlock one by hand. Partner payouts use single-use pay tokens that expire.
- Result
- A remittance can be claimed by one agent at a time and paid out only once.
1×Insurance platformBackground jobs that run oncejob registration at startup, not per request
- Problem
- Recurring jobs were scheduled from request handling, so every request re-registered all of them, and full-year report exports timed out inside the web request.
- What I did
- Moved scheduling into a hosted service that runs once at startup. Built background exports on a dedicated Hangfire queue with concurrency caps, streaming rows to Excel or CSV and storing files in MinIO for seven days.
- Result
- Jobs register once, and large reports finish in the background instead of timing out.
3Online portalsPayments partners can't send twicepayment gateways behind one interface
- Problem
- The portals needed eSewa, Khalti and ConnectIPS, and partners retrying a request could have created the same policy twice.
- What I did
- Put the three gateways behind one payment interface, and gave the merchant API HMAC request signing and idempotency keys stored per partner, method, path and key, with expiry.
- Result
- Adding a gateway doesn't touch purchase code, and a retried partner request returns the first result instead of acting again.
Experience
From first commit to platform owner.
BSc Computer Science and Information Technology
Vedas College, 2017 – 2022
Jan 2025 – presentnow
Senior Software Engineer
Himalayan Everest Insurance, Kathmandu, hybrid
- Own backend reliability and modernisation for an insurance platform of 50 .NET projects on Kubernetes, across policy, claims, accounting, reinsurance and reporting.
- Rebuilt nine services on .NET 8 and moved Hangfire to .NET 8; replaced 397 per-call HttpClients with IHttpClientFactory to end socket exhaustion.
- Main backend author of the online portals: 298 endpoints for customers, agents, corporates, merchants and staff, with eSewa, Khalti and ConnectIPS payments.
- Built one Serilog and Seq logging library with correlation IDs; reporting on PostgreSQL views, dblink and partitioned tables; hands-on with the production Kubernetes cluster.
- Integrate and review the team's work: about 320 teammates' branches merged from some 13 developers since 2025.
Jul 2022 – Nov 2024
.NET Developer
EKbana Solutions, Lalitpur, on-site
- Vendor engineer on the insurance platform: co-insurance accounting, receipts, trial balance optimisation, and goAML anti-money-laundering reports built in PostgreSQL and read with Dapper.
- Backend engineer on a digital wallet and agent network (2,119 commits); designed its partner API and stopped double remittance payouts with a claim lock.
- Reviewed and merged 165 merge requests from seven teammates, and promoted releases through qa, uat and live.
- Built a QuickBooks Online integration for an accounting firm, and the base API (auth, permissions, Hangfire, audit and request logging, rate limiting) later used for the insurer's portals.
Jan 2022 – Jun 2022
.NET Developer
NBIT, Lalitpur
- Built and maintained .NET Framework applications for client projects.
Anatomy of an incident
The day the ports ran out.
A real production incident on the insurance platform, replayed as the trace I followed. Step through it: what users saw, how I found the cause, and what the fix changed.
POST/api/underwriting/policies
trace 7f3a…c91e
500
2,140ms
- ingress POST /policies2,140ms
- └underwriting issue policy2,120ms
- └http → accounting2,030ms
- └postgres open connection52ms
- └serilog → seq12ms
Step 1/5
What users saw
Under load, issuing a policy started failing with 500 errors, and every retry made it worse.
Read the deep diveHttpClient, done properly
Timings are illustrative; the chain of failures is the real one.
Journey
From the controller to architecture, and back.
One feature, issuing a policy, written the way engineers write it from their first year to the architect's chair. Pick a stage.
Make it work.
Everything lives in the controller: SQL, business rules, email, all of it. It works on my machine and nobody knows why.
Controllers/PolicyController.cs
[HttpPost("save")]
public IActionResult Save(PolicyDto dto)
{
var conn = new SqlConnection("Server=prod;User=sa;Password=123");
conn.Open();
var sql = "INSERT INTO Policy VALUES ('" + dto.Name + "', " + dto.Premium + ")";
new SqlCommand(sql, conn).ExecuteNonQuery();
if (dto.Premium > 100000)
new SmtpClient("mail").Send("boss@co", "Big policy!", dto.Name, "");
return Ok("saved"); // works on my machine
}stage 01/07
Writing
Notes from production.
14 articles on .NET, infrastructure and running a team. Most started as an incident.
Places
Where I've been.
Every place I've lived, worked or travelled to. Mostly the Kathmandu valley so far, so the map has plenty of room.
Drag to spin the globe
2 places in 1 country so far
Bookshelf
What's on the shelf.
Books that shaped how I build and run systems. Pull one off the shelf.
Contact
Got a system that needs a steady hand?
Open to senior and lead roles in backend architecture and .NET modernisation, in Kathmandu or remote.
PDF, A4, every project and article in one file
Local time in Kathmandu: KTM, UTC+5:45
ejan@kathmandu: ~
➜ ~ whoami
Ejan Shrestha, Senior Software Engineer, Kathmandu, Nepal
➜ ~ contact --email
➜ ~