Milind Dethe's personal site — posts, projects, and build logs from the things I build, break, and automate.https://milinddethe15.tech/en-ushttps://milinddethe15.tech/posts/understanding-backend/finale/https://milinddethe15.tech/posts/understanding-backend/finale/The checkout worked. So I put four copies of it in containers behind a load balancer, kept traffic flowing, and tried to change it. The first deploy was an outage, a column rename broke a perfect rolling deploy, and the rollback made it worse. This is what production turned out to mean.Thu, 17 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/observability/https://milinddethe15.tech/posts/understanding-backend/observability/Every checkout returned 200 and every order shipped. Then a user said it was slow sometimes, and I couldn't tell where. Then I killed a process and nothing on the request path noticed. Logs, metrics, traces and one alert, in the order I needed them.Wed, 16 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/correctness/https://milinddethe15.tech/posts/understanding-backend/correctness/I saved the order on one line and published the event on the next. Then I killed the process between them, swapped them, and put the publish inside the transaction. Nothing fixed it, until the event became a row.Mon, 14 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/distributed-systems/https://milinddethe15.tech/posts/understanding-backend/distributed-systems/I put a breakable network between two services and asked one to charge a card. A timeout looked the same whether the payment never happened, already happened, or was still happening. That changed how I think about distributed systems.Mon, 14 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/queue/https://milinddethe15.tech/posts/understanding-backend/queue/I moved the confirmation email out of the request and checkout went from 300 ms to under 1 ms. Then I followed one message through crashes, duplicates, dead letters, and backpressure.Sun, 13 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/concurrency/https://milinddethe15.tech/posts/understanding-backend/concurrency/My inventory handler passed every test. Then twenty buyers hit it at the same moment and it sold twenty units from a stock of five. A transaction didn't help. A lock did, until I ran two copies of the app.Sat, 12 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/retries/https://milinddethe15.tech/posts/understanding-backend/retries/My dependency got slow for three seconds. Without retries, users saw three bad seconds. With three retries, they saw eight, and I watched the retries turn a small problem into a bigger one.Fri, 11 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/caching/https://milinddethe15.tech/posts/understanding-backend/caching/My dashboard query was already as fast as it could get. Then everyone asked for it at once, and the database could only answer nine times a second. A cache fixed that in six lines. Then it started lying to me.Thu, 10 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/why-database-get-slow/https://milinddethe15.tech/posts/understanding-backend/why-database-get-slow/I thought a database was a big dictionary. Give it a key, get a row back. So I built a table of a million users, asked for one by email, and watched Postgres read every single row to find it.Wed, 09 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/understanding-backend/journey-of-a-request/https://milinddethe15.tech/posts/understanding-backend/journey-of-a-request/I thought calling an API was client, server, database, response. I traced one request, counted thirteen hops, and measured that most of the time was gone before the server ran a line of my code.Mon, 07 Sep 2026 00:00:00 GMTunderstanding-backendmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day7/https://milinddethe15.tech/posts/mlops/day7/Learn how model serving, monitoring, drift detection, and retraining complete the MLOps lifecycle after deployment.Fri, 19 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day6/https://milinddethe15.tech/posts/mlops/day6/Learn about feature stores, training-serving skew, and why machine learning teams use platforms like Feast to keep training and production data consistent.Thu, 18 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day5/https://milinddethe15.tech/posts/mlops/day5/Learn why machine learning workflows evolve from Python scripts into pipelines and how Kubeflow Pipelines orchestrates reproducible ML workflows on Kubernetes.Wed, 17 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day4/https://milinddethe15.tech/posts/mlops/day4/Learn how machine learning teams improve model performance using cross validation, hyperparameter tuning, AutoML, and parallel experimentation.Tue, 16 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day3/https://milinddethe15.tech/posts/mlops/day3/Learn why machine learning teams need data versioning and experiment tracking, and how tools like DVC and MLflow help make experiments reproducible.Mon, 15 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day2/https://milinddethe15.tech/posts/mlops/day2/Learn why Jupyter notebooks are great for experimentation but insufficient for reproducible machine learning workflows.Sun, 14 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/mlops/day1/https://milinddethe15.tech/posts/mlops/day1/Why DevOps alone is not enough for machine learning and how MLOps addresses data, models, and training-serving skew.Sat, 13 Jun 2026 00:00:00 GMT7-days-of-mlopsmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-7-ai-platform-engineering/https://milinddethe15.tech/posts/day-7-ai-platform-engineering/A complete AI platform: layers, services, infrastructure, and where tools like vLLM, Ray, and Kubernetes fit together.Wed, 10 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-6-ai-platform-engineering/https://milinddethe15.tech/posts/day-6-ai-platform-engineering/Why vLLM exists: batching, PagedAttention, and efficient GPU memory management for LLM serving.Tue, 09 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-5-ai-platform-engineering/https://milinddethe15.tech/posts/day-5-ai-platform-engineering/Why Ray exists: scheduling computation (tasks & actors) vs infrastructure for distributed AI workloads.Mon, 08 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-4-ai-platform-engineering/https://milinddethe15.tech/posts/day-4-ai-platform-engineering/Why AI scheduling differs from traditional apps: gang scheduling, topology awareness, and specialized schedulers.Sun, 07 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-3-ai-platform-engineering/https://milinddethe15.tech/posts/day-3-ai-platform-engineering/Why GPU utilization is hard: resource fragmentation, sharing strategies (MIG, time-slicing, batching), and cost impacts.Sat, 06 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-2-ai-platform-engineering/https://milinddethe15.tech/posts/day-2-ai-platform-engineering/How Kubernetes discovers GPUs: drivers, device plugins, and making GPUs schedulable resources.Fri, 05 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/day-1-ai-platform-engineering/https://milinddethe15.tech/posts/day-1-ai-platform-engineering/Why AI workloads demand different platform engineering: GPUs change bottlenecks and resource priorities.Thu, 04 Jun 2026 00:00:00 GMT7-days-of-ai-platform-engineeringmilinddethe15@gmail.com (Milind Dethe)https://milinddethe15.tech/posts/lfx-mentorship/https://milinddethe15.tech/posts/lfx-mentorship/Wed, 04 Dec 2024 00:00:00 GMTexperienceLFXmilinddethe15@gmail.com (Milind Dethe)