Most engineering content is noise. We filter for what matters: - Backend Architecture - System Design - AI in Production Real decisions. Real trade-offs.
Canada
Joined December 2012
- Tweets6.3K
- Following82
- Followers1.1K
- Likes43
Decoupling to Scale: how Grab moved to Aerospike with zero downtime — plus Meta's 28% latency cut and LinkedIn's 75x GPU speedup.
#SoftwareEngineering #SystemDesign #Aerospike #Rust
the infrastructure decisions you made at moderate scale become active liabilities at production scale.
Scale Finds Every Shortcut You Took open.substack.com/pub/thepra…
thepracticalstack461.substac…
Agentic AI Is Breaking System Boundaries — And Your Org Chart Is Next
Behind the Outages: Redis, CPU Starvation, and Kafka at Petabyte Scale open.substack.com/pub/thepra…
Agent Systems Are Crossing Lines — Your Stack Isn’t Ready
open.substack.com/pub/thepra…
Rate limiting is not about blocking bad actors.
It's about protecting good ones from each other. Most teams pick one without knowing why.
#RateLimiting #APIDesign #SystemDesign
thepracticalstack461.substac…
Most engineers use LLMs daily but can't explain what happens between prompt and response.
Here's the full pipeline — 8 concepts, one image. Save this.
🔁 Tokenization → Embeddings → Attention → Transformer → KV Cache → Sampling → Context Window → Inference Optimisation
![Hello world" enters an LLM as [9906, 1917]. By the time output appears, it's been through tokenization, embeddings, 96 transformer layers, KV cache, and a probability sampler. Most people call this "AI." Engineers should know what's actually happening.](https://pbs.twimg.com/media/HFzIceYbAAEnx0N.jpg?name=orig)
ALT Hello world" enters an LLM as [9906, 1917]. By the time output appears, it's been through tokenization, embeddings, 96 transformer layers, KV cache, and a probability sampler. Most people call this "AI." Engineers should know what's actually happening.
Most engineers pick a caching strategy based on
what they've used before — not what the system
actually needs.
8 strategies. Different tradeoffs. Cache-aside ≠ always the answer.
🧵Which one do you default to ?
#DistributedSystems #SystemDesign
🤖 Made with AI
Grab lost 40% of eng time to repetitive queries.
They didn't build a bigger model. They decomposed it like microservices.
Schema drift hits one agent. Not the whole system.
Boundaries matter more than models.
The Practical Stack 👇
thepracticalstack461.substac…
#DistributedSystems
Tier-1 bank: sub-5ms p99 at 1.6M msg/sec with strict ordering intact.
The unlock wasn't just tuning.
Multi-DC replica placement done right — skip this and tail latency spikes under load.
The Practical Stack 👇 thepracticalstack461.substac…
Your K8s scheduler is blind to GPU utilization.
A job at 3% looks identical to one that finished.
Utilization-aware preemption fixes this — but the idle threshold is the entire bet.
The Practical Stack 👇 thepracticalstack461.substac…
#Kubernetes #GPUComputing
Spring Boot Logging mechanism along with a few of configuration details.
#Java #CoreJava #Spring #SpringBoot #development
buff.ly/32LtNDB
Learn Spring Boot With Ehcache to enhance project performance!!
#Java #Spring #SpringBoot #CoreJAva
buff.ly/32UclMV
