Sushruth

hi, i'm sushruth 👋 cs undergrad interning at a fintech company. outside work, i build things (rn it's a LIMS app for a civil lab.) i write about prod bugs, building from scratch, and the messy in-between. follow along if that sounds like something you wanna read abt!

@devXritesh Precomputation and edge caching using a trie its near zero latency. a trie helps with prefix matching but ranking what shows up uses popularity, trending and your search history

As a developer, Have you ever wondered : You type just "how to" in Google search and it instantly shows full suggestions like "how to make money", "how to cook pasta" etc... There are 8.5+ billion searches globally every day. How is this autocomplete so fast?

We use microservices at work for a fintech app. The complexity overhead involved is no way something that can be handled by a 5-person team. good luck with that. Separate CI/CD pipelines, distributed tracing and cross-service coordination would eat more time than actual feature dev start with a monolith, modularize well, split only when traffic and team size demand it

If microservices enable teams to deploy independently and scale components separately, why not split our 5-person startup into 50 microservices?

So how kafka works is, Kafka has topics. Each topic has a bunch of partitions which are commonly compared with queues but are not exactly queues. Each partition can only be consumed by one consumer. So if you have 3 partitions and scale to 10 consumers, 7 just sit idle doing nothing The true way to improve processing time is to increase partition count along with consumer count.

Your Kafka consumer is processing messages very slowly. Increasing consumer count from 3 to 10 doesn't help. Why not and what's the bottleneck?

It's a combination of message queues, autoscaling and building the applications using event-driven architecture. Message queues help by buffering the incoming traffic, protecting the downstream services from getting flooded Autoscaling can be configured to ensure more instances of the applications are spun up with increasing traffic Load balancing also plays a key role in ensuring that none of the instances have unequal load Unlike sports streaming, Black Friday has predictable traffic spikes so teams pre-warm instances ahead of time, caches loaded, connection pools ready, everything initialized so theres no cold start latency when the spike hits

Interviewer: How does Amazon handle millions of orders during Black Friday without crashing?

@levelsio @steipete telegram having a billion MAU is not surprising at all lol, the rest of the world has been on it for years

SF's tech scene is finally discovering Telegram has 1 billion MAU Which why it's finally in Claude (And of course thanks to @steipete)

@0xlelouch_ Yep they will, as long as the SLA is breached

An interesting question if you are into tech infra and observability! Your service has 99.9% uptime SLA (43 minutes downtime/month). You had 2 outages: 20 minutes and 30 minutes. Customer demands refund. Do they get it?

@mysticwillz adaptive bitrate streaming, the video is pre-encoded in multiple resolutions and the client switches between them in real time based on bandwidth and buffer health

Interviewer: How does Netflix automatically adjust video quality during playback?

built a collaborative notepad using CRDTs and websockets so this is a fun one There are 2 main approaches here 1. OT(Operational Transformation) where a central server transforms conflicting ops to converge (google docs uses this) 2. Conflict-free Replicated Data Types (CRDTs) where the data structure itself stores change type, location and timestamp to merge without conflicts or a central server CRDT is a more modern and, imo, a better way to deal with concurrent editing

You're building a collaborative doc editor like Google Docs. When two users edit the same paragraph simultaneously, one user's changes disappear. How do you handle concurrent editing?

@javarevisited The output: true false part is a variable so concatenation happens in runtime, creating a new object

What prints? String a = "Ja" + "va"; String b = "Java"; System.out.println(a == b); Now try this: String part = "Ja"; String c = part + "va"; System.out.println(c == b); Why different results?

stateless app servers are easy to scale because theres no shared state but databases hold state and keeping that consistent across nodes is the real problem. read scaling is doable with replicas but write scaling is painful since every node needs to agree on write order. sharding helps but cross shard queries add a whole new layer of complexity

Interviewer: If horizontal scaling is easy with cloud platforms, why is scaling databases still hard?

@yongfook honestly if someone can ship(and maintain) working products with vibe coding who cares how they got there.

There's a new type of dev. They don't get dev humor because they only know vibe coding. And they think they are experts because they have the latest .md files from wherever. Insufferable.

@_khusheyyy gm ^^

goodd morningg y'all <3

@devXritesh in my own projects, not yet but its on my radar, currently most of my endpoints dont paginate deep enough for it to hurt but i know itll bite eventually when data grows

@TreeApostle Offset pagination silently kills perf by forcing the DB to scan + discard every prior row (pure O(n) at depth). Cursor-based (e.g. `WHERE id > :last_id LIMIT 20`) stays O(1) index-fast forever. Switched any of your APIs yet?

System Design Series - Day 4/90 API RESPONSE DESIGN: Pagination and Filtering You built an API. User asks: Get all orders. You return 100,000 records. Result: API dies, 30-second delay, memory explodes, user leaves. Thread below 👇 #SystemDesign #APIPagination

@SumitM_X Looks correct to me? To remove during iteration, iterator.remove() after calling next() is safe unlike list.remove() inside a for each loop.

What's the issue with this code : Iterator<String> it = list.iterator(); while (it.hasNext()) { if (it.next().equals("B")) it.remove(); }

@0xlelouch_ it sucks in Hyd as well, thats approx the same time it takes for my commute as well. 2 hrs gone for commute which couldve been spent more in gym and hobbies. Or just getting a better sleep :(

Tuesday in Bangalore as a software dev: WFH vs WFO and money saved/spent. 1. WFO - Commute: 2 hrs/day - Fuel/metro + parking: ₹150–₹300/day - Lunch + coffee: ₹250–₹500/day - “One quick smoke/tea”: another ₹100 2. WFH - Extra power + AC: ₹60–₹120/day - Internet: ₹20/day - Lunch at home: ₹80–₹150 Net: WFH saves ~₹300–₹900/day and 2 hours. If you go office 12 days/month, that’s ₹3.6k–₹10.8k and ~24 hours gone. Track it for 2 weeks atleast. Then decide what culture is worth to you!

@javarevisited collaborative filtering mostly, find users with similar listening patterns and recommend what they liked that you havent heard yet. they also analyze the actual audio features of songs and layer that on top to handle new tracks with no listens

Interviewer: How does Spotify recommend the perfect song for you?

@levelsio the real servitude is thinking there's only one path, most people never even consider that they have options outside the default script

People born in servitude these days Your primary goal should be to exit servitude

@javarevisited Refactor a legacy service saving the team hours every week and nobody notices but fumble one leetcode medium and suddenly I'm "not a strong candidate"

Software engineering is the only career where real impact isn’t always measurable… but interview performance is brutally measurable.