Faiz Halde

Breadcrumbs: Distributed System

Tue, 23 Jun 2026 10:40:42 +0200

What’s the difference between Fault tolerance, Replication, Consistency & Consensus?

These four concepts are tightly coupled and form a causal chain.

Fault tolerance is the goal: keep serving correctly even when machines crash or the network breaks.

Replication is the mechanism: to not lose data when one machine dies, keep copies on several machines.

Consistency is the problem replication creates: once you have copies, they can disagree. What is a client allowed to observe? (Always the latest write? Maybe a stale read? Never a “torn” state?)

Consensus is the algorithm that solves it: getting N machines to agree on one value / one ordering of operations, despite some of them failing or messages being lost.

How Feldera handles moving windows

Tue, 16 Jun 2026 07:21:25 +0200

Feldera is an incremental SQL engine. You give it a query, and instead of running it over and over, it efficiently keeps the result continuously up to date as the underlying data changes.

As I keep coming across more of SQL, I realized the language is very expressive. That’s great for an analyst, and for a traditional batch engine it’s just another query to execute, but anyone around the incremental view maintenance space would wonder “hey, how does this work?”.

Sizing Your LLM Inference Cluster

Wed, 27 May 2026 11:26:15 +0200

GPU capacity planning boils down to two questions: what’s the theoretical minimum GPU count for my workload, and how much extra capacity is needed once queueing and workload variability are accounted for?

In this post, I’d like to give an overview of this methodology. Here’s a toolkit that automates most of it: howmanygpus

Start with model facts

The following references were used to derive the formulas in the toolkit:

For a model, the first useful quantities are the weight footprint and KV-cache cost per token.

Foo ≠ Hostname

Tue, 17 Mar 2026 14:59:34 +0100

A common recurring pattern I’ve observed is generating ingress hostnames from app names:

app: foo
hostname: ${app}.<cluster-domain>

It guarantees uniqueness and simplifies setup. For internal services or ephemeral environments, this is often perfectly fine.

But when this pattern becomes the default for externally visible endpoints, it starts to break down — because it fundamentally misunderstands what a hostname represents in that context.

A hostname is not just a unique identifier. It’s a contract — with browsers, with TLS, with DNS, and with your users. By reducing it to a derived label, you lose control over things that matter in production.

Making JSON Awaitable

Sun, 01 Mar 2026 22:05:50 +0100

When working with LLMs for building Agentic Apps, one thing quickly becomes obvious:

You want structured outputs.
The model streams tokens.
Your code waits for a finished JSON blob.

That mismatch accumulates latency. jsontap closes that gap.

This post explains the design behind jsontap.

JSON is a tree

JSON is not just a string. It is a hierarchical structure.

{
 "user": {
 "name": "Alice",
 "scores": [10, 20, 30],
 "friends": [
 {
 "name": "Bob",
 "email": "bob@example.com"
 }
 ]
 }
}

As a tree: