SkyRaun / Studies

Studies.

Raun means experiment and proof, and this is where the proof gets published. Each study takes one question, runs the experiment, and shows the evidence — including the runs that did not go the way I expected. Free, permanently, no signup wall.

evidence, not opinion no email gate every claim reproducible

Latest

http 14 July 2026 · 9 min read

Your HTTP client timed out. Did the request actually fail?

A client gives up at two seconds. The server committed at 1.87. Both behaved exactly as designed, and now two systems disagree about whether a payment happened. This is the cheapest distributed-systems lesson available, and almost nobody learns it before it costs them money.

Read the study →

from the experiment

1.870s  server  committed
2.000s  client  read timeout
2.001s  client  reported: failed
2.410s  server  200 OK → nobody

2 July 2026

TCP gave you bytes. It never promised you messages.

One 20-byte write, an 8-byte read buffer, and three chunks that respect none of your message boundaries. Why recv() is not a message queue.

networking

11 min

19 June 2026

The framing bug that passes code review

The same decoder, the same process, the same buffer size — correct on one connection and wrong on the next. Only the write timing changed. A short study in why "it worked" is not evidence.

networking

7 min

5 June 2026

What does an acknowledgement actually acknowledge?

Received? Persisted? Processed? Replicated? Four brokers, four different answers, and the gap between them is where your data goes missing.

messaging

8 min

21 May 2026

A queue and a retained log are not the same product

You can rewind one of them. That single difference decides your replay story, your consumer model, and how you recover from a bad deploy at 4am.

messaging

10 min

8 May 2026

Visibility timeout: the setting that quietly duplicates your work

Set it below your p99 processing time and the queue will hand the same job to a second worker — correctly, and by design. Reproduced locally, with the metrics that reveal it.

sqs

6 min

24 April 2026

Retry storms: how three polite clients take down one service

Every client retries three times. Nobody is misbehaving. The downstream sees a load multiplier it was never sized for. What jitter actually buys you, measured.

distributed systems

9 min

10 April 2026

DNS TTL is a promise about caching, not about truth

Lowering the TTL an hour before a cutover does not do what you think it does. Where the stale answer is really being held, and how to prove which resolver is lying.

dns

7 min

Why these are free

Don't paywall the proof. Paywall the structured journey.

A study shows you that something is true. A lab makes you prove it yourself, break it on purpose, repair it, and defend the repair against conditions you did not choose. Those are different products, and only one of them costs money.

Studies are the evidence

One question, one experiment, the actual output — including the runs that contradicted my prediction. If a claim here is wrong, it is wrong in public.

Labs are the practice

The same material as a system on your machine with a broken decoder, a failing test, and no answer key. Reading about a framing bug and repairing one are not the same skill.

Challenges are the exam

Symptoms only, misleading signals left in, nobody telling you which layer lied. This is the part that shows up in interviews and on call.

Reading is the cheap part.

Every study here has a lab behind it where you cause the failure yourself and prove the repair.

Get Foundations $99 once Read a full lab free