<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Distributed-Systems on Tushar Choudhary</title><link>https://tushar-c23.github.io/tags/distributed-systems/</link><description>Recent content in Distributed-Systems on Tushar Choudhary</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sun, 20 Sep 2026 19:30:00 +0530</lastBuildDate><atom:link href="https://tushar-c23.github.io/tags/distributed-systems/index.xml" rel="self" type="application/rss+xml"/><item><title>The batch is the unit: what Kafka actually writes to disk</title><link>https://tushar-c23.github.io/posts/kafka/batch-is-the-unit/</link><pubDate>Sun, 20 Sep 2026 19:30:00 +0530</pubDate><guid>https://tushar-c23.github.io/posts/kafka/batch-is-the-unit/</guid><description>&lt;p&gt;&lt;em&gt;First in the &lt;a href="https://tushar-c23.github.io/posts/kafka/"&gt;Kafka Internals&lt;/a&gt; series: reading the design doc against the actual source (Kafka trunk, 4.5.0-SNAPSHOT, KRaft-only). Every number below came off a single-node broker running on my laptop, and you can reproduce all of them.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Kafka&amp;rsquo;s API is message-shaped. You call &lt;code&gt;producer.send(record)&lt;/code&gt; with one record. You call &lt;code&gt;consumer.poll()&lt;/code&gt; and iterate &lt;code&gt;ConsumerRecord&lt;/code&gt;s one at a time. The docs say &amp;ldquo;message&amp;rdquo; everywhere. So the obvious mental model is: a message goes in, a message sits on disk, a message comes out.&lt;/p&gt;</description></item></channel></rss>