Batching

The "Postal Service" Problem (Overhead)#

Imagine you have 100 letters to send. - Individual Sending: Driving to the post office 100 times. You spend more time "driving" (network overhead) than "mailing" (data transfer). - Batch Sending: Waiting until you have 100 letters, putting them in one box, and making one trip.

At Google scale (100,000 clicks/sec), sending individual messages would bombard the Network Card with too many "packets per second," causing the system to choke.

How Batching Works in the Producer#

When your code calls producer.send(), the Kafka Library doesn't send it immediately. It places the message in a local memory buffer (RAM).

The Producer sends the batch when: 1. Batch Size is reached: (e.g., the "box" hits 16 KB). 2. Linger Time is reached: (e.g., the producer has waited 5ms, and the box isn't full yet, but we can't wait any longer).

graph LR
    App[Your App Code] -->|send| Buffer[Producer RAM Buffer]
    Buffer -->|Wait 5ms or 16KB| Batch[Full Batch]
    Batch -->|One Network Trip| Broker[Kafka Broker]

The Trade-off: Throughput vs. Durability#

Batching is a "speed vs. safety" decision.

1. High Throughput (Ad Clicks)#

• Settings: linger.ms = 5, batch.size = 16KB.
• The Risk: If the App Server crashes while 9 clicks are in RAM, they are lost forever.
• The Reward: We can handle 100,000 clicks/sec because the Network and CPU are relaxed.

2. High Durability (Bank Transfers)#

• Settings: linger.ms = 0.
• The Risk: System is slower and harder to scale.
• The Reward: Every single cent is sent to Kafka the millisecond it happens. No data is left sitting in volatile RAM.

Why this doesn't "Bombard" your RAM#

A common concern is that batching will eat up all your server's memory. However, the math shows otherwise:

• 100 clicks × 1 KB each = 100 KB.
• Even with 1,000 active batches, that's only 100 MB of RAM.
• Most modern servers have 16 GB to 64 GB of RAM. 100 MB is a tiny fraction.