Introduction

When using AWS, you likely choose the Tokyo region (ap-northeast-1) quite often. This choice is typically based on the following assumptions:

  • Most users are within Japan
  • Data needs to be stored domestically
  • Latency should be minimized

I myself was aware of these reasons, but I kept selecting the Tokyo region somewhat arbitrarily. This time, I want to focus on latency and verify whether Tokyo is truly the fastest option.

What We're Validating

I'll create ALBs in the following regions and measure latency by sending HTTP requests:

  • Tokyo
  • Osaka
  • Singapore
  • Northern Virginia
  • São Paulo

The measurement source will be from the Tokyo suburbs (my home).

Validation Results

Here are the results:

Region Latency
Tokyo 33.1292 ms
Osaka 139.517 ms
Singapore 225.019 ms
Northern Virginia 508.59 ms
São Paulo 619.534 ms

As expected, the Tokyo region had the smallest latency. The results also confirm that latency is proportional to physical distance.

What About CloudFront?

Next, let me verify the case with CloudFront. I specified an ALB as the origin under CloudFront. Here are the results...

Region Latency
Tokyo 31.8299 ms
Osaka 31.7758 ms
Singapore 33.6764 ms
Northern Virginia 31.4756 ms
São Paulo 33.1176 ms

The results showed minimal differences across all regions. When requests go through CloudFront, they are processed at edge locations, so regional differences in the origin are mostly absorbed.

This means that using any region produces similar results.

Warning: These measurements were taken with CloudFront caching enabled. When cache misses occurred, latency was proportional to physical distance, similar to direct ALB access.

Summary

From the validation results, we learned:

  • Tokyo region is the fastest for access from Japan
  • For direct ALB access, latency is proportional to physical distance
  • CloudFront can absorb regional differences

Therefore, whether the "Tokyo region" is truly optimal depends on your use case. Region selection should not be based on "intuition" but rather on latency and architectural characteristics.

The results this time turned out mostly as expected, but I want to keep questioning whether "commonly accepted truths" are really correct.

Info: The code used in this verification is published here.