We Picked the Region Because It Was the Default in the Dropdown

The proof of concept took just an afternoon to create. An account was made using whatever region was selected by default, a few services were deployed, and the setup was presented to a stakeholder who approved of it. A year and a half later, that same region hosted the production database, object storage, queues, and four years' worth of regulatory data, with no one having made a decision about it at any point.

The repercussions unfolded gradually, each one minor enough to ignore. Most users were located on the other side of the world, adding roughly a hundred and forty milliseconds to every round trip. Since the application communicated frequently rather than being slow, this resulted in web pages that users described as sluggish, but no one filed a ticket about it.

Later, legal inquired during an unrelated review about where customer records were physically stored, and the truthful answer required explaining that no one had made a choice. A managed service that was desired was not available in that region, so a self-managed older version was used instead, leading to the operational work that the platform was intended to prevent.

The analytics team subsequently set up their warehouse closer to headquarters, resulting in nightly data transfers across regions that continued indefinitely. What makes region selection more complex than it initially appears is that it is one of the few choices that truly impacts a system in an environment that markets itself as elastic.

Compute resources can be easily shifted. Stateful services require an outage, a migration project, and a real risk of data loss before they can move. By the time someone is a senior enough to worry about the answer, changing it would cost more than simply living with it, so that's what they do, and they refer to it as their architecture.

In the end, I now view a few choices as one-way streets and insist they receive five minutes of deliberate thought, even during a prototype, especially concerning region and data residency, account and subscription boundaries, the identity provider, and the primary datastore. These are not decisions with wrong answers; they are the ones that start out being fine, become permanent without anyone realizing, and later need to be defended by people who were not present during the original decision-making process. This is the work of Serguey Shinder.