Satire

We Scaled to Forty Instances. The Report Still Takes Fourteen Minutes.

The infrastructure spend doubled. The bottleneck is in a stored procedure. The two facts are unrelated.

2026 Apr 222 min readamdahls-lawperformancescalingbottleneck

The reporting service is slow. P95 response time is unacceptable. The ops team is given budget to scale horizontally. Amdahl's Law is not in the capacity planning document.


What Actually Happens

The service scales from four instances to forty. P95 latency improves by 9%. The business report that takes fourteen minutes now takes twelve minutes and forty seconds. An infrastructure engineer declares a win. The product team does not.

Three weeks later, a database engineer is handed the problem. She looks at the stored procedure generating the report. It loops through fifty-three thousand records sequentially, applies a transformation to each one, writes intermediate results to a temp table, then aggregates. It runs on a single thread. It has run on a single thread since 2018. No amount of additional application instances touches it — the procedure runs on the database server, not the application servers.

Rewriting the procedure takes four days. Report generation drops to forty seconds.


Blocky· Senior Pragmatist

You scaled everything except the constraint. The constraint did not notice.

DivineForge Advisory

Recognise this pattern in your organisation? I help teams cut through the governance, make the right technical calls, and actually ship.

Let's talk architecture →

How a Builder Should Respond

Profile before you scale. Identify which portion of the work is inherently sequential — that portion defines your ceiling regardless of how many instances you add. A small improvement from horizontal scaling tells you the sequential fraction is large. Find it, fix it, then scale. Scaling parallel capacity around a serial bottleneck is expensive infrastructure theatre.


This scenario illustrates Amdahl's Law. See all 13 laws →

ShareXLinkedIn

Stay in the loop

New satire drops when the enterprise does something worth documenting. No spam — just the next article.

20 articles about enterprise dysfunction. None of it billable. Buy me a coffee.

Buy me a coffee