Field note · November 20, 2026 · Maks Danylenko

We budgeted $10,000 for the first month of logging. GolfNext’s bill came to $200 a month.

Suppose a startup makes ten infrastructure decisions. Each time, the team picks a $1,000-a-month solution when $100 a month would cover the need. That adds up to $10,000 a month for infrastructure the startup could run for $1,000. Logging at GolfNext is a good example of how one of those decisions gets made.

GolfNext’s applications wrote a new log file each day on every machine. When a customer reported a problem, we had to search those files by hand, machine by machine. Support needed one place to search the logs, track recurring errors, and get an alert when the error rate spiked.

Why we considered ELK for GolfNext

Back in 2023, if you asked an architect or consultant how to centralize your logs, the usual answer was the ELK Stack. That was our plan for GolfNext, too. Before we proposed it to the client, we estimated what it would cost to set up and what we’d spend each month keeping it running.

Our November 2026 cost model allows 80 hours to set up ELK. At $100 an hour, that came to $8,000. For a workload of 1 TB of logs per month with 90 days of retention, we estimated another $1,250 a month for infrastructure. That put the first month at $9,250, so we rounded the budget to $10,000.

Had we run ELK ourselves, we would have budgeted 20 engineering hours a month to maintain the cluster. At the same rate, that would come to $2,000 for maintenance plus $1,250 for infrastructure, for an estimated total of $3,250 a month.

Would a hosted logging platform cost less?

We compared monthly costs for 1 TB of incoming logs. Our ELK estimate includes 90 days of searchable history. The ELK baseline is $3,250 per month, including 20 hours of maintenance at $100 an hour and the modeled infrastructure.

Estimated monthly cost for 1 TB of incoming logs (November 2026)
PlatformEstimated monthly costComparison with ELK
Self-hosted ELK$3,250Baseline
Logz.io$3,410–$4,092$160–$842 more
Splunk Cloud$3,300–$4,500 (third-party estimate)$50–$1,250 more
Datadog~$2,60030 days of retention. A 90-day comparison requires a quote.
Sumo Logic Continuous$3,750–$6,250 for ingestion$500–$3,000 more before storage.

Why we chose CloudWatch Logs

GolfNext already ran on AWS, so we checked what CloudWatch Logs would cost. For the same 1 TB a month and 90 days of retention, we estimated CloudWatch Standard at about $590 a month. We could start smaller and pay for actual usage. GolfNext’s logging bill came to about $200 a month.

One problem showed up in our Java logs. A single stack trace could arrive as dozens of separate events, which made it hard to read the exception from start to finish. Over a weekend, I wrote a Logback appender that sends the whole stack trace to CloudWatch as one event. If you’ve run into the same issue, the appender is open source and you can use it in your own application. Logback CloudWatch appender on GitHub.

I gave four AI agents the same logging problem

Would an AI agent recommend a separate logging stack? I gave four agents the same two-service project, asked each to plan centralized logging, and left the hosting details out of the brief.

All four recommended Grafana Loki. One suggested a managed service. If we hosted Loki ourselves, we’d face much the same question we’d faced with ELK: who would run it, and what would their time cost? If we budgeted 20 engineering hours a month at $100 an hour, that would come to $2,000 a month for maintenance alone, before hosting and storage.

Then I added one sentence to the README: “Services run on AWS ECS (Fargate).” I gave the agents the same task again. This time, all four recommended CloudWatch. The logging requirements hadn’t changed. The agents now knew what the client already had.

What happens when you make that choice ten times?

We’ve seen clients running a demo app on Kubernetes when a single instance on AWS Elastic Beanstalk would have done the job. The demo ran, but someone still had to look after the cluster. That’s what I mean by hype-driven development: choosing a tool on the strength of its reputation before you can name the client requirement that justifies it.

Logging is one decision. Where to run the app is another. A small team can easily make ten decisions like these while getting a product off the ground. Each one adds to the monthly bill and leaves the team with something else to maintain.

That’s why the quality of each decision matters.

Whether a recommendation comes from an architect I respect or an AI agent, I still have to explain it to the client. For GolfNext, I could explain why support needed searchable logs and alerts. I couldn’t name anything they needed that justified running a separate cluster.

Who responds when a release goes wrong?

If errors spike after a deployment, Mavka AI SRE investigates and alerts your team. You decide in advance when it can roll back the release or open a pull request with a fix. Our on-call engineers can step in when an incident needs a person.

Book an AI SRE intro call

Research notes

  • AWS CloudWatch pricing and log classes.
  • 90-day infrastructure estimate: three EC2 nodes, about 3 TB of gp3 disks, S3 snapshots and inter-AZ traffic. The self-hosted comparison adds 20 hours at $100 an hour.
  • Logback appender source at the reviewed revision.
  • AI experiment: four agent configurations each planned with and without the ECS Fargate detail. The 8 runs are a small field note, not a benchmark.

Related reading