Google Cloud Run cost me $4,676 in 6 weeks with zero traff

> I’d appreciate hearing what others would do differently.

the answer is staring you right in the face:

> I fed my setup, budget, and constraints as context into Gemini CLI

> The commit message claimed “60% cost savings.”

don't outsource your critical thinking to a chatbot.

or, if you feel you simply must have the chatbot do this work for you, supervise it more closely. instead you ignored it for 6 weeks:

> From Nov 2 to Dec 14, Cloud Run accrued ~$4,676.

Multiplying complexity by nondeterminism. AI is going to democratize the way people don't know what they're doing.

Some honest, maybe hard to hear feedback: as a solo developer with no customers and no traffic you probably shouldn’t be doing anything with cloud services. Cloud providers have many scalable services that mostly come down to them being paid by traffic costs or CPU/RAM costs and they don’t care what you’re doing, if you have 0 customers or 1000, if its “dev” or “production” - its all production to them.

I’ve seen runaway cloud costs at my employer a few times, with a few different services and it took a fair amount of time to figure out the monitoring/alerting. They may change their service agreements or pricing structure in hard to decipher ToS, etc. If the cloud provider won’t refund or credit a company that has a representative, they aren’t going to pay any attention to a solo dev or small team.

I would build locally on your laptop and start with a $5/month VM until you get a paying customer and know what size your system needs.

Exactly right! Start with a $5/month VPS, then as needed upgrade to $10, $30, $60 a month. Then switch to scalable cloud.

As a solo dev, I use $5 VPS on OVH or Hetzner for peace of mind. I remember trying out GCP for some simple html or django thing a decade ago and it cost me $30/mo or something and I said nah. I've also seen an ex-coworker rack up $50k with no recourse. I've already read enough horror stories like this.

Also don't outsource your thinking to LLMs, it's a useful tool, but once you do, it's brainrot for programmers.

Most hobby and small business sites can easily run on a $5-10/mo VPS. If you need a bit more, hetzner server auctions should suffice. It's always safer to use a fixed cost service, instead of the cloud hosts that don't let you set a hard quota on spend.

I am all in on cloud, AWS specifically. I have been doing AWS consulting for five years.

But if I were doing a side project or starting a business, I would personally use a simple VPS. In my case I would use AWS Lightsail - it’s a simple fixed priced service with no surprises.

I’m not saying you should use AWS. But you definitely shouldn’t be taking advice from a chatbot when you don’t know what you’re doing.

It's a bit like when I needed ChatGPT to help me revise an article yesterday, and it automatically toned down the harsh criticisms I had written about ChatGPT in the article. Asking Gemini how to save money on Google Cloud would likely yield similar results.

With all due respect, what?

First of all, thank you so much for obviously writing part of this via a Large Language Model. Second of all, what kind of argument is "The commit message claimed '60% cost savings'" - do you have any idea what you were actually doing? And lastly, addressing your question:

> Do you set hard budget caps and accept downtime?

If you have no clue what you're doing, yes! Especially for early prototyping, why not? IaaS offerings will also just create downtime for you as well if you need more resources than you've provisioned. It's normal. Either you set up a system where you can rely on dynamic scaling or you don't and set hard limits.

You asked your cloud provider to provision resources, and you were billed for them. If you can't handle working with a cloud provider, you might want to look into less scalable but in turn more cost stable infrastructure solutions.

Appreciate the directness, and fair point. That’s exactly what I find confusing about their setup and why I’m here trying to learn.

A little more context: I’ve been on GCP for 4 years, App Engine for the majority of it. Expensive but stable. I’ve used Gemini in the past to reduce costs successfully, so this wasn’t my first attempt at optimizing.

I take ownership of the outcome, but the config behavior still doesn’t match my mental model and Google support hasn’t been able to clarify how to properly scope this either, which is why I turned here.

> the config behavior still doesn’t match my mental model

Could you -learn- how to self-host a version of your app to expand your mental model in doing so? You outsourced the thinking part to an LLM - a bag of words - and are surprised the outcome didn't just work?

> Google support hasn’t been able to clarify how to properly scope this either

More outsourcing of thinking, no? Is Plan A really asking the vendor selling you compute how to use less compute and make them less money, instead of figuring out how to use just enough of it yourself?

If you're taking ownership, who could have effected the outcome the most here? Maybe the person who keeps outsourcing thinking to LLMs, support requests, and forums? I'd argue ownership would look more like figuring out how to handle Top Cost #1 yourself and reduce burn rate, starting by doing less outsourcing.

You just did what some chatbot told you to do? One provided by the people whose profits you were trying to minimize?

I wouldn’t have expected that to go well.

> I’d appreciate hearing what others would do differently.

Don't use AI to build something you are on the hook for financially. AI knows no consequence.

> I’d appreciate hearing what others would do differently

Very naive answer: before doing anything I would start by playing around in the GCP pricing calculator [1] to figure out how much it was going to cost during development and in production. Did you use the calculator tool? If so, were its estimates accurate?

[1] https://cloud.google.com/products/calculator

This is the modern-day equivalent of asking the genie from the magic lamp for something, and getting something else that meets the letter but not the spirit of the request.

The smart thing to do is realize you don't know what you're doing, and don't rely on the genie at all. Or hire someone who knows how to tame the genie. Or whatever; someone else put it better: you fucked around and you found out. lol

I'd say rent a VPS or Dedicated server, you can get some serious hardware on a dedi box for $100-200 a month. And there's no 'overages' just because something uses more CPU or RAM.

Cloud stuff is a complete waste of money unless you really specifically need the scaling it offers and have wads of cash.

Thanks to everyone with constructive feedback.

A few things that stood out to me:

One alert is never enough. Treating alerting as a system, not a single email.

GCP pricing calculator in the deployment checklist, not an afterthought. You almost have to be obsessive about checking.

Hard caps: GCP doesn't have them. Budgets only send alerts, they don't stop charges. You can wire automation to disable billing but that's a hack, not a feature.

Cloud providers benefit financially from complexity and opaque defaults. I own my config mistakes, but the suggestions around capacity planning and realistic cost modeling were genuinely helpful.

Sorry to hear this happened to you. Everyone's piling on, like they've never made mistakes. With no "hard" limits, it's almost like google wants this to happen. Have you tried contacting your credit card company? They may be able to chargeback google. Maybe you can make a case about idle capacity?

I have enabled some google services to play with (maps, calendars), that required me to put my credit card down. I'm thinking I should get a prepaid credit card so if things get out of hand, they'll shut me down rather than keep charging me until my max credit limit.

Wait you’re still paying $5 per day on an idle service?

That’s absurd.

I know, right? I solo a small business (6 figures revenue) off a static site and backend on $1/month VPS.

$0 revenue, 0 customers, and thinks $5/day is acceptable. LOL.

I came here to say this.

OP, what are you even doing that requires $150/month of spend, especially given things are mostly idle?

Photo culling app. Photographers upload 1000+ images, ML clusters similar shots so clients can pick favorites faster. Needs serious parallel compute (but infrequently). Cloud SQL is the baseline (~$40-50/month, doesn't scale to zero).

My fuckup was setting --cpu 4 --memory 16Gi during initial migration and only optimizing instance counts, not resource sizing.

Didn't help that GCP has no hard spending cap. Budgets only send alerts (in my case 1 easy to miss alert saying you spent more than $50 this month), they don't stop charges... Looking at my credit card statement there are multiple $1-2k charges.

The incentives aren't exactly aligned toward helping you spend less.

> Cloud SQL is the baseline

Does GCP not have something equivalent to Aurora Serverless? [0]

It's more expensive if you are constantly using the database over the course of a month, but much less expensive if you only have bursty traffic.

0: https://aws.amazon.com/rds/aurora/serverless/

your alerts seemed to have worked if you got one alert right? I used to work with situation alerts, and the "one alert" thing is a trap, you miss one and you're done. I coded alert recaps, alert reminders, custom scheduling etc etc, if you have opportunity for getting alerts, don't write them off because you got just a single alert, stop counting on perfectly response to a single alert and lean into multiple alerts wherever possible. take care of the human tasked with responding to alertable situations, we're imperfect error prone blobs of protoplasm lol

I understand the "everybody does it that way" aspect of the cloud, but the old way of subscribing to a dynamic DNS and hosting locally still works just fine and you have full control of things.

Later, once you have more traffic and/or paying customers, it would be worth looking into cloud hosting. And even then, you probably don't need as much horsepower as they're trying to sell.

I self host. I also have no paying customers and negligible compute needs. It's free minus the power cost which is, again, negligible. If it was on docker on my main computer instead of 18 year old junk, it would probably be even less.

I don't understand why any solo dev would use aws. Its awfully priced and 99% of people can just use a vps and that should be enough. By the time you blow up and need cloud scale you can afford it.

> Is App Engine still preferable purely for cost predictability?

How is Cloud Run cost not predictable?

It is fairly simple arithmetic in a spreadsheet to estimate the upper bound with # of max instance times the unit prices times the per-instance resources. Exactly the same as you do with App Engine standard & flexible environment.

> budget caps

Doesn’t exist in GCP and most cloud providers. You can fix the usage or hard cap the usage autoscaling, but not the spend incurred by the usage.

> What guardrails work that don’t depend on constant manual billing checks?

Start conservatively with max instances and instance resource, and iterate based on the actual performance and needs. Say, you know, put the number 1 in everything.

Do your capacity planning and cost estimation and understand them. “Solo dev” or not, you need these things to run the business. The root cause was that you outsourced your business and budgeting decision to LLM without verifying it and understanding it.

I am sorry for the OP but this is a perfect example of "You don't need cloud at your stage". Get a $10 VPS and call it a day. Unless you have 1000s of daily active users, a $10 VPS with a solid backup strategy will work. Heck, you can even do more than that with a decent VPS. You could setup a docker compose on 1 single decent server and call it a day when you have 0 users.

I like using VMs for their predictable pricing. And docker-compose is easy enough to iterate with.

Any favorite providers?

Hetzner

Stories like this scare me and why I stick to PAAS.

Isn't Google Cloud Run a PAAS?

Sorry, I meant things like Render and PythonAnywhere

> I fed my setup, budget, and constraints as context into Gemini

OP, you fucked around and found out.