Service Limits and Quotas Explained
Quotas decide what your account can create before you ever hit a technical ceiling. This guide covers soft and hard limits, where to find yours, and how increase requests work.
You can have a perfectly designed architecture and still be unable to deploy it, because the account you are deploying into is not allowed to create that many of something. Quotas are the least interesting part of cloud infrastructure right up until they are the only thing standing between you and a launch.
The mechanics are not complicated, but the vocabulary is inconsistent across providers and the numbers live in places people do not think to look. This guide sorts out the terms, explains which limits tend to bite first, and sets out what actually happens when you ask for more.
Three words that are not synonyms
Limit, quota and capacity get used interchangeably and they are not the same thing. A quota is an administrative allowance: your account may create up to this many of this resource. A limit is often the same idea, though some providers use it for technical ceilings, such as the maximum size of a single object or the maximum number of rules in a firewall. Capacity is physical: whether the provider has hardware free in that location right now.
The distinction matters because the remedies differ. A quota problem is solved by a request to the provider. A technical limit is solved by changing your design. A capacity problem is solved by waiting, changing instance type, or changing location. When a deployment fails, the first question is which of the three you have hit, and the error message usually says, if you read past the first line.
Where quotas come from
Providers apply quotas for three reasons, and understanding them tells you how a request will be received. The first is fraud control: a brand new account that can immediately spin up a large fleet is an attractive target for abuse, so new accounts start low. The second is capacity planning, since the provider needs some predictability about demand in each location. The third is protecting you from your own automation, which is the reason behind many of the smaller limits.
Because of the first reason, quotas tend to rise over time for accounts with payment history. That is the legitimate kernel of truth behind the emphasis on account age in listings, discussed in how to read a cloud account listing. It is also why quotas are set per account: they are a statement about the account holder’s relationship with the provider, not about the technology.
Soft limits and hard limits
A soft limit is a default that the provider will adjust on request. Most compute, storage and networking counts are soft. A hard limit is a ceiling built into the service that no support ticket will move, usually because it reflects an architectural constraint in the platform itself.
The practical approach is to find out which category each constraint in your design falls into before you build. If a soft limit stands between you and your target scale, that is a scheduling problem: submit the request early and carry on. If a hard limit does, that is a design problem and no amount of planning ahead will help. Discovering the difference during a launch window is avoidable and painful.
Providers rarely label the two categories explicitly in the console, so the documentation is where the answer lives. As a rough guide, anything expressed as a count of resources you can create is usually adjustable, and anything expressed as a maximum size, a maximum depth, or a maximum number of entries inside a single object is usually fixed. That heuristic is not perfect, but it points you at the right question quickly.
The quotas that bite first
In practice a short list accounts for most of the blocked deployments people encounter.
- Total vCPU count per region, often broken down by instance family, which catches anyone scaling horizontally or using larger machine sizes.
- Accelerator availability, which is both a quota and a capacity question and is usually the tightest constraint on the platform.
- Public IP addresses, which run out quietly when every instance gets one by default.
- Virtual network count per region, which constrains how many isolated environments you can build inside one account.
- Managed database instances and storage volume totals, which are easy to forget because they grow slowly.
- API request rates against the control plane, which infrastructure as code tooling can exhaust during a large apply.
Note how many of these are per region. A fleet that deploys fine in your usual location can fail entirely in a second one, which is a frequent surprise during expansion. If you are planning a multi region footprint, how regions and availability zones work explains why the boundaries fall where they do.
Finding your current numbers
Every major provider has a page that lists quotas with current usage against each one. It is usually called service quotas, limits, or something similar, and it sits in the account or management section rather than inside each individual service. Learn where that page is for your provider before you need it.
Read it with the region selector in mind, since the page almost always shows one region at a time. Also read the usage column, not just the limit column. A quota at eighty percent consumption is a future outage with a date on it, and catching that during a routine review is far cheaper than catching it during an incident. Where the provider exposes quota metrics to its monitoring service, alert on the handful that matter, in the same spirit as the alerting described in cloud monitoring and alerting.
Requesting an increase
An increase request is a short piece of persuasion, and the providers that handle them well are responding to a human reading your justification. Say which quota, in which region, what the new value should be, and what you are building. A specific number with a reason is approved faster than a vague request for more.
A few things help. Ask for a value that covers the next six months rather than exactly what you need today, because repeated small requests take longer in total. Submit before you are blocked, since lead times vary from minutes for routine increases to days for large ones or scarce resource types. Keep a payment history on the account, because an account that has never paid an invoice is a weaker case than one that has.
Expect some requests to be declined or partially granted, particularly for accelerators, and have a plan that does not depend on the answer. Providers also differ in how generous their defaults are in the first place, which is one of the less visible distinctions between platforms in a comparison like AWS against Google Cloud.
Capacity is a separate conversation
Having quota does not guarantee you can launch. If a provider has no free hardware of the type you asked for in the zone you asked for, the request fails regardless of your allowance. The error wording differs from a quota error, and learning to tell them apart saves you from filing a pointless support ticket.
Mitigations are design choices rather than paperwork: accept more than one instance type, spread across zones, use capacity reservations where the provider offers them, and build retry with backoff into whatever provisions your infrastructure. Scarcity is most acute for accelerated compute, which is why anyone planning that work should read GPU instances for training and inference before committing to a schedule.
Designing around a limit you cannot move
When a hard limit blocks you, the options are structural. Shard the workload across more than one region, which spreads the demand across separate regional allowances. Shard across more than one account, which is one of the arguments in the case for account isolation. Or change the resource shape: fewer larger machines instead of many small ones, or a managed service whose limits are expressed differently.
Whichever route you take, record the limit and the workaround somewhere your future self will find it. Quota constraints are among the most commonly rediscovered facts in infrastructure work, and the second discovery is just as expensive as the first. A single file in your repository listing every quota you have hit, the region it applied to, and what you did about it will save someone a working day within the year. It is also the first thing a new engineer should read before planning anything that scales. If you are choosing a platform partly on how much room its defaults give you, product pages such as Buy AWS Account and Buy Oracle Cloud Account are a reasonable place to start comparing.