There are two distinct things people mean by a denial of service attack, and buying protection for one when you have the other is the most common way this purchase goes wrong.
The first is a flood. Packets arrive faster than your pipe can carry them or faster than your network stack can process them. Your application is irrelevant to this attack; it never sees a request. The traffic is often reflected and amplified off third-party servers, it comes from spoofed sources, and the only place it can be stopped is upstream of your bandwidth, in a network large enough to absorb it. You cannot solve this yourself. Nothing you configure on your origin matters.
The second is a request flood. Well-formed HTTPS requests, in modest volume, aimed at the expensive parts of your application: a search endpoint that scans, a report that aggregates, a login that hashes a password, a page that renders after six database queries. There is no bandwidth problem at all. Your origin falls over because it is doing real work that it was asked to do, and every request individually looks like something a customer might send.
The defences are unrelated. The first needs network capacity and packet-level filtering. The second needs request inspection, rate limiting, caching, bot classification and application-level judgement. A provider can be excellent at one and unremarkable at the other, and the marketing for both uses the same three letters.
Then there is the third thing nobody discusses until it happens: what the attack costs you in billing. Several protection models meter on traffic delivered or requests processed, and an attack is by definition a large amount of both. The clause that determines whether an attack is an incident or an incident plus an unbudgeted invoice is in the contract, and buyers typically read it for the first time during the attack.
Key takeaways
- Layer 3 and 4 volumetric floods and layer 7 request floods are different attacks needing different defences. Network capacity does not protect an expensive endpoint, and request inspection does not stop a link being saturated.
- Always-on protection removes the detection and diversion delay that on-demand mitigation always has. On-demand costs less and means the first minutes of every attack land on you.
- Surge billing during an attack is the clause to negotiate before signing. Ask specifically whether attack traffic is billed, whether there is cost protection, and what conditions attach to it.
- Your origin must be unreachable except through the protection, or all of it is decorative. Origin address exposure through DNS history, certificate transparency logs, mail headers and old subdomains is the standard bypass.
Layer 3 and 4: absorbing a flood
Volumetric attacks work by arithmetic. If the attacker can direct more traffic at your address than your connectivity can carry, the link saturates and legitimate packets are dropped at the upstream router, regardless of anything you do downstream.
The mechanics worth understanding:
Amplification and reflection. The attacker sends small requests with your address as the forged source to third-party servers running protocols that answer with much larger responses. The victim receives the amplified responses from thousands of innocent servers. This is why attack traffic often arrives from legitimate infrastructure and why blocking source addresses is futile.
State exhaustion. Rather than saturating bandwidth, the attacker exhausts a finite table: connection state on a firewall or load balancer, half-open connections, or NAT entries. Volume can be modest while the resource fills. A device sized by throughput can fail here while the link is far from full.
Protocol abuse. Malformed or unusual packets that cost far more to process than to send, targeting the asymmetry between generating traffic and parsing it.
The only durable defence is to have the traffic arrive somewhere with more capacity than the attacker can generate, and to filter it there. That means one of three architectures.
Always-on proxy. Your DNS points at the provider, traffic is inspected and forwarded to your origin. Filtering is immediate because the traffic was never going anywhere else. This covers web traffic naturally and requires the provider to terminate TLS.
BGP-based diversion for whole networks. You announce your address space through the provider, permanently or on demand, and traffic is routed through scrubbing centres before reaching you. This protects everything you run, not just HTTP, which is the only option for non-web protocols, game servers and voice. It requires you to own address space and speak BGP, which excludes most SaaS companies.
Cloud provider network protection. If you run in a major cloud, the provider already absorbs common volumetric attacks at the network edge as a baseline for all customers, with an enhanced paid tier adding detection tuning, response support and billing protection.
The detection and diversion delay is the thing to ask about for anything not always-on. On-demand mitigation means traffic flows normally until detection triggers, then routing changes, then scrubbing begins, and route propagation is not instant. During that window your origin absorbs the attack. Vendors publish target times for this; treat them as targets rather than guarantees, and ask what triggers detection, whether it is automatic, and what happens outside business hours.
Layer 7: when the attack is well-formed requests
This is the attack most SaaS companies actually experience, and it is the one network capacity does nothing about.
The attacker does not need volume. They need to find the requests that cost you the most and send enough of them. Every application has these endpoints:
Search and filter endpoints that scan rather than use an index, especially with unusual parameter combinations that defeat your cache.
Report generation and export that aggregates over a large range, holds a connection, and consumes memory.
Authentication endpoints that deliberately perform expensive password hashing, which is correct security design and an attacker’s gift.
Anything with cache-busting parameters. A query string the attacker varies per request bypasses your CDN entirely and lands every request on the origin.
Endpoints that call third parties. Your rate limit with the downstream provider becomes the resource being exhausted, and the failure appears in a system you do not control.
The defences are all application-aware. Rate limiting per address, per account, per token and per endpoint, with the awareness that a distributed attacker defeats per-address limits. Caching aggressively enough that most requests never reach the origin, which means cache keys that ignore parameters that should not vary the response. Bot classification, since much of this traffic is automated and identifiably so, covered in bot management tools. Request inspection through a web application firewall for the patterns that are identifiably abusive. And on your side, timeouts, concurrency limits and queue bounds so an overloaded endpoint degrades rather than taking the process down with it.
The uncomfortable truth about layer 7 defence: at some sensitivity you will block real customers, and the population you block is not random. The same tradeoff described in bot management applies here, with the added pressure that you are usually tuning it during an incident.
A word on the slow attacks, which are the opposite of a flood. Opening many connections and sending data as slowly as the server will tolerate exhausts connection slots with almost no bandwidth. Modern reverse proxies handle this with timeouts and connection limits, and any proxy-based protection absorbs it structurally, but an origin exposed directly is still vulnerable to something a single laptop can run.
Needs first-hand data: Load test your own application to find its actual breaking point per endpoint: requests per second at which latency degrades, which endpoint fails first, and what the failure looks like. Do it against a production-shaped environment with a production-shaped dataset. Without this you cannot set a meaningful rate limit, you cannot size anything, and during an attack you will not know whether the traffic you are seeing is even large enough to be the cause.
The surge billing clause
This is the section buyers wish they had read before signing.
An attack generates traffic, and several parts of your stack bill for traffic. If the protection provider bills on delivered bandwidth or requests processed, and your origin bills on egress and compute, and your autoscaler responds to load by adding instances, then the attack produces an invoice as well as an incident. The invoice arrives weeks later, after the technical incident is closed and the team has moved on.
Ask these questions before signing, and get the answers in the contract rather than in an email from a sales engineer.
Is attack traffic billable? Some providers explicitly exclude mitigated attack traffic from metered charges. Others bill everything that traverses the network and offer a credit process afterwards. A credit process is not the same as an exclusion, because it is discretionary and it happens after you have received the bill.
Is there cost protection, and what triggers it? Several providers offer a form of billing protection for charges caused by a documented attack. Read the conditions: it commonly applies only to specific services, only on specific plan tiers, only when the attack is confirmed by the provider, and only when you followed the recommended architecture. A protection that requires you to have already deployed things you have not deployed is not protection.
What is unmetered, and is it in the contract? Unmetered mitigation is a common and valuable commitment, and it means mitigation capacity rather than everything you use. Confirm exactly which meter it covers.
What happens to your origin bill? The protection provider’s billing is one meter. Your cloud egress, your compute autoscaling and your per-request serverless charges are others, and they are not covered by the protection vendor’s cost protection regardless of what it says. The mitigation here is architectural: make sure the origin only accepts traffic from the protection network, so attack traffic cannot reach a metered origin at all.
Does the contract distinguish attack traffic from legitimate growth? A successful product launch and an attack both look like a traffic spike. Know who decides which one it was, and on what evidence.
Is there a commitment tier that changes during an attack? Some agreements price on a committed volume with overage rates. An attack blows through the commitment, and the overage rate is the number that matters, not the headline rate.
The architectural control that makes most of this moot: lock the origin. Accept traffic only from the protection provider’s published address ranges, or better, use an authenticated tunnel or private connectivity so there is no public origin address at all. If attack traffic cannot reach your infrastructure, it cannot generate your bill.
Which leads to the other thing nobody checks.
Your origin is probably still reachable
Every proxy-based protection is bypassed completely if the attacker can address your origin directly, and origin addresses leak in predictable ways.
DNS history. Passive DNS services retain the address your domain pointed at before you moved behind a proxy. This is the first place anyone looks and it requires no skill.
Certificate transparency logs. Every publicly trusted certificate is logged, so every subdomain you have ever requested a certificate for is enumerable. The one you forgot to proxy is in there.
Subdomains that were never proxied. Mail servers, staging environments, monitoring endpoints, VPN gateways, legacy applications. Any of them resolving to the same infrastructure exposes it.
Outbound connections that reveal the source. Mail sent directly from the application, webhooks delivered from the origin, and error reports containing the origin address all disclose it to a recipient the attacker can arrange to be.
Server responses that leak. Redirects to an internal hostname, error pages containing an address, or an application that renders its own hostname in a header.
The fix is a firewall rule at the network boundary that permits only the protection provider’s address ranges, plus removing every direct public entry point to the same infrastructure. Then verify it from outside: try to reach your application by address rather than by name, and confirm it fails. Almost nobody runs this check, and it is the difference between protection that works and protection that an attacker steps around in ten minutes.
Needs first-hand data: Enumerate your own attack surface the way an attacker would. Pull certificate transparency records for your domains, check passive DNS history for your apex and main subdomains, and test whether your origin responds to a direct connection from outside the protection provider’s ranges. Record every exposed path you find and close it. Re-run this quarterly, because every new subdomain is a new chance to leak the origin.
Cloudflare

Cloudflare’s model is always-on proxy protection across a very large global anycast network, with volumetric mitigation included at every plan level rather than sold as an upgrade. Because traffic already flows through the network for DNS, CDN and WAF purposes, there is no detection or diversion delay: attack traffic arrives at an edge location near its source and is dropped there. For anything HTTP-shaped, this is the lowest-effort protection available.
Pros
- Always-on by architecture with no diversion step, so mitigation begins at the first packet rather than after detection
- Unmetered volumetric mitigation across plans, which removes the surge billing question for the mitigation itself
- Layer 3 and 4 and layer 7 defence in one policy surface alongside WAF, rate limiting and bot scoring, so response during an incident happens in one place
- Anycast means attack traffic is distributed across many locations rather than concentrating on one scrubbing centre
Cons
- Requires proxying and TLS termination by the provider, which some regulated organisations cannot accept
- Non-HTTP protocols need the network-level or Spectrum-style products, which are separate and commercially different
- Layer 7 defence still requires you to tune rate limits and bot policy, and an untuned deployment stops volumetric attacks and not much else
- Concentration risk, since DNS, CDN, WAF, bot and DDoS defence all depend on one provider’s control plane
Best for: Web-facing applications that want always-on protection and layered application controls without a network engineering project.
Pricing: Volumetric mitigation included across plans with no metering on attack traffic, and advanced application-layer controls sold with higher tiers.
AWS Shield
AWS Shield comes in two forms. The standard tier is automatic, free and applied to all AWS customers, absorbing common network and transport layer attacks at the edge. The advanced tier adds detection tuning, access to the response team, richer visibility, and the billing protection that is the main reason to buy it. It protects AWS entry points specifically, which is both the strength and the boundary.
Pros
- Baseline network layer protection is automatic and free for every AWS customer, so the floor is already higher than most people assume
- The advanced tier includes cost protection against scaling charges caused by a documented attack, which addresses the surge billing problem directly
- Access to a response team during an attack, which is worth a great deal at 3am when nobody on your side has done this before
- Integrates natively with CloudFront, Route 53, Global Accelerator and load balancers, with WAF policy in the same account
Cons
- The advanced tier carries a substantial monthly commitment with a minimum term, which is hard to justify until the first attack
- Protection applies to AWS resources, so anything outside AWS needs a separate answer and a second contract
- Cost protection has conditions attached, and the conditions assume an architecture you may not have deployed
- Layer 7 protection still comes from AWS WAF, which you configure and tune yourself, so Shield alone does not stop a request flood
Best for: AWS-centric organisations who want native protection plus billing protection and an escalation path during an incident.
Pricing: Standard tier included at no charge, advanced tier as a fixed monthly subscription with a minimum commitment period plus data transfer charges.
Akamai Prolexic
Prolexic is Akamai’s network-layer DDoS service, built around scrubbing centres that clean traffic for entire networks rather than for individual web properties. It handles any IP protocol, supports always-on and on-demand routing, and includes a security operations centre that will take action on your behalf under a pre-agreed runbook. This is the enterprise answer for organisations protecting more than a website.
Pros
- Protects whole networks and any protocol, not only HTTP, which is the only workable option for non-web services
- Always-on or on-demand BGP diversion, so you can choose the latency and cost tradeoff per network segment
- A staffed operations centre with pre-agreed playbooks, so mitigation decisions during an attack do not wait for your engineers to wake up
- Very large scrubbing capacity and long operational history against large attacks
Cons
- Requires owning address space and running BGP, which rules it out for most SaaS companies entirely
- Enterprise procurement and onboarding measured in weeks or months, so it is not a control you adopt during an incident
- On-demand mode carries a detection and diversion window during which traffic still reaches you
- Layer 7 application defence is a different Akamai product, so full coverage means buying more than Prolexic
Best for: Enterprises with their own address space protecting multiple protocols and services, not just a web application.
Pricing: Enterprise contracts based on protected address space, clean bandwidth commitments and always-on versus on-demand routing.
Fastly

Fastly provides DDoS protection as a property of its edge platform: traffic is absorbed and filtered across its points of presence, with its programmable edge allowing custom logic to shed or shape load during an attack. The combination of an edge network and the ability to write policy that runs there is the distinctive part, because it means your mitigation can be specific to your application rather than generic.
Pros
- Programmable edge means you can write mitigation logic specific to your application, such as serving a cached response instead of rejecting outright
- Always-on absorption at the edge with no diversion step for traffic already flowing through the platform
- Combines with its next-generation WAF for layer 7 defence under one vendor and one view of the request
- Real-time visibility into edge traffic, which shortens the diagnosis step during an attack considerably
Cons
- Focused on HTTP and similar traffic, so network-wide protection for arbitrary protocols is not the model
- Writing effective edge logic requires engineering effort and someone comfortable with the platform
- Commercial model is oriented toward larger traffic volumes, so small sites find better value elsewhere
- As with any proxy, the origin lock is your responsibility and the protection is worthless without it
Best for: Engineering-led teams already using Fastly who want mitigation they can program rather than configure.
Pricing: Platform subscription with traffic-based components, with DDoS mitigation included as a property of the service rather than metered separately.
Azure DDoS Protection
Azure provides basic network-layer protection to all resources by default and a paid tier adding tuned detection based on your traffic patterns, attack analytics, mitigation reports and cost protection. It applies to Azure virtual networks and public addresses, and it is the native choice for an Azure-centric estate, with the same argument as every native service: nothing new to procure, deploy or integrate.
Pros
- Baseline platform protection applies to all Azure public addresses with no configuration or charge
- The paid tier tunes thresholds to your observed traffic rather than applying generic ones, which reduces both false triggers and detection delay
- Attack analytics and post-incident mitigation reports are useful evidence for stakeholders and auditors alike
- Cost protection for scaling charges incurred during a documented attack addresses the surge billing exposure
Cons
- Protects Azure resources, so a hybrid estate needs a second product and a second operational model
- The paid tier is priced per protection plan with resource-count components, which is expensive for small deployments
- Network-layer focus means application-layer floods need Azure WAF configured and tuned separately
- Protection is per virtual network or plan scope, so coverage gaps appear if new resources land outside the protected scope
Best for: Azure-centric organisations wanting tuned network-layer protection and cost protection without a third-party contract.
Pricing: Basic protection included, with a paid tier sold as a monthly protection plan covering a number of public addresses plus overage per additional address.
Imperva
Imperva offers DDoS protection for websites, for individual addresses and for whole networks via BGP, alongside its WAF and application security portfolio. Its positioning centres on rapid mitigation commitments backed by service level agreements, and on the ability to protect both web properties and network infrastructure from one vendor, which suits organisations that have both and do not want two contracts.
Pros
- Covers website, individual address and full network protection modes, so one vendor spans web and infrastructure
- Contractual mitigation time commitments, which gives procurement something concrete rather than a best-effort statement
- Shares a platform with its WAF, so layer 3 and 4 and layer 7 defence are configured and observed together
- Long operational history and mature reporting, which matters for regulated buyers needing evidence after an incident
Cons
- Enterprise sales and pricing model, with the evaluation and onboarding timeline that implies
- Management interface carries a lot of accumulated complexity, which raises the cost of acting quickly during an incident
- Network protection modes require BGP and address space, with the same constraints as any network-level service
- Breadth across a large portfolio means depth varies, and some adjacent modules are weaker than the core
Best for: Enterprises needing both application and network protection with contractual mitigation commitments from one vendor.
Pricing: Enterprise subscription by protected properties, clean bandwidth and protection mode, with mitigation commitments in the agreement.
OVH

OVH includes DDoS mitigation with its hosting, dedicated servers and cloud services at no additional charge, filtering through its own infrastructure as a standard property of the service rather than a product you buy. For teams already hosting there, this is the most cost-effective protection in this list by a wide margin, and the tradeoffs are about depth and control rather than capacity.
Pros
- Included at no extra charge with hosting services, which removes the entire commercial decision for existing customers
- Substantial network capacity and long experience mitigating large volumetric attacks
- Protects any protocol on the hosted infrastructure, not only web traffic, which suits game servers and non-HTTP services
- No surge billing exposure for the mitigation itself, since it is not separately metered
Cons
- Applies to traffic destined for their infrastructure, so it is not a protection product for services hosted elsewhere
- Layer 7 application defence is thin compared with the dedicated vendors, so request floods need your own answer
- Less granular control and tuning than a dedicated provider, and less visibility into what was mitigated and why
- Support and escalation during an incident are not equivalent to a dedicated security operations centre
Best for: Teams already hosting on OVH who need solid volumetric protection without a separate vendor or a separate bill.
Pricing: Included with hosting and infrastructure services at no separate charge, with no metering on mitigated attack traffic.
How to choose
These questions in order will produce the answer.
Is the thing you are protecting HTTP? If yes, a proxy-based provider is simplest and fastest and you can be protected this week. If you are protecting game servers, voice, VPN endpoints or arbitrary protocols, you need network-level protection and BGP, which is a different product and a different timeline.
Do you own address space? No is the common answer for SaaS companies, and it eliminates the BGP-based services entirely. Say so early and stop taking meetings about them.
Always-on or on-demand? Always-on if you can, because the diversion window in on-demand mitigation is real and the attack lands on you during it. On-demand exists because always-on costs more; if you choose it, know the detection trigger, the diversion time, and who can invoke it at 3am.
Where is your bill exposed? Map every meter an attack would move: protection provider, cloud egress, autoscaled compute, per-request serverless, third-party API calls made per request. Then lock the origin so attack traffic cannot reach the metered parts, and get the protection provider’s billing terms in writing.
Have you verified the origin is unreachable? Do this before anything else, because it is free, it takes an afternoon, and without it every other decision on this list is theoretical.
What is your layer 7 answer? Volumetric protection does nothing about a request flood against an expensive endpoint. Rate limiting, caching, bot classification and application-level timeouts are the defence, and they need to exist before the attack.
| Provider | Layer 3 and 4 | Layer 7 | Model | Picks itself when |
|---|---|---|---|---|
| Cloudflare | Always-on, unmetered | Strong, same policy surface | Proxy | Web application, want protection this week |
| AWS Shield | Baseline free, advanced tier | Via AWS WAF | Cloud native | AWS estate, cost protection matters |
| Akamai Prolexic | Scrubbing centres, any protocol | Separate product | BGP diversion | You own address space and protect a network |
| Fastly | Always-on at edge | Via its WAF plus edge logic | Proxy | You want programmable mitigation |
| Azure DDoS Protection | Baseline free, tuned paid tier | Via Azure WAF | Cloud native | Azure estate, tuned thresholds wanted |
| Imperva | Website, address or network modes | Strong, same platform | Proxy or BGP | Enterprise needing both with SLA commitments |
| OVH | Included with hosting | Thin | Hosting property | Already hosting there, volumetric is the concern |
The two things that matter most are not on that table. Lock your origin so the protection cannot be stepped around, and know your own application’s breaking point per endpoint so you can tell an attack from a launch. Providers sell capacity; those two are the parts you own.
Frequently asked questions
Do I need dedicated DDoS protection if I am already behind a CDN?
For volumetric attacks, a CDN with global anycast capacity gives you most of what a dedicated service would, provided your origin only accepts traffic from the CDN. The gaps are non-HTTP services, which the CDN does not carry at all, and layer 7 floods against uncacheable endpoints, which pass straight through unless you have configured rate limiting and bot policy. So the honest answer is that the CDN covers the flood and you still owe an answer for the request flood, and that answer is configuration rather than another vendor.
What is the difference between always-on and on-demand mitigation?
Always-on means traffic is already flowing through the protection, so filtering starts at the first malicious packet. On-demand means traffic reaches you normally until detection triggers, routing is changed to divert it through scrubbing, and routes propagate. That sequence takes time, and during it the attack is hitting your infrastructure. On-demand costs less and is a reasonable choice for services that can absorb a few minutes of degradation. For anything customer-facing and revenue-carrying, the diversion window is usually the whole outage.
Will an attack increase my cloud bill?
It can, in several places at once, and this is the exposure most teams have not modelled. Metered egress, autoscaled compute responding to load, per-request serverless invocations, and per-request charges from third-party APIs your application calls all move during an attack. Some providers offer cost protection for charges attributable to a documented attack, with conditions. The reliable control is architectural: restrict your origin to the protection provider’s address ranges so attack traffic never reaches a metered resource, and confirm the billing terms in the contract rather than in a conversation.
How do I know whether this is an attack or a traffic spike?
Look at shape rather than volume. Legitimate spikes have a referrer or a campaign behind them, hit the pages a person would hit, carry a plausible mix of clients and geographies, and convert at some rate. Attacks concentrate on few endpoints, often the expensive ones, show unusual client fingerprints, arrive from address ranges with no history with you, and convert at zero. Having your own baseline recorded in advance is what makes this a five-minute question instead of an hour of argument, which is why the load test and traffic baseline work belongs in the calendar rather than in the incident.
Does DDoS protection help with scrapers and credential stuffing?
Only incidentally. Those are low-volume, well-formed, application-level problems and a volumetric mitigation layer will not notice them. They need bot classification, per-account rate limiting and authentication-layer defences, which are covered in bot management tools. The overlap is that the same vendors often sell both, which is a convenience in the control plane and not evidence that one product solves the other problem.
Related reading
- Best application security tools — where availability defence sits relative to the rest of the programme.
- Best web application firewalls — the request inspection layer, and why blocking mode on day one is a mistake.
- Best bot management tools — low-volume automated abuse that volumetric mitigation never sees.
- Best rate limiting solutions — the layer 7 control that actually stops a request flood.
- Best CDN providers — the delivery network that absorbs most volumetric traffic as a side effect.
- Best DNS providers — the other thing an attacker targets, and the history that leaks your origin.