[jart]

Use a token bucket on your web server to catch abusive IPs and then blackhole them using `iptables -t raw -I PREROUTING -s ip -j DROP`. I know. I run https://ipv4.games/ which invites hackers to unleash their botnets, and the service runs on a small VM with only a few cores. It's been attacked by botnets with 49,131,669 IP addresses. There's no Cloudflare frontend or anything like that, because back when I used Cloudflare, the people who attacked the service would actually bring down the Cloudflare nodes before they brought down my web server. I doubt I've ever paid more than $100/month to operate the service. Please note that your service provider needs to have free ingress in order for this strategy to be effective.

[wielebny]

This strategy may work for a (D)DoS that is targeted to an application layer, but won't work if the attack is designed to exhaust your bandwidth.

Once you're receiving more traffic than you network cards can handle, it does not matter if you'll drop the packets with iptables or not.

I was the target of attacks that caused Hetzner to terminate my contract. I was leasing physical servers there, so I assume the attacks were overwhelming their infrastructure.

[zettabomb]

These days it seems that DDoS attacks are often not targeted at bandwidth either, but rather packets per second. It is (apparently) much easier to exhaust routing capacity with an inordinate number of tiny packets than with a still large number of large packets. Cloudflare has some fun ways to deal with this [0].

[0] https://blog.cloudflare.com/mitigating-a-754-million-pps-ddo...

[jart]

What they did to me was flood the Linux Kernel with TCP connections. That's why it's so important to block IPs in the raw PREROUTING table. You need to nip it in the bud before Linux starts allocating any memory to the attacker.

[jart]

I rent a GCE VM and there's not many if any people out there who can exhaust Google's network infrastructure. The only thing I have to worry about is making sure my server doesn't respond to abusive traffic.

[bravetraveler]

Eventually you're probably going to want an ipset, at least. Otherwise processing your chain will continuously cost more, and more, and more.

[jart]

I just declare firewall jubilee every now and then, where I flush the iptables and let people try again. It's also because people usually only control the IPs they use temporarily, so I don't want someone innocent later on to be blocked from using the service because someone abusive used their IPs beforehand. But even if I didn't do this, it doesn't cost much for Linux to iterate over an array of blocked int32's. It's really only allocated TCP connection resources that are problematic.

[bravetraveler]

I'm glad you make a point to flush the chain/let things retry. I often hear about people adding drops.. to just then forget about them

I saw millions and started to feel my heart race a little