upstream hash consistent seems to depend on order of DNS entries (Maxim Dounin)

[Ian Morris Nieves]

Hi Maxim,

Thank you for your response.
OK, I understand what you have said.

This seems slightly strange to me.  We both understand the purpose/point of hashing (consistent or not consistent).  Why does it then make sense to fall-back to round robin?

In my case, the number of hosts will change over time, and I can’t update the nginx config.  So I thought it would make sense to use a hostname that resolves to many IPs.  This would be a scalable solutioin.

I would think that the more correct solution would be to hash the IP addresses that all the names resolve to…. not the names themselves.

I must be missing/mis-understanding something.  Does the current implementation/solution make sense to you?

Best,
Ian

> 
> Message: 2
> Date: Mon, 3 Feb 2020 16:05:38 +0300
> From: Maxim Dounin <mdounin at mdounin.ru>
> To: nginx at nginx.org
> Subject: Re: upstream hash consistent seems to depend on order of DNS
> 	entries
> Message-ID: <20200203130538.GR12894 at mdounin.ru>
> Content-Type: text/plain; charset=utf-8
> 
> Hello!
> 
> On Sat, Feb 01, 2020 at 11:48:00AM -0800, Ian Morris Nieves wrote:
> 
>> Here is the setup:
>> - I am running nginx in a single docker container and it has an 
>> upstream to a docker service which is composed of 3 docker 
>> containers (which happens to be php-fpm)
>> - the service is configured to _not_ expose a single virtual ip 
>> address (vip), instead the service exposes the ip addresses of 
>> all 3 containers through docker?s built-in DNS.  When this DNS 
>> is asked for the IP address of the service it will respond with 
>> a list of 3 IP address but the list will rotate in round-robin 
>> fashion each time a lookup is performed.  Thus the first IP in 
>> the list will not be the same for any 2 consecutive lookups.
>> 
>> My first question is:
>> Is it the correct behavior that consistent hashing depends on 
>> the order of IP addresses in the DNS query response?  I can 
>> imagine arguments either way, and it is possible that this 
>> critical detail is outside the scope of consistent hashing.  I 
>> will also forward this question to the author of Ketama.
> 
> Consistent hashing uses the _name_ as written in the 
> configuration, not the IP addresses the name resolves to.
> 
> If a name resolves to multiple IP addresses, these addresses are 
> considered equal and requests are distributed between them using 
> round-robin balancing.
> 
> That is, to balance multiple servers (containers) using consistent 
> hashing, you have to configure an upstream block with multiple 
> "server" directives in it.
> 
> -- 
> Maxim Dounin
> http://mdounin.ru/
>