consistent hashing using split_clients

rmalayter nginx-forum at nginx.us
Wed Oct 31 16:46:09 UTC 2012


Maxim Dounin Wrote:
> 
> Percentage values are stored in fixed point with 2 digits after 
> the point.  Configuration parsing will complain if you'll try to 
> specify more digits after the point.
> 
> > How many "buckets" does the hash table for split_clients
> > have (it doesn't seem to be configurable)?
> 
> The split_clients algorithm doesn't use buckets, as it's not a 
> hash table.  Instead, it calculates hash function of the 
> original value, and selects resulting value based on a hash 
> function result.  See http://nginx.org/r/split_clients for 
> details.
> 

So clearly I am down the wrong path here, and split_clients just cannot do
what I need. I will have to rethink things.

The 3rd-party ngx_http_consistent_hash module appears to be un-maintained,
un-commented. It also uses binary search to find an upstream instead of a
hash table, making it O(log(n)) for each request. My C skills haven't been
used in anger since about 1997, so updating or maintaining it myself would
probably not be a fruitless exercise.

Perhaps I will have to fall back to using perl to get a hash bucket for the
time being. I assume 4096 upstreams is not a problem for nginx given that it
is used widely by CDNs.

A long time ago Igor mentioned he was working on an variable-based upstream
hashing module using MurmurHash3:
http://forum.nginx.org/read.php?29,212712,212739#msg-212739

I suppose other work took priority. Maybe Igor has some code stashed
somewhere that just needs testing and polishing.

If not, it seems that the current "ip_hash" scheme used in nginx could be
easily adapted to fast consistent hashing by simply
  -using MurmurHash3 or similar instead of the current simple
multiply+modulo scheme
  -allowing arbitrary nginx variables as hash input instead of just the IP
address during upstream selection
  -at initialization utilizing a hash table of 4096 or whatever configurable
number of buckets
  -fill the hash table by sorting the server array on
murmurhash3(bucket_number + server_name + server_weight_counter) and taking
the first server

Is there a mechanism for sponsoring development along these lines and
getting it into the official nginx distribution? Consistent hashing is the
one commonly-used proxy server function that nginx seems to be missing.

Posted at Nginx Forum: http://forum.nginx.org/read.php?2,232428,232434#msg-232434



More information about the nginx mailing list