Hi! I write on behalf of Immobiliare.it, an Italian company leader in
the real estate services and advertising market, we are using almost
exclusively HAProxy for our load-balancing. In particular, we are using
a patched version of HAProxy to balance requests to our cache servers.
Long story short, we set up an improved version of the replicated,
sharded cache pattern: each server both maintains its cache entries and
a subset of entries of the other servers. Doing that, we ensure that a
request for an entity can be sent either to its server (the one provided
by the HAProxy's consistent hashing algorithm) or to its second server
(the next server in the tree after the one chosen by the consistent
hashing algorithm), so we can ensure that all the entries are still
present if a single server crashes, and when many server crashes, only a
little subset of those entries were lost. To do that we created a new
consistent hashing algorithm (consistent-2x) that, once found the server
that has to serve the request, also looks for another server in the tree
(which should reside on different host) and sends the request either to
the first or to the second in a random fashion, still considering
server's weight.
The whole implementation of the above mentioned algorithm requires the
following changes:
the tree we build is not driven by weights but it is static, so if a
server crashes HAProxy does not re-balance requests (consistent-2x
algorithm does this); weights are used instead to chose one of the two
selected servers; servers' IDs should follow a given pattern: servers on
the same machine should give the same result when their IDs is divided
by 1000; consistent-2x algorithm. We are looking forward to hearing from
you. We wish our contribution could be useful for as many people as
possible.
Best,
Dario Di Pasquale
Immobiliare.it
