Initial hex release.
A libcluster strategy that uses
HyParView for membership and connects
only the nodes in the local active view via Erlang distribution — i.e.
partial-mesh BEAM distribution.
Why
libcluster's built-in strategies all assume a full mesh: every discovered
node calls Node.connect/1 for every other discovered node. That works fine
to ~50–100 nodes; past that you start hitting net_kernel pressure, partial
partitions, and chatty heartbeats. HyParView gives each node a bounded
active view of log(N) + c peers — this strategy connects only those.
Install
def deps do
[
{:libcluster_hyparview, "~> 0.1"}
]
endSee the README
for the pre-flight checklist (-connect_all false), full topology config,
and integration notes with other libcluster strategies.
Companion projects
- hyparview — the underlying
membership protocol. Bumped to~> 0.2in this release. - hyparview_pubsub_bench —
benchmark suite comparing partial-mesh + flooding vsPhoenix.PubSub
full-mesh. Documents the connection-count / failure-resilience trade-offs.