GossNet.Discovery.Gcp 0.10.0

GossNet.Discovery.Gcp

Google Cloud Compute Engine node discovery for GossNet.Protocol.

Instances are found by a label they already carry, so nothing registers or deregisters itself — the same shape as GossNet.Discovery.Aws and GossNet.Discovery.Azure. An instance that goes away simply stops matching.

Labels, not network tags. Compute Engine has both and they are not interchangeable. Network tags are unkeyed strings used for firewall rules; labels are the key/value metadata that corresponds to an AWS or Azure tag, and are what this provider matches on.

Installation

dotnet add package GossNet.Discovery.Gcp

Usage

Label every cluster member, then point discovery at that label:

gcloud compute instances add-labels node-a --labels=gossnet-cluster=production

using GossNet.Discovery.Gcp;
using GossNet.Protocol;

var gcpOptions = new GcpDiscoveryOptions
{
    ProjectId = "my-project",
    LabelKey = "gossnet-cluster",
    LabelValue = "production",
    Port = 9055
};

var configuration = new GossNetConfiguration
{
    Hostname = "10.128.0.4",
    Port = 9055,
    DiscoveryProviderFactory = cfg => new GceLabelNodeDiscovery(cfg, gcpOptions)
};

await using var node = new GossNetNode<MyMessage>(configuration);
node.Start();

Hostname must be the address other nodes reach this instance on — normally its internal IP, matching UseInternalIp. Discovery excludes the node from its own results by comparing against it, so a mismatch makes a node gossip to itself.

Credentials and permissions

Credentials come from Application Default Credentials: the attached service account on a VM, GOOGLE_APPLICATION_CREDENTIALS, or gcloud auth application-default login locally.

The identity needs compute.instances.list, which roles/compute.viewer grants:

gcloud projects add-iam-policy-binding my-project \
  --member=serviceAccount:gossnet@my-project.iam.gserviceaccount.com \
  --role=roles/compute.viewer

Running on GKE?

Use GossNet.Discovery.Kubernetes instead. This package is for instances on Compute Engine; on GKE the pods are what you want to discover, not the nodes hosting them.

Options

Compute Engine describes instances, not services, so the port cannot be discovered and must be the same on every node.

Instances are listed aggregated across every zone, since a gossip cluster normally spans zones and asking per-zone would mean knowing the zone list up front. The request sets ReturnPartialSuccess, so one unreachable zone degrades the result rather than failing the whole lookup, and paging is followed in full so large clusters are not truncated.

Only instances in the RUNNING state are returned — stopped and terminated ones keep their labels, but their addresses are gone or reassigned. Instances with no address of the requested kind are skipped rather than returned as unreachable neighbours; note that an external address exists only when the instance has an access config.

A failed API call — throttling, missing permission, no credentials — throws NodeDiscoveryException rather than returning an empty list, so an API problem is never mistaken for a cluster of one.

License

MIT