feat: agent-initiated gRPC stream, decoupled image build#7
Merged
Conversation
The control plane no longer dials each agent's HTTP API. Instead every agent holds one persistent gRPC bidirectional stream open to the control plane, which pushes VM lifecycle commands down it; the agent runs them against its local Runtime and answers on the same stream. Agents now need no inbound reachability, and the open stream both delivers commands and serves as the host's liveness signal. - proto/agentlink + generated internal/agentlink/pb: AgentLink.Connect bidi stream. Command/result payloads are JSON of the existing agent.VMSpec/agent.VM types; heartbeats ride the same stream. - internal/agentlink/hub.go: CP-side gRPC server + per-host connection registry; registers hosts (reconstructing committed capacity), correlates results, stamps liveness on heartbeats, and marks a host down the instant its stream drops. - internal/agent/link.go: agent-side dial/register/serve loop with heartbeat ticker and reconnect backoff. - provisioner/remote.go calls the hub by host id (Commander interface); repository.HostRepository gains MarkDown for immediate down-on-disconnect. - Remove the old HTTP path: agent /vms server + client, HTTP register/heartbeat handlers, AgentAuth middleware, and ADVERTISE_ADDR. Agent dials CONTROL_PLANE_GRPC_ADDR; control plane listens on GRPC_PORT (8090). Co-Authored-By: Afshin Arani <afshin@arani.dev>
A cold image build pulls layers lazily, so the layer download streams during the squashfs write and is bound to the context passed in. A short request/command-scoped deadline (a dropped agent stream, the reconciler's 30s batch timeout) cancelled that context mid-stream, aborting a build that legitimately takes minutes — surfacing as "stream layer tar: write file ...: context canceled". Since the .tmp is removed on error, every retry restarted from zero. Build the rootfs under the runtime's lifetime context plus a generous ImagePullTimeout instead of the command ctx. The build is content-addressed, idempotent, and shared across every VM booting the ref, so a cancelled provision now leaves a populated cache for the next attempt rather than killing the work. ImageStore becomes an ImageEnsurer interface so the non-KVM tests can assert the build context. Also drop the reconciler's 30s per-batch timeout: nothing in the reconciler should impose a command/request-scoped deadline on provisioning. Operations run under the reconciler lifetime, bounded by the agent-side pull timeout and process shutdown. Co-Authored-By: Afshin Arani <afshin@arani.dev>
The e2e suite was wired for the old "control plane dials agent over HTTP" model (agent.NewRouter/NewClient, provisioner.NewRemote(hostRepo, ...), and HTTP /agent/hosts register+heartbeat endpoints), all removed when the control channel inverted to an agent-dialed gRPC stream. The suite no longer compiled. - main_test.go: stand up the real agentlink.Hub on a localhost gRPC listener and an in-process placement agent via agent.RunLink over a FakeRuntime, exactly as production wires it; provisioner.NewRemote(hub). The host registers itself over the stream, so wait on the fleet inventory for it to come ready before running tests. Adds startAgent() for ad-hoc per-test agents. - agent_test.go: the agent has no inbound API anymore; query the in-process FakeRuntime directly to confirm the VM landed on the agent. - hosts_test.go: rewrite host lifecycle/identity as real gRPC flows (connect->ready, disconnect->down, reconnect->ready; agent-supplied id authoritative and updated in place on reconnect). Drop the HTTP request-validation tests (bad id, missing fields, unknown heartbeat) — no analog on the typed gRPC Register path; that seam is unit-covered in internal/agentlink/hub_test.go. Full suite passes against a live Postgres testcontainer. Co-Authored-By: Afshin Arani <afshin@arani.dev>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.