From 795345e5fd7974a1f5227d507a58bb3ed75eafd5 Mon Sep 17 00:00:00 2001 From: Stefan Boberg Date: Mon, 13 Apr 2026 16:38:16 +0200 Subject: Compute OIDC auth, async Horde agents, and orchestrator improvements (#913) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Rework of the Horde agent subsystem from synchronous per-thread I/O to an async ASIO-driven architecture, plus provisioner scale-down with graceful draining, OIDC authentication, scheduler improvements, and dashboard UI for provisioner control. ### Async Horde Agent Rewrite - Replace synchronous `HordeAgent` (one thread per agent, blocking I/O) with `AsyncHordeAgent` — an ASIO state machine running on a shared `io_context` thread pool - Replace `TcpComputeTransport`/`AesComputeTransport` with `AsyncTcpComputeTransport`/`AsyncAesComputeTransport` - Replace `AgentMessageChannel` with `AsyncAgentMessageChannel` using frame queuing and ASIO timers - Delete `ComputeBuffer` and `ComputeChannel` ring-buffer classes (no longer needed) ### Provisioner Drain / Scale-Down - `HordeProvisioner` can now drain agents when target core count is lowered: queries each agent's `/compute/session/status` for workload, selects candidates by largest-fit/lowest-workload, and sends `/compute/session/drain` - Configurable `--horde-drain-grace-period` (default 300s) before force-kill - Implement `IProvisionerStateProvider` interface to expose provisioner state to the orchestrator HTTP layer - Forward `--coordinator-session`, `--provision-clean`, and `--provision-tracehost` through both Horde and Nomad provisioners to spawned workers ### OIDC Authentication - `HordeClient` accepts an `AccessTokenProvider` (refreshable token function) as alternative to static `--horde-token` - Wire up `OidcToken.exe` auto-discovery via `httpclientauth::CreateFromOidcTokenExecutable` with `--HordeUrl` mode - New `--horde-oidctoken-exe-path` CLI option for explicit path override ### Orchestrator & Scheduler - Orchestrator generates a session ID at startup; workers include `coordinator_session` in announcements so the orchestrator can reject stale-session workers - New `Rejected` action state — when a remote runner declines at capacity, the action is rescheduled without retry count increment - Reduce scheduler lock contention: snapshot pending actions under shared lock, sort/trim outside the lock - Parallelize remote action submission across runners via `WorkerThreadPool` with slow-submit warnings - New action field `FailureReason` populated by all runner types (exit codes, sandbox failures, exceptions) - New endpoints: `session/drain`, `session/status`, `session/sunset`, `provisioner/status`, `provisioner/target` ### Remote Execution - Eager-attach mode for `RemoteHttpRunner` — bundles all attachments upfront in a `CbPackage` for single-roundtrip submits - Track in-flight submissions to prevent over-queuing - Show remote runner hostname in `GetDisplayName()` - `--announce-url` to override the endpoint announced to the coordinator (e.g. relay-visible address) ### Frontend Dashboard - Delete standalone `compute.html` (925 lines) and `orchestrator.html` (669 lines), consolidated into JS page modules - Add provisioner panel to orchestrator dashboard: target/active/estimated core counts, draining agent count - Editable target-cores input with debounced POST to `/orch/provisioner/target` - Per-agent provisioning status badges (active / draining / deallocated) in the agents table - Active vs total CPU counts in agents summary row ### CLI - New `zen compute record-start` / `record-stop` subcommands - `zen exec` progress bar with submit and completion phases, atomic work counters, `--progress` mode (Pretty/Plain/Quiet) ### Other - `DataDir` supports environment variable expansion - Worker manifest validation checks for `worker.zcb` marker to detect incomplete cached directories - Linux/Mac runners `nice(5)` child processes to avoid starving the main server - `ComputeService::SetShutdownCallback` wired to `RequestExit` via `session/sunset` - Curl HTTP client logs effective URL on failure - `MachineInfo` carries `Pool` and `Mode` from Horde response - Horde bundle creation includes `.pdb` on Windows --- src/zencompute/runners/windowsrunner.cpp | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'src/zencompute/runners/windowsrunner.cpp') diff --git a/src/zencompute/runners/windowsrunner.cpp b/src/zencompute/runners/windowsrunner.cpp index 92ee65c2d..e643c9ce8 100644 --- a/src/zencompute/runners/windowsrunner.cpp +++ b/src/zencompute/runners/windowsrunner.cpp @@ -48,7 +48,9 @@ WindowsProcessRunner::WindowsProcessRunner(ChunkResolver& Resolver, if (m_JobObject) { JOBOBJECT_EXTENDED_LIMIT_INFORMATION ExtLimits{}; - ExtLimits.BasicLimitInformation.LimitFlags = JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE | JOB_OBJECT_LIMIT_DIE_ON_UNHANDLED_EXCEPTION; + ExtLimits.BasicLimitInformation.LimitFlags = + JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE | JOB_OBJECT_LIMIT_DIE_ON_UNHANDLED_EXCEPTION | JOB_OBJECT_LIMIT_PRIORITY_CLASS; + ExtLimits.BasicLimitInformation.PriorityClass = BELOW_NORMAL_PRIORITY_CLASS; SetInformationJobObject(m_JobObject, JobObjectExtendedLimitInformation, &ExtLimits, sizeof(ExtLimits)); JOBOBJECT_BASIC_UI_RESTRICTIONS UiRestrictions{}; -- cgit v1.2.3