Repeated Bazel test targets with --runs_per_test at 10, 20, 50, 100, and
200 runs. The tests failed intermittently when multiple instances ran
concurrently. This indicates timing, shared tracing state, socket readiness,
or parallel-safety issues that need focused follow-up before re-enabling the
tests.
A normal non-repeated bazel test //... run passes with these tests ignored.
Disabled tests:
- sandbox_forward_foreground_fails_when_ssh_exits_before_listener_opens
- sandbox_forward_background_terminates_owned_child_when_listener_never_opens
- podman_socket_probe_accepts_successful_ping_response
- podman_socket_probe_rejects_docker_ping_response
- docker_socket_probe_accepts_successful_ping_response
- docker_socket_probe_rejects_podman_ping_response
- docker_socket_detection_returns_the_responsive_candidate
- podman_socket_detection_returns_the_responsive_candidate
- driver_watch_events_are_roots_and_store_operations_have_parents
- reconcile_sweeps_are_roots_and_operations_have_parents
- gateway_listeners_bind_ipv6_wildcard_and_ipv4_callback_on_same_port
- watch_producer_releases_request_span_when_client_disconnects
- expected_conflicts_leave_the_span_unmarked
- store_spans_record_what_they_touched_as_attributes
- store_operations_export_spans_with_parents
- refresh_worker_ticks_are_roots_and_store_operations_have_parents
Signed-off-by: Simon Scatton <sscatton@nvidia.com>
Continue distributed traces across the gateway-to-driver process boundary and
export VM driver spans to the same OTLP/gRPC collector. The driver reports as
the distinct openshell-driver-vm service.
Updated the gateway architecture and configuration reference with a generic
external-driver forwarding contract.
Instrumented:
- Every RemoteComputeDriver RPC injects the active W3C trace context into
tonic metadata. Managed VM readiness and runtime initialization give startup
capability probes stable parent operations rather than isolated root spans.
- A tonic service layer creates fixed, low-cardinality server spans for every
ComputeDriver RPC. New handlers inherit tracing automatically; failures record
OpenTelemetry error status and the gRPC status code.
- Background provisioning remains attached to CreateSandbox after the RPC
returns without extending the RPC span lifetime.
- Provisioning records image preparation, bootstrap image resolution, overlay
preparation, lifecycle configuration, pre-launch hooks, guest preparation,
and launcher spawn as child spans.
- VM startup reconciliation roots one trace for the persisted-sandbox scan,
with per-sandbox restore and provision operations beneath it. The root remains
open until all spawned restore tasks finish.
- Delete cleanup records its own child operation.
Design notes:
- The gateway forwards its configured OTLP endpoint to managed external
drivers. SDK `OTEL_*` variables continue to own sampling, batching, limits,
headers, and transport tuning.
- The VM driver has its own tracer provider and service resource so trace
backends preserve the service boundary.
- RPC operation names come from an explicit method mapping, keeping cardinality
bounded without parsing the protobuf descriptor set at runtime.
- Propagation uses a remote SpanContext for spawned provisioning. This keeps
one trace while allowing the CreateSandbox server span to finish when the
RPC response is sent.
- Startup restoration is independent of gateway requests. It begins at the VM
driver reconciliation span rather than attaching to an unrelated RPC.
- Existing tracing events remain on the logging path. The OpenTelemetry layer
exports spans only and excludes the SDK exporter callsites to avoid recursive
traces.
- Export configuration failures do not prevent the driver from serving, and
buffered spans are drained during graceful shutdown.
- Trace fields identify drivers, sandboxes, images, lifecycle phases, and gRPC
outcomes without recording credentials, sandbox tokens, or request query
parameters.
Refs #2507
Signed-off-by: Kris Hicks <khicks@nvidia.com>
Give each span-assertion test its own thread-scoped in-memory exporter so
parallel tests cannot contaminate or reset captured spans. Keep a bare global
tracing registry only to preserve callsite interest, and serialize scoped
subscriber changes because tracing caches that interest process-wide.
Remove the test-only OTLP collector, polling delivery barrier,
transport-specific test, and direct opentelemetry-proto dependency. Seed the
expected persistence conflict before tracing begins so its assertion window
contains only the operation under test.
Signed-off-by: Kris Hicks <khicks@nvidia.com>
Add an opt-in OTLP/gRPC trace exporter to the gateway. Export is enabled
by the presence of an `[openshell.gateway.otlp]` table with an endpoint;
there is no separate toggle.
Instrumented:
- Inbound request server spans, named for the RPC (`$service/$method`) or
`{method} {path}` for plain HTTP. They continue valid W3C `traceparent`
context when present and start a new trace otherwise. gRPC spans also
carry `rpc.system`, `rpc.service`, `rpc.method`, and trailer-derived
`rpc.grpc.status_code`.
- Compute driver calls (create, delete, list, get, validate, watch) as
client spans anchored on the `ComputeDriver` contract.
- Store reads and writes as children of the current request or loop span.
- Work with no inbound request: compute driver initialization, the sandbox
reconcile sweep, provider credential refresh tick, and driver watch events.
Each roots one operation trace so its child work does not arrive as anonymous
single-span traces.
This is deliberately not exhaustive. Auth, policy evaluation, and
middleware remain uninstrumented, as do store lifecycle calls (`ping`,
`close`) that a readiness poll would turn into a span per tick. The aim is
a useful trace tree at a reviewable size; coverage can grow against real
traces.
Design notes:
- The TOML table owns whether and where to export. The SDK `OTEL_*`
variables own how; sampling, batching, and limits are not mirrored into
gateway config.
- The OpenTelemetry layer exports spans only. Existing `tracing` events
remain on the stdout and sandbox-log paths and are not copied into trace
payloads.
- Telemetry never blocks the gateway. A malformed endpoint logs an error
and disables export rather than failing startup, and buffered spans are
drained during graceful shutdown.
- Failed spans carry error status without a separate `error.type` attribute.
Request spans use HTTP status and gRPC response trailers; driver spans use
the returned gRPC status; autonomous loop spans record failed results
explicitly. Store spans exempt `UniqueViolation` and `Conflict`, because
those errors report expected contention such as a held lease or an
optimistic-concurrency retry.
- The compute driver is reachable only through `TracedDriver::call`, so a
call cannot skip its span. This is the client half of a client/server pair
and the single place to inject context if drivers move out of process.
- Tests share one process-wide subscriber and in-memory exporter because
`tracing` caches callsite interest globally.
Inbound W3C trace context is propagated into gateway request spans. Context
is not yet injected into outbound driver calls, so a future out-of-process
driver would still need propagation at the `TracedDriver` seam.
The Helm chart is intentionally unchanged, so OTLP export cannot yet be
enabled on a chart-deployed gateway.
Refs #2507
Signed-off-by: Kris Hicks <khicks@nvidia.com>
* refactor(compute): extract create_sandbox_record and update_sandbox_record helpers
Split apply_sandbox_update_locked into named helpers to make the two
distinct paths explicit: create_sandbox_record for first-observation events
and update_sandbox_record for subsequent driver snapshots on existing sandboxes.
The dispatcher now uses a match on the existing record rather than an
early-return guard.
No behavior change.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* refactor(compute): make sandbox readiness gateway-owned across all drivers
Introduce compute_phase_components and apply_readiness_conditions to
centralise the gateway's phase composition logic. The public SandboxPhase
is now determined by combining the backend phase reported by the driver with
supervisor session presence, independent of the driver implementation.
Remove SupervisorReadiness from the driver contract. Running containers
always report BackendReady; the gateway owns the Ready decision. Rename
the dispatcher match to three arms so that status-less events for existing
sandboxes are a documented no-op rather than a silent pass-through.
Drop backend_ready_no_session and the SupervisorNotConnected condition.
The BackendReady driver condition plus the Provisioning phase already
communicates that the backend is up but the supervisor has not connected.
The redundant condition added noise without new information.
Closes#1951
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* fix(compute): expose disconnected supervisor readiness
Signed-off-by: Drew Newberry <anewberry@nvidia.com>
* docs(sandboxes): clarify supervisor readiness lifecycle
Signed-off-by: Evan Lezar <elezar@nvidia.com>
---------
Signed-off-by: Evan Lezar <elezar@nvidia.com>
Signed-off-by: Drew Newberry <anewberry@nvidia.com>
Co-authored-by: Drew Newberry <anewberry@nvidia.com>
The Kubernetes driver's default workspace PVC never set storageClassName,
so on clusters with no default StorageClass the PVC stayed Pending and
sandbox creation failed.
Add a workspace_storage_class option to KubernetesComputeConfig, wired
through SandboxPodParams into the generated volumeClaimTemplates. When
non-empty it sets storageClassName; empty preserves the current behavior
of relying on the cluster default StorageClass.
Expose it via the OPENSHELL_K8S_WORKSPACE_STORAGE_CLASS env var on both
the standalone driver and the embedded gateway runtime defaults, and via
the server.workspaceStorageClass Helm value.
Closes#2442
Signed-off-by: lr90 <qiuweimin@matrixorigin.cn>
* fix(core): avoid i64 underflow on expired GCP token
When a cached GCP access token was already expired (expires_at_ms < now),
the remaining lifetime calculation used plain subtraction:
(expires_at_ms - now) / 1000. This underflows in debug builds and wraps
in release builds.
Use saturating_sub so an expired token simply yields a non-positive
remaining lifetime and is skipped. Add a regression test.
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(server): avoid i64 underflow in lease age calculation
The lease steal path computed age as now_ms() - record.updated_at_ms.
If the stored updated_at timestamp is in the future (e.g. clock skew),
the subtraction underflows. Use saturating_sub to treat a future
timestamp as age 0, keeping the lease considered fresh.
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(server): avoid i64 overflow on large credential max lifetime
max_lifetime_seconds is only validated as >= 0. The code multiplied the
raw i64 value by 1000 to compute max_expires, which overflows for
values above i64::MAX / 1000. The capped i32 value used for the STS
request was already computed but not reused.
Compute max_lifetime_ms from the capped value with saturating_mul and
use it for both expires_at fallback and max_expires.
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(server): avoid hardcoded indexing into base_url_config_keys
resolve_vertex_ai_route assumed every inference provider profile has at
least two base URL config keys by indexing [0] and [1]. Only the Vertex
profile satisfies that today; other profiles would panic here.
Iterate base_url_config_keys with find_map instead, and add a test with
an empty key list to ensure no panic.
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(server): clamp lease age at zero for future timestamps
Address review feedback: add regression coverage for future-timestamp
lease clock skew. Extract lease_is_expired() and clamp the age at zero —
i64::saturating_sub saturates at i64::MIN, not zero, so a future
updated_at_ms must be clamped explicitly to be treated as age zero and
remain unstealable with a positive TTL.
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(server): fall back past blank preferred Vertex base URL
Address review feedback: find_map stopped at the first present key even
when its value was blank, and the outer filter discarded it without
considering lower-priority aliases. Filter blank values inside the
closure so a blank preferred key falls back to a valid alias, matching
the documented priority order. Add a regression test with a blank
preferred key and a valid fallback.
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(server): saturate AWS STS expiry cap on saturated clock
Address review feedback: now_ms + max_lifetime_ms could still overflow
when current_time_ms() saturates near i64::MAX. Compute
max_expires = now_ms.saturating_add(max_lifetime_ms) once and use it for
both the conversion fallback and the expiry cap. Add a boundary unit
test covering a saturated clock.
Signed-off-by: Andrew White <andrewh@cdw.com>
---------
Signed-off-by: Andrew White <andrewh@cdw.com>
* fix(driver-podman): resolve Podman socket via auto-detection
Align Podman socket selection with the existing Docker model: explicit
config wins, otherwise probe openshell-core for a responsive socket.
This also fixes the original HOME-unset panic, since resolution no
longer hardcodes a per-OS default path.
- add detect_podman_socket() in openshell-core, mirroring
detect_docker_socket
- PodmanComputeConfig.socket_path is now Option<PathBuf>, no default
- remove default_socket_path() (podman driver) and podman_socket_path()
(vm driver), both replaced by the shared detector
- update server env override and CLI for the new Option type
- add tests: responsive-candidate detection in openshell-core, and
config-error (not panic) when no socket is configured or reachable
* fix(driver-podman): address review feedback on socket resolution
Extract socket resolution into resolve_socket_path, taking the
detector as a parameter so tests do not depend on real env vars or
the host's actual Podman state. Replace the flaky env-mutating test
with three deterministic cases: explicit wins, detected is used when
absent, and neither source errors.
Fix the socket_path doc comment to describe it from a config user's
point of view, matching DockerComputeConfig's docstring. Drop a
comment that only made sense next to the Docker driver code.
Update the Podman README and gateway docs: they described a fixed
per-OS default path that no longer exists, replace with the actual
probe-then-fail behavior.
* docs(driver-podman): simplify socket default description
Previous wording was self-contradictory (says auto-detect on unset,
then lists the same var as a probed candidate) and omitted the Linux
/run/user/uid/podman/podman.sock candidate.
* feat(workspace): implement workspace model (Phase 1 of RFC 0011)
Implements workspace and membership model providing hard isolation
boundaries for multi-player OpenShell deployments.
Workspace CRUD with Kubernetes-style Terminating phase for graceful
deletion. All resources scoped by workspace via ObjectMeta. Membership
RPCs for workspace access control. Persistence migration shifts name
uniqueness to (object_type, workspace, name). Provider profiles support
platform and workspace scoping. Service routing uses workspace-prefixed
DNS labels. Inference routes renamed and workspace-scoped with
DeleteInferenceRoute RPC. Python SDK with WorkspaceClient, two-method
list pattern (workspace-scoped and for_all_workspaces), and workspace
parameter on all methods. CLI workspace flags, TUI workspace cycling.
K8s driver filters unmanaged CRs and uses delete preconditions. Podman
driver uses immutable container IDs. Label serialization fixed across
all put_if call sites.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(cli): delegate sandbox upload command to existing upload function
The standalone `sandbox upload` command reimplemented upload logic
inline with two bugs: it used `Path::exists()` which follows symlinks
(rejecting dangling symlinks), and it ran git-aware filtering on
symlink sources. The `run::sandbox_upload()` function already handles
both cases correctly via `sandbox_upload_plan()`. Replace the inline
logic with a call to the existing function.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(e2e): shorten sandbox names and fix test compatibility
Shorten the sandbox name in initial_sparse_policy_is_acknowledged_as_loaded
from 'e2e-2159-sparse-enrich' (22 chars) to 'e2e-sparse-enrich' (17 chars)
to comply with MAX_ROUTABLE_NAME_LEN (19 chars).
Also capture stderr in create_keep_with_args so future sandbox creation
failures include the actual CLI error instead of reporting empty output.
Signed-off-by: Derek Carr <decarr@redhat.com>
* test(workspace): add test coverage for workspace CRUD and persistence isolation
Add unit tests for workspace create happy path, get round-trip, get
not-found, get empty-name rejection, already-exists error, and
resolve_workspace not-found. Add persistence test proving cross-workspace
name uniqueness (same name in different workspaces produces separate
records). Add workspace name max-length boundary tests. Fix e2e harness
to include stderr in name-parse-failure error path. Align Python e2e
test_workspace_crud with try/finally pattern. Document provider profile
catalog workspace scoping gap in RFC 0011.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(examples): update examples for workspace model compatibility
Shorten sandbox names in demo scripts to fit the 19-character
MAX_ROUTABLE_NAME_LEN limit: policy-demo prefix to pd-, multi-agent
notepad derives a short SANDBOX_TAG from the run ID, governance
interceptor uses gs-PID-RANDOM. Update vscode-remote-sandbox.md SSH
host aliases from openshell-{name} to openshell-{name}.{workspace}
format.
Signed-off-by: Derek Carr <decarr@redhat.com>
* feat(sdk): add workspace-scoped client and workspace CRUD
Add WorkspaceScopedClient modeled after kube::Api::namespaced — captures
workspace once and injects it into every sandbox request. Add workspace
CRUD methods (create, get, list, delete) and list_sandboxes_all_workspaces
on OpenShellClient. Extend SandboxRef with workspace field and add
WorkspaceRef type. Include mock tests for all new operations.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(lint): resolve clippy warnings in workspace test assertions
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(docs): convert indented code blocks to fenced in RFC 0011
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(lint): resolve clippy warnings and apply cargo fmt across workspace
Auto-format with cargo fmt and fix clippy warnings exposed by the
reformat: unnecessary qualifications, map_unwrap_or, identical match
arms, unused variable prefix, dead code annotations, and let-unit-value
in e2e harness.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(workspace): address workspace scoping issues from review
- Add workspace field to settings JSON output (CLI)
- Skip Podman containers missing workspace label instead of defaulting
to empty string, matching K8s driver behavior
- Add resource_version to list_by_scope SELECT in both SQLite and
Postgres backends, with regression test
- Gate PolicyLocalContext proposal/lookup routes on workspace readiness,
returning 503 when workspace is not yet discovered
- Block sandbox and provider creation in TUI all-workspaces mode
- Clear workspace vectors in TUI reset_sandbox_state
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(workspace): make provider profile catalog workspace-aware
Thread workspace through snapshot_catalog so the
EffectiveProviderProfileCatalog enforces workspace boundaries on both
read and write paths. UserProviderProfileSource now loads platform-scoped
profiles (workspace "") plus the target workspace's profiles, preventing
cross-workspace duplicate profile ID collisions that previously caused
global catalog failures.
Update RFC 0011 to reflect catalog scoping is implemented in Phase 1
rather than deferred to future work.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(persistence): include workspace column in atomic policy revision INSERT
put_policy_revision_atomic omitted the workspace column from the INSERT
into the objects table in both SQLite and Postgres backends, causing
atomically-written policy revisions to lose their workspace association.
Add workspace field to AtomicPolicyRevisionWrite and thread it through
both backend INSERT statements, matching the non-atomic put_policy_revision
path which already included it.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(proxy): skip ancestor walk when socket owner is the entrypoint
collect_ancestor_identities walked the entire process tree above the
entrypoint when the connecting process was the entrypoint itself,
SHA256-hashing every ancestor binary (IDE, shell, container runtime).
On dev machines with large binaries in the ancestor chain this exceeded
the 30-second test timeout. When start_pid == stop_pid there are no
intermediate ancestors to verify, so return an empty list immediately.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(workspace): make provider profile catalog scope-aware
Allow the same profile ID at platform and workspace scopes by
introducing layered catalog entries where workspace profiles shadow
platform profiles. Add source and scope fields to the ProviderProfile
proto and CLI output. Migrate List/Get handlers to the catalog,
fixing divergence with runtime profile resolution.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(e2e): align podman e2e labels with centralized driver constants
The podman driver moved its container labels to the centralized
openshell.ai/ prefix, but the e2e test harness and cleanup script
still referenced the old openshell.sandbox-* keys, causing the
local_driver_token_restart test to fail on container lookup.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(e2e): align python profile isolation test with scope-aware catalog
Platform profiles are now visible in workspace listings as fallbacks
per the layered catalog design. Update the assertion to match.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(workspace): honor profile_workspace in runtime profile resolution
Runtime profile lookups now consult provider.profile_workspace via
get_type_profile_for_scope. Providers created with --global-profile
(profile_workspace="") resolve to the platform profile even when a
workspace profile shadows the same ID. All 6 runtime call sites
updated; type-only call sites remain scope-agnostic.
Signed-off-by: Derek Carr <decarr@redhat.com>
---------
Signed-off-by: Derek Carr <decarr@redhat.com>
Previously, Docker was auto-detected when the CLI was installed or a candidate
Unix socket existed. Neither check verified that the Docker API was responsive.
A similar check was done when auto-detecting Podman in the past, but was
replaced in 1f07bf04 with a probe of candidate Podman sockets instead.
This change applies the functional API probing approach introduced for Podman
in 1f07bf04 to Docker. It also makes Docker driver initialization use the same
socket-selection mechanism as Docker auto-detection instead of Bollard’s local
defaults. This means the previously auto-detectable Docker socket paths
$HOME/.docker/run/docker.sock and $XDG_RUNTIME_DIR/docker.sock will actually be
usable.
When no working compute driver can be auto-detected, the gateway exits early
with a message saying as much:
> configuration error: no compute driver configured and auto-detection found no
> suitable driver; set --drivers or OPENSHELL_DRIVERS to kubernetes, podman,
> docker, or vm
This makes for a better user experience when installing OpenShell without an
available supported compute driver.
Signed-off-by: Kris Hicks <khicks@nvidia.com>
* feat(cli)!: fold gateway metadata into list
BREAKING CHANGE: openshell gateway info no longer shows local gateway registration metadata. Use openshell gateway list or openshell gateway list -o json for local registration details.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* feat(gateway): add elevated gateway info
Signed-off-by: Evan Lezar <elezar@nvidia.com>
---------
Signed-off-by: Evan Lezar <elezar@nvidia.com>
The field was added during the Kubernetes driver extraction refactor
(#817) as a pass-through mechanism, but was never wired up to a CLI
flag, Python SDK helper, or any documentation. The only reachable
user path was raw gRPC construction.
The Kubernetes driver now always injects the default workspace PVC,
removing the branching logic that checked for a user-supplied VCT.
Field number 9 is reserved in the proto to prevent reuse.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* feat(server): support remote compute driver endpoints
Add named remote compute driver endpoint support to the gateway.
Remote drivers are selected by a non-reserved compute driver name and either a CLI/env socket endpoint or [openshell.drivers.<name>].socket_path. The VM driver now enters ComputeRuntime through the same acquired remote endpoint path, while Docker, Podman, and Kubernetes retain their in-process drivers.
Require --drivers/OPENSHELL_DRIVERS when pairing an ad-hoc socket endpoint so the socket does not imply a magic driver name, and keep reserved in-tree names unavailable for unmanaged socket endpoints.
Co-authored-by: Evan Lezar <elezar@nvidia.com>
Signed-off-by: st-gr <38470677+st-gr@users.noreply.github.com>
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* test(server): cover remote compute driver UDS lifecycle
Signed-off-by: Evan Lezar <elezar@nvidia.com>
---------
Signed-off-by: st-gr <38470677+st-gr@users.noreply.github.com>
Signed-off-by: Evan Lezar <elezar@nvidia.com>
Co-authored-by: Evan Lezar <elezar@nvidia.com>
* feat(gpu)!: add resource requirements
BREAKING CHANGE: SandboxSpec.gpu and DriverSandboxSpec.gpu were replaced with resource_requirements.gpu, changing protobuf field 9 from a bool to a message for both public and driver APIs.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* refactor(gpu): pass requirements through sandbox create
Pass the coupled GPU requirement object through the CLI sandbox_create boundary instead of splitting presence and count into separate arguments.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* refactor(gpu): pass requirements to timeout message
Pass ResourceRequirements into the provisioning timeout message helper so GPU hints are derived from the same nested request object used to create the sandbox.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* refactor(gpu): pass driver requirements through helpers
Thread Option<GpuResourceRequirements> through driver validation and rendering helpers instead of splitting GPU presence and count into separate arguments.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
* fix(gpu): validate exact device requests
Require exact driver GPU device lists to be tied to a GPU request, allow a single exact device to use the default countless request, and require explicit matching counts for multi-device lists.
Signed-off-by: Evan Lezar <elezar@nvidia.com>
---------
Signed-off-by: Evan Lezar <elezar@nvidia.com>
Introduce a database-backed lease that ensures only one gateway replica
runs the watch and reconcile loops. Includes lease primitives with CAS
safety, cooperative cancellation via watch channels, SQLite bypass for
single-replica deployments, and integration tests covering failover,
contention, and CAS chain integrity.
Signed-off-by: Derek Carr <decarr@redhat.com>
Legacy or never-initialized sandbox rows decode to the proto-default
Unspecified phase. The startup resume sweep skipped them, leaving the
backend unreconciled: never resumed, never errored. Unspecified was the
only non-terminal phase skipped, since unrecognized values already fall
through to Unknown and get resumed.
Treat Unspecified as running so the sweep reconciles it from live driver
state: resume when the backend resource exists, mark Error when it is gone.
Tests
- cargo test -p openshell-server \
resume_persisted_sandboxes_resumes_running_phases
Signed-off-by: Shiju <shiju@nvidia.com>
* refactor(proto): move phase and current_policy_version into SandboxStatus
Move phase and current_policy_version from SandboxSpec into
SandboxStatus to correctly model mutable runtime state. Update all
callers in the gateway server, TUI, and Python SDK to read and write
these fields through SandboxStatus accessors.
Signed-off-by: Derek Carr <decarr@redhat.com>
* fix(server): preserve sandbox status on statusless driver updates
When a driver update arrives without a status payload (e.g. before
Kubernetes populates the status subresource), preserve the stored
phase, conditions, and current policy version instead of resetting
them. Adds a regression test covering the edge case.
Signed-off-by: Derek Carr <decarr@redhat.com>
---------
Signed-off-by: Derek Carr <decarr@redhat.com>
In the Rust ecosystem there's largely three ways to do system calls:
- raw libc
- nix
- rustix
Of the three, libc is almost all `unsafe` and really 95% of use
cases should be either nix or rustix. nix is the original one,
but after having looked at the code of both, I think rustix
is just better designed and organized. It's also reached 1.0,
whereas nix is still making semver-breaking changes (in fact
we're behind here in this project).
Now in practice, we have both *transitively* in the depchain
already, and that's true for quite a lot of projects.
But I think rustix is better, so let's add rustix as
a workspace dependency (process feature) and migrate
a few use cases to it - it's especially better than the raw
libc which is suprisingly widespread.
If we agree to do this, then many other calls can be ported.
Signed-off-by: Colin Walters <walters@verbum.org>
Move default_sandbox_image() and SUPERVISOR_IMAGE_BINARY_PATH into
openshell-core so all compute drivers share a single authoritative
definition. Eliminates four identical copies of default_sandbox_image()
and two identical copies of SUPERVISOR_IMAGE_BINARY_PATH.
Implement ProviderPlugin for ProviderDiscoverySpec to remove the
repeated three-method delegation boilerplate from eight simple
provider modules. Each now declares only a SPEC constant; the blanket
impl handles id(), discover_existing(), and credential_env_vars().
Add resource_version-based optimistic concurrency control to the
persistence layer. Every write now requires an explicit WriteCondition
(MustCreate or MatchResourceVersion), enforced at compile time by
gating unconditional put/put_message behind #[cfg(test)].
- Add WriteCondition enum and put_if for conditional writes
- Add update_message_cas for atomic read-modify-write operations
- Add list_messages/list_messages_with_selector helpers that hydrate
resource_version from authoritative DB rows
- Convert all production write paths to CAS-aware methods
- Gate put/put_message behind #[cfg(test)] to prevent non-CAS writes
- Use structured PersistenceError::UniqueViolation matching instead of
string matching for duplicate detection
- Hydrate resource_version from WriteResult directly on creates,
eliminating unnecessary read-after-write round trips
Signed-off-by: Derek Carr <decarr@redhat.com>
* feat(gateway): add TOML configuration file (RFC 0003)
Introduces an opt-in --config / OPENSHELL_GATEWAY_CONFIG flag that loads a
TOML file with gateway-wide settings and per-driver tables. Source
precedence is CLI > env > file > built-in default, implemented via clap's
ValueSource so existing flags and env vars keep their priority.
Driver crates (kubernetes, docker, podman, vm) now derive Deserialize on
their config structs. SupervisorSideloadMethod gains Deserialize with
kebab-case rename. A per-driver inheritance allowlist on the loader side
overlays [openshell.gateway] shared defaults (default_image,
supervisor_image, image_pull_policy, guest_tls_*, ssh_handshake_skew_secs,
client_tls_secret_name, host_gateway_ip, enable_user_namespaces) onto
each [openshell.drivers.<name>] table before deserialization.
The Helm chart renders a new gateway-config ConfigMap and mounts it at
/etc/openshell/gateway.toml. The migrated OPENSHELL_* env entries are
dropped from the StatefulSet — only the Secret-backed
OPENSHELL_SSH_HANDSHAKE_SECRET remains. database_url stays on --db-url.
Adds examples/gateway/gateway.example.toml and updates architecture/gateway.md
with the source precedence and inheritance rules.
* docs(gateway): drop ssh_handshake_skew_secs and ssh_handshake_secret from examples
Both fields are scheduled for removal. Remove the example values and the
env-only note so the gateway.toml example and the architecture doc stop
recommending settings that will not exist much longer.
* docs(rfc): correct OPENSHELL_CONFIG to OPENSHELL_GATEWAY_CONFIG in RFC 0003
* docs(gateway): add per-driver TOML example configurations
Adds focused single-driver examples next to the comprehensive
gateway.example.toml: kubernetes, docker, podman, and microvm. Each one
demonstrates the realistic settings for that driver plus how shared
[openshell.gateway] defaults inherit into the driver table.
A new unit test (`checked_in_examples_parse`) loads every example through
the config_file loader so schema drift fails CI rather than silently
shipping a broken example.
* refactor(gateway): drop image_pull_policy from shared inheritance
Kubernetes and Podman use mutually-incompatible vocabularies for the same
TOML key:
- Kubernetes: `Always | IfNotPresent | Never` (free-form string passed
verbatim to the K8s API).
- Podman: `always | missing | never | newer` (strict lowercase enum
deserialised into `ImagePullPolicy`).
No value means the same thing in both drivers. Sharing the key at
`[openshell.gateway]` scope and inheriting it into every active driver's
table meant any value safe for one driver was either wrong or silently
dropped for the other (`IfNotPresent` → `ImagePullPolicy::Missing` after
`.unwrap_or_default()`). Operators run one driver per gateway, so the
"shared default" never pays for itself.
Make `image_pull_policy` driver-local:
- Remove the field from `GatewayFileSection` and from
`inheritable_keys()` for both Kubernetes and Podman.
- Drop the file→`RunArgs` merge for the gateway-scope key.
- Stop unconditionally clobbering the driver value with
`config.sandbox_image_pull_policy` in the runtime wiring — only apply
the CLI/env override when it was set (and, for Podman, only when it
parses into the lowercase enum).
- Move the key under `[openshell.drivers.kubernetes]` and
`[openshell.drivers.podman]` in every example, the RFC, the
architecture doc, and the Helm-rendered gateway ConfigMap.
The supervisor pull policy follows the same shape: it is K8s-only and
moves into `[openshell.drivers.kubernetes]` alongside `image_pull_policy`
in the Helm template.
* fix(gateway): address review feedback on TOML configuration
Resolves the P1 and P2 issues raised in PR #1317:
- Helm gateway ConfigMap moves `grpc_endpoint` under
`[openshell.drivers.kubernetes]` so the default install no longer fails
the gateway's `deny_unknown_fields` schema check.
- `kubernetes_config_from_file` and `podman_config_from_file` only let
the gateway-wide CLI/env `grpc_endpoint` overwrite the driver-table
value when it was actually supplied, preserving file-only configs.
- Kubernetes driver default `image_pull_policy` is now empty (was Podman
vocabulary "missing"), so default deployments let the Kubernetes API
apply its own policy instead of being rejected.
- New `disable_tls` gateway field plumbs `.Values.server.disableTls`
through the TOML ConfigMap instead of relying on env vars dropped from
the StatefulSet.
- StatefulSet pod template now carries a `checksum/gateway-config`
annotation so `helm upgrade` rolls pods when the ConfigMap changes.
- Auxiliary listener resolution preserves the full `SocketAddr` from
`health_bind_address` / `metrics_bind_address`, so a loopback-pinned
health port is not silently relocated onto the public bind address.
- `ssh_session_ttl_secs` from the file is now applied to `Config` (it
was previously accepted by the loader but never read).
New regression coverage: cli-level merge tests for the new fields plus
helm-unittest assertions for the ConfigMap shape, checksum annotation,
and `disable_tls` rendering.
* docs(gateway): consolidate gateway TOML examples into docs reference
Replaces the per-driver example files under examples/gateway/ with a
single published reference page at docs/reference/gateway-config.mdx
covering source precedence, layout, the full example, and the four
per-driver examples (Kubernetes, Docker, Podman, microVM). Drew flagged
during PR #1317 review that the examples belong with the user-facing
docs rather than in a sibling examples/ directory.
The cross-references in architecture/gateway.md and RFC 0003 are updated
to point at the new docs page; the round-trip test in config_file.rs is
removed (schema coverage stays on the inline parses_full_example test
and per-field merge tests — doc-snippet drift belongs in a separate
docs-lint, not in a cross-tree Rust unit test).
* refactor(core): move DEFAULT_K8S_NAMESPACE into K8s driver
The constant is Kubernetes-specific (used only by KubernetesComputeConfig's
Default impl) and does not belong in openshell-core. Relocate it to the
driver crate that owns the K8s vocabulary; openshell-core retains only
truly cross-cutting defaults.
* refactor(core): move Podman bridge default into Podman driver
DEFAULT_NETWORK_NAME is Podman vocabulary, consumed only by the Podman
driver. Also drops the unused DEFAULT_IMAGE_PULL_POLICY constant.
* docs(auth): scrub remaining SSH handshake secret references
Sweeps the trailing mentions left after the rebase: the gateway
config-file module doc, the Helm gateway-config ConfigMap header,
the gateway-config.mdx env-only note, and the RPM systemd unit
comment for init-gateway-env.sh.
* docs(gateway): clarify OPENSHELL_GRPC_ENDPOINT applies to all drivers
The previous comment implied the callback endpoint was Kubernetes-only,
but the value is propagated to every compute driver (Kubernetes, Docker,
Podman, VM) and must be reachable from wherever the sandbox runs.
* refactor(gateway): move driver options into config (#1394)
Signed-off-by: Drew Newberry <anewberry@nvidia.com>
* fix(e2e): regenerate gateway config via TOML for docker + podman harnesses
The gateway CLI flags moved into TOML config tables in 560550d2 (#1394),
which made every existing e2e/with-{docker,podman}-gateway.sh invocation
fail with "unexpected argument '--sandbox-namespace'" (and a long tail
of similar driver-specific options) before the gateway could even bind.
Replace the obsolete CLI flags with a synthesized
`[openshell.drivers.<driver>]` table written to `${STATE_DIR}/gateway.toml`
and passed via the new `--config` flag. Only the gateway-wide flags
that survived 560550d2 (bind-address, port, drivers, db-url, tls-*,
disable-tls, log-level, health-port) stay on the command line.
Both scripts get a small `toml_string` helper to properly TOML-quote
the values (the previous `%q` printf format produced bash-escape, not
TOML-escape). The Docker harness also corrects two field names that
diverged from the driver schema: `docker_network_name` →
`network_name`, and the supervisor binary/image plumbing now reads
through to `supervisor_bin` / `supervisor_image` in the same table.
The Podman harness drops `--ssh-gateway-port` (deleted in 560550d2 —
gRPC + SSH are multiplexed on the same port now) and substitutes
`network_name` + `gateway_port` for the obsolete `--sandbox-namespace`
(which the Podman driver never had as a typed field).
* fix(core): swap bind-only 0.0.0.0 SSH gateway host for cluster URL host
CLI's resolve_ssh_gateway treated 0.0.0.0 as a loopback "keep as-is"
when the cluster URL was also loopback, so the SSH proxy connected to
0.0.0.0:port. The unspecified address is never a valid connect target
and is not present in any TLS cert SAN, which produced BadCertificate
TLS handshake failures during `openshell sandbox create -- ...` in
docker/podman e2e (e.g. bypass_detection).
Resolution: when the server returns 0.0.0.0 or :: as the gateway host
and both endpoints are loopback, fall back to the cluster URL's host
(which the CLI is already using to reach the gateway, so it must
resolve and match the cert).
* fix(e2e): repair podman harness on macOS
Podman 5.x with the applehv/libkrun provider no longer creates the legacy
~/.local/share/containers/podman/machine/podman.sock symlink, and
`podman system service` is a Linux-only subcommand — the macOS client
delegates the API service to the VM. Both assumptions in the harness
were stale, so the script tried to start a temporary service that podman
rejected with "unknown flag: --time".
- Discover the macOS socket via `podman machine inspect` instead of the
hardcoded path.
- On Darwin, fail fast with a "start podman machine" message rather than
attempting the Linux-only `podman system service` fallback.
- Write socket_path into [openshell.drivers.podman] so the in-process
driver picks up the discovered socket; the driver reads TOML only
after the config refactor (560550d2), so OPENSHELL_PODMAN_SOCKET alone
was no longer enough.
* fix(server): use clone_from for TLS client CA assignment
clippy 1.95.0 rejects assigning the result of `Clone::clone()` to an
existing variable under `-D warnings` (`assigning_clones`). Switch to
`clone_from(&...)` to satisfy the lint and avoid the redundant
allocation.
---------
Signed-off-by: Drew Newberry <anewberry@nvidia.com>
Co-authored-by: Drew Newberry <anewberry@nvidia.com>
Make --tls-client-ca optional and make client certificates always
optional when a CA is configured. This decouples HTTPS encryption
from mTLS authentication, allowing mTLS and OIDC bearer tokens to
coexist as parallel authentication mechanisms.
When --tls-client-ca is provided, client certificates are validated
against the CA when presented but never required. Clients may connect
with or without a certificate — authentication is handled at the
application layer (e.g. OIDC).
Two TLS modes are now supported:
- HTTPS with optional mTLS (--tls-client-ca provided)
- HTTPS-only (--tls-client-ca omitted)
The --disable-gateway-auth flag is preserved for backward
compatibility but is now a no-op. The allow_unauthenticated field
has been removed from TlsConfig. The Helm chart conditionally
includes the client-ca volume and env var based on whether
clientCaSecretName is configured.
* refactor!(auth): drop SSH handshake secret in favor of mTLS
The OPENSHELL_SSH_HANDSHAKE_SECRET / x-sandbox-secret mechanism was
misnamed: it does not authenticate SSH (which flows over the
RelayStream gRPC RPC and is gated by mTLS plus supervisor Unix-socket
permissions). It only gated a small set of sandbox-to-gateway
control-plane RPCs, and production deployments already enforce mTLS
on that channel — so the shared secret was redundant.
Replace the secret check with an mTLS-presence marker. Sandbox-class
methods (ReportPolicyStatus, PushSandboxLogs,
GetSandboxProviderEnvironment, SubmitPolicyAnalysis, GetSandboxConfig,
GetInferenceBundle) accept callers without a Bearer token; the gRPC
mTLS handshake is the trust boundary. Dual-auth methods treat
Bearer-present as full-scope CLI access and Bearer-absent as
sandbox-restricted scope via validate_sandbox_caller_update.
Drops the secret from all drivers (K8s, Podman, VM), the sandbox gRPC
interceptor, the Helm chart (values + pre-install hook + StatefulSet
env), the RPM bootstrap script, the man pages, and the
debug-openshell-cluster skill. Also removes the never-read
ssh_handshake_skew_secs flag and config field.
BREAKING CHANGE: --ssh-handshake-secret / OPENSHELL_SSH_HANDSHAKE_SECRET
and --ssh-handshake-skew-secs / OPENSHELL_SSH_HANDSHAKE_SKEW_SECS are
removed from the gateway, sandbox, and all driver binaries. The
openshell-ssh-handshake K8s Secret is no longer managed by the chart;
operators may delete the orphan. Deployments using
--disable-gateway-auth must enforce caller authentication at the
fronting proxy, since the gateway no longer validates a per-request
secret on sandbox-class methods.
Refs OS-174.
* docs(auth): scrub residual SSH handshake secret references
Sweep across docs, e2e scripts, the Podman driver README/NETWORKING
notes, the RPM/Helm/setup guides, the gateway man page, and RFC 0003
to remove instructions and examples that still referenced
OPENSHELL_SSH_HANDSHAKE_SECRET / --ssh-handshake-secret /
ssh_handshake_skew_secs. The mechanism is gone; nothing should still
suggest setting it. Negative-assertion regression tests are kept so
the env var cannot silently be re-introduced.
* test(drivers): drop SSH handshake secret negative-assertion tests
The supporting code, env vars, CLI flags, and config plumbing are
gone — these tests asserted absence of strings that no longer have
any path to being set. Remove the guards from the Docker, Podman,
Kubernetes, and VM driver test modules.
Consolidate shared constants and helpers into openshell-core to avoid
duplicate definitions scattered across crates. Adds OPENSHELL_SANDBOX
to the sandbox_env module so all drivers and the sandbox supervisor
reference it via a typed constant rather than string literals, matching
the existing pattern for SANDBOX_ID, ENDPOINT, and the other env vars.
Add opt-in support for Kubernetes user namespace isolation on sandbox
pods. When enabled, container UID 0 maps to an unprivileged host UID
and capabilities become namespaced, providing defense-in-depth for the
supervisor process.
Configuration is two-layered: a cluster-wide default via
OPENSHELL_ENABLE_USER_NAMESPACES (default false) and a per-sandbox
override via the new `user_namespaces` field on SandboxTemplate.
When user namespaces are active, the pod security context is extended
with SETUID, SETGID, and DAC_READ_SEARCH capabilities to match the
bounding-set requirements inside a user namespace.
Introduces SandboxPodParams struct to replace long argument lists on
sandbox_to_k8s_spec and sandbox_template_to_k8s.
Validated end-to-end on OCP 4.22 (K8s 1.35.3, CRI-O 1.35, RHEL
CoreOS, kernel 5.14) with full SSH tunnel and non-identity UID mapping.
Piotr Mlocek