Description
Written by an AI agent (Codex) on behalf of @jasonkuhrt.
Problem
Users with an instrumented remote cache server cannot connect its requests to the vp run clients that made them. Native cache fetch, store and blob-download requests currently have neither exported OpenTelemetry client spans nor propagated W3C trace context, so those callers are missing from trace-based service maps.
Proposed capability
Allow vp run to export its native remote-cache HTTP client spans through standard OTLP/HTTP configuration. For example, with a collector configured in the command's environment:
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 \
OTEL_SERVICE_NAME=build-client \
vp run build --remote-cache read-writeThe instrumented cache server should receive a traceparent identifying the exported client span. A tracing backend can then connect the caller to its cache server.
The proposed behavior is opt-in when an OTLP endpoint is configured, honors the standard service/resource/header/protocol variables and export opt-outs, and flushes at session completion, including a failed task. A normal cache miss remains successful cache behavior. Exported attributes exclude cache credentials, keys, payloads and error-response contents; an unavailable collector must not fail the task run.
Current implementation
The HTTP transport lives in crates/vt_remote_cache/src/lib.rs; RemoteClients in crates/vt/src/session/cache/remote.rs composes it for a run. The current transport calls reqwest directly without an OpenTelemetry tracer or W3C propagation.
A focused draft implementation follows those boundaries: the session owns the optional SDK/exporter lifetime, and the remote-cache transport accepts a tracer for its HTTP operations. It does not change cache keys, payloads or authentication. Existing internal Rust performance tracing is a separate capability.
Metadata
Metadata
Assignees
Labels
Type
Fields
Priority
Activity