SQLite Metrics

• sqlite_commit_phase_duration_seconds{phase,path}: Engine-side histogram for commit request phases. path is fast or slow. phase is decode_request, meta_read, ltx_encode, pidx_read, udb_write, or response_build.
• sqlite_commit_stage_phase_duration_seconds{phase}: Engine-side histogram for staged commit uploads. phase is decode, stage_encode, or udb_write.
• sqlite_commit_finalize_phase_duration_seconds{phase}: Engine-side histogram for staged commit finalize work. phase is stage_promote, pidx_write, or meta_write.
• sqlite_commit_dirty_page_count{path}: Histogram of dirty page counts per commit path.
• sqlite_commit_dirty_bytes{path}: Histogram of raw dirty-page bytes per commit path.
• sqlite_udb_ops_per_commit{path}: Histogram of UniversalDB operations per commit path.
• envoy_sqlite_commit_dispatch_duration_seconds: Pegboard-envoy histogram for websocket frame arrival to depot dispatch.
• envoy_sqlite_commit_response_duration_seconds: Pegboard-envoy histogram for depot return to websocket response send.
• sqlite_commit_phases: Actor inspector labeled timing metric exposed from /inspector/metrics. Values are request_build, serialize, transport, and state_update.

Scrape Points

• Engine and pegboard-envoy Prometheus metrics come from the shared /metrics endpoint on port 6430.
• Actor-local commit timings come from GET /inspector/metrics on the actor gateway route.

Tracing

• Set RUST_LOG=depot=debug,pegboard_envoy=debug,sqlite_v2_vfs=debug to emit per-commit phase spans and VFS phase logs.
• depot commit handlers use debug spans for the high-level request and sub-phase work.
• The VFS logs request-build, serialize, transport, and state-update timings after each successful commit.

Diagnosis

• High decode_request or envoy_sqlite_commit_dispatch_duration_seconds usually means envoy-side validation or actor lookup is slow before storage work starts.
• High meta_read or pidx_read points at UniversalDB read pressure or cache misses.
• High ltx_encode means commit encoding and compression are doing real work. Check dirty page counts and raw dirty bytes together.
• High udb_write, meta_write, or envoy_sqlite_commit_response_duration_seconds points at write-path latency after encode.
• A healthy actor should show non-zero sqlite_commit_phases totals after commits in /inspector/metrics. If SQL runs but those timings stay zero, the native VFS metrics path is broken.