Skip to content

ROADMAP.md

Latest commit

 

History

History
122 lines (83 loc) · 5.85 KB

ROADMAP.md

File metadata and controls

122 lines (83 loc) · 5.85 KB
Raw

Roadmap

This page records what works today, what we learned from early validation, and what to build next. The charter (CHARTER.md Part I) remains the long-range product plan; this doc is the practical engineering backlog.

Current capabilities

Area Status
BPF collector sched_switch / sched_waking, task storage, ringbuf, per-CPU stats, targeted TGID
Go uprobes runtime.casgstatus → tid_to_task / task_id on events
Userspace Ringbuf drain, trace v2 (CRTC), CLI record / analyze / export / eval
Symbolize Trace stack_id → frames (PC placeholders; ELF maps planned)
Export criticast export --pprof (gzip profile, critical_wait samples)
Attribution E1 lineage, E2 sudog-elem (offline GT), E3 resource suppress in analyze, E4 naive baseline
Analyzer Segments, SCC, longest path, false-wakeup filter, dominant-wait aggregation
Regression httpgo workload, adversarial server + interleaved load, scripts under scripts/

Published numbers: results/README.md.

Validated learnings (follow these in future work)

Overhead

  • At ~12k req/s (Go HTTP, wrk), full and sampled modes stayed under ~1% median throughput loss with zero ringbuf drops on a 6.1 kernel.
  • bpftrace spike peak ~12k wakes/s — well below the 1M/s danger zone.
  • Filter order matters: target → block vs preempt → min_block → sample → stack.

Attribution

Mechanism Lineage only (E1) Sudog elem (E2, GT) Notes
spawn-lineage 1.0 1.0 parentGoid / spawn sites
conn-pool, mutex 1.0 1.0 Waiter's own lineage; do not inherit waker cookie
chan-work-handoff ~0.55 1.0 (GT replay) Shared worker pool breaks per-goid cookie
chan (trace-joined E2) — ~0.78 BPF does not emit sudog.elem yet (aux always 0)

Product rule (from validation):

  • Ship Tier-0/1 (scheduler wait-for + lineage) as the default.
  • Use Tier-2 (channel/work handoff) only when sudog.elem (or equivalent) is present in the event stream.
  • Otherwise label request-ambiguous with confidence — never a confident wrong token.

E2 offline vs production: GT logs carry a per-item elem id at send/recv; attribution logic is validated. End-to-end trace join still needs kernel-side elem capture.

BPF gaps found in validation

bpf/collector.c reads last_sudog_elem / futex_uaddr into event.aux but does not write them yet. Trace-joined E2 therefore falls back to waker-token heuristics (~78% chan precision). Closing this gap is the top BPF attribution task.

Phase 1 — status

Plumbing (Bar A): validated on Linux 6.1 cloud — record/analyze/pprof, zero ringbuf drops under wrk (results/p1-smoke.md).

Thesis (Bar B): not validated on live httpgo demo — unscoped longest path, all WC_UNKNOWN, idle/sentinel dominant waits. Mechanism precision remains on P0 GT + eval, not this trace.

See P1_COMPLETION.md (required reading before claiming “P1 proves Criticast”).

P1 charter scope Shipped
End-to-end Tier-0/1 data path Yes
E3: no confident Tier-2 chan without aux Yes (code; not exercised when class unknown)
CI + golden tests Yes
Request-scoped critical path ≈ wall clock on live trace No (Bar B)
sudog.elem / real wait classes on live BPF No (P2)

Remaining: merge P1 to main if needed; all new work on phase2/tier2-product.

Next work: branch phase2/tier2-product · gate: scripts/validate-bar-b.sh (stub until P2 lands).

Next milestones (Phase 2 — active)

Bar B acceptance and sprint registry live in this file and PHASES.md. Summary:

1. Kernel refinement (L2) — P2 priority

  • Populate last_sudog_elem on Go channel block path (gopark / wait reason)
  • Optional futex_uaddr for mutex waits
  • TTL or generation on sudog map (pointer reuse after free)
  • Subject stack at block site (not only waker stack on EV_BLOCK_END)
  • wait_class refinement beyond classify_prev

2. Attribution & join (L3)

  • Wire trace aux → E2 in production path; re-run trace-joined eval (target chan ≥90%)
  • parentGoid from DWARF or runtime probe (reduce reliance on offsets.json only)
  • Cookie TTL enforcement in userspace replay
  • Broadcast / netpoll sites in adversarial fixture + matrix rows

3. Symbolization & export (L1 / L5)

  • internal/symbolize — Resolver + trace STACKS (P1)
  • ELF /proc symbolization + build-id cache
  • internal/export — pprof (P1)
  • OTLP-Profiles default export (P3)
  • CLI: analyze, export --pprof (P1); top TUI (P3)

4. Analyzer (L4)

  • SCC + longest path + dominant waits (P1)
  • Full wPerf cascade redistribution (E.2)
  • Confidence + ambiguous buckets in analyze (P1)
  • Golden trace tests (P1)

5. Operations

  • CI: kernel matrix compile, attribution regression thresholds in go test
  • Container image rename / polish (criticast-dev vs legacy tag)
  • Operator docs for cgroup targeting, capabilities, multi-tenant scoping

Regression expectations

When touching BPF or attribution:

  1. ./scripts/verify.sh
  2. Re-run overhead bench if hot path changed (GETTING_STARTED.md)
  3. Re-run adversarial eval --mode all if L3 changed
  4. Update results/phase0/ reports with date, commit, and tables

Do not regress: spawn/pool/mutex precision; overhead <5% at charter load; zero ringbuf drops on benchmark runs.

Non-goals (unchanged)

Distributed tracing, L7 payload parsing, metrics TSDB, Go uretprobes, blocking kernel on userspace backpressure, naive waker-cookie propagation at shared resources.