Repository navigation
feat(ai/sidecar): online SGD+EMA trainer with replay buffer (ADR-0781) - #178
Merged
Merged
Conversation
lusoris
force-pushed
the
feat/sidecar-online-trainer
branch
from
May 29, 2026 12:12
7fa9ed8 to
f369e94
Compare
Implements Phase 4b item 4b.7 (Research-0733): a Python sidecar container co-located with vmafx-node that continuously fine-tunes a tiny-AI ONNX model from live (features, true_score) pairs emitted during scoring jobs. - ai/sidecar/replay_buffer.py: thread-safe bounded ring buffer (10 000 samples, FIFO eviction) for experience-replay batching. - ai/sidecar/sgd_ema.py: online SGD + EMA trainer (beta=0.999, gradient clipping, 100-step LR warmup, atomic ONNX export at opset 17). - ai/sidecar/online_trainer.py: Unix-socket server (/tmp/vmafx-sidecar.sock), newline-delimited JSON, mixed-batch training, periodic checkpoint export with SHA-256 sidecar file. - cmd/vmafx-node/online_feedback.go: non-blocking Go FeedbackClient with a 1000-entry in-process ring buffer; fire-and-forget from the scoring path. - deploy/helm/vmafx/templates/sidecar-trainer.yaml: named template vmafx.sidecarContainer, gated by sidecar.trainer.enabled (default false), k8s 1.29+ native sidecar (restartPolicy: Always). - Tests: test_replay_buffer.py (thread-safety, eviction, stats), test_sgd_ema.py (EMA formula, warmup, ONNX export round-trip). - ADR-0781, docs/ai/sidecar-online-training.md, changelog, rebase-notes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
lusoris
force-pushed
the
feat/sidecar-online-trainer
branch
from
May 29, 2026 15:54
f369e94 to
fb60a88
Compare
lusoris
marked this pull request as ready for review
May 31, 2026 13:38
lusoris
added a commit
that referenced
this pull request
Jun 12, 2026
…CI) (#868) * fix(codeql): resolve HIGH-severity security-cpp-high alerts (23 sites) - cpp/integer-multiplication-cast-to-long (11): pre-cast one operand to size_t / double / ptrdiff_t before int*int multiplications in cambi.c, float_vif.c (log message), iqa/convolve.c (img_offset), moment.c, psnr.c, and vif_tools.c (four memcpy size expressions). Add stddef.h to convolve.c for ptrdiff_t. - cpp/incomplete-parity-check (3): change `% 2 == 1` to `% 2 != 0` in vif_tools.c (assert), svm.cpp (powi loop), pdjson.c (JSON object key/value alternation). The == 1 form is wrong for negative operands; != 0 is always correct. - cpp/wrong-type-format-argument (2): fix float_vif.c error log that printed size_t fields scaled_w/scaled_h with %d; change to %zu. - cpp/world-writable-file-creation (1): in vmaf.cpp replace bare fopen("wb") with open(O_WRONLY|O_CREAT|O_TRUNC, 0644)+fdopen() on POSIX so the created file is never world-writable independent of the caller's umask. Add <fcntl.h>. - cpp/path-injection (4): in test_output.c resolve the mkstemp-created path through realpath() immediately after creation, breaking the taint chain from getenv("TMPDIR") to the vmaf_write_output call site. - cpp/toctou-race-condition (2): skipped — both sites are in test cleanup (RMDIR after stat assertion). The stat result drives a test assertion, not a security-sensitive access decision; no atomic replacement of open() is applicable to rmdir. Reported as skipped. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(codeql): resolve security-python-and-ci CodeQL alerts Fixes 18 open CodeQL alerts across the Python and CI categories: - yaml.github-actions.security.run-shell-injection (#661): move github.event_name, github.base_ref, and github.event.before from inline ${{...}} interpolation to env: vars in the SYCL clang-tidy detect step of lint-and-format.yml. - python.lang.security.use-defused-xml-parse (#216, #217): replace xml.etree.ElementTree with defusedxml.ElementTree in feature_extractor.py and quality_runner.py; add defusedxml>=0.7.1 to python/pyproject.toml and python/requirements.txt. - py/undefined-export (#352, #353, #354, #616): restructure aiutils/__init__.py to do a conditional eager import of the parquet helpers so the names are defined when pyarrow is present, and only include them in __all__ when the import succeeded. - py/stack-trace-exposure (#178, #179, #585): log exception detail server-side and return a generic message to the HTTP client in http_transport.py _handle_score (invalid JSON, bad params, scorer error branches). - python.lang.security.audit.dangerous-subprocess-use-tainted-env-args (#227, #372): add shlex.quote() around user-supplied path arguments passed into shell strings in extract_ugc_features.py and test_bbb_e2e_v5_bug_cluster.py. - py/file-not-closed (#677, #678): replace bare open() calls with context managers in test_coverage_round3.py. - py/redundant-comparison (#427, #431): remove redundant assert not (x != y) lines that duplicate the preceding assert x == y. - py/equals-hash-mismatch (#182): convert RdPoint to frozen=True dataclass so __eq__ and __hash__ are generated consistently. - py/inheritance/signature-mismatch (#197): add result_dict=None default to EnsembleVmafQualityRunner._populate_result_dict so the signature is compatible with the base class. - py/multiple-definition (#201): drop redundant assignment to feature_found in feature_extractor.py wildcard discovery path. - py/str-format/surplus-named-argument (#204): remove unused dataset= kwarg from the format() call in routine.py. Skipped: python.lang.security.audit.insecure-file-permissions (#373) — the Unix socket at 0o660 is intentional (Go sidecar node must write to it and runs as the same UNIX group); tightening to 0o644 would break the IPC channel. Skipped: py/path-injection (#180, #181) — _validate_path() already resolves the path and checks it against an allowlist before any file operation; the data flow is secure and the CodeQL dataflow trace is a false positive on this allowlisted pattern. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
3 of 6 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
vmafx-nodethat continuously fine-tunes a tiny-AI ONNX model from live(features, true_score)pairs produced during scoring jobs./tmp/vmafx-sidecar.sock(sharedemptyDir), newline-delimited JSON — simpler and lower overhead than gRPC loopback for the expected < 100 pairs/s rate.beta=0.999, Mean Teacher) stabilises checkpoints against noisy burst updates.FeedbackClientis non-blocking from the scoring path (1000-entry ring buffer + background drainer).Deliverables checklist (ADR-0108)
docs/research/0733-vmafx-sidecar-training-architecture.md(pre-existing, authored in Phase 4b planning)## Alternatives considered(gRPC vs Unix socket, dedicated training pool vs sidecar, Go-native SGD)AGENTS.mdinvariant note:ai/sidecar/AGENTS.md(socket path sync, wire protocol, ONNX opset, buffer capacity)pytest ai/sidecar/tests/(replay buffer + SGD+EMA; PyTorch optional, skipped if absent)changelog.d/added/sidecar-sgd-ema-online-trainer.mddocs/rebase-notes.md§ADR-0781Documentation (ADR-0100 / ADR-0042)
docs/ai/sidecar-online-training.md— operator guide covering quick-start, architecture diagram, all config env vars, checkpoint format, k8s 1.29+ native sidecar semantics, Go metrics, and v1 limitations.docs/adr/0781-sidecar-sgd-ema-online-trainer.md— design rationale (ADR status: Proposed — architecture review needed before Accepted).Architecture review questions
node-deployment.yamlvsnode.yaml— the new Helm chart layout onmasteraddednode.yaml; this PR still modifiesnode-deployment.yaml. Do the sidecar container stanzas need to move tonode.yaml? Reviewer to advise.0.999— appropriate for 50 encodes/hour workloads (~1000-step half-life ≈ 20 hours); may need tuning for lower-traffic deployments. Recommend operator override viaVMAFX_SIDECAR_EMA_DECAY.VmafxModelTrainingCRD + atomic session swap (Research-0733 §3.4) is a follow-up; confirm scope boundary.Test plan
pytest ai/sidecar/tests/test_replay_buffer.py— no PyTorch needed, runs in < 1 spytest ai/sidecar/tests/test_sgd_ema.py— requires PyTorch; skipped if absentgo build ./cmd/vmafx-node/...— verifiesonline_feedback.gocompiles cleanhelm template . -f values.yaml --set sidecar.trainer.enabled=true— verify sidecar stanza rendersState / bug tracking
no state.md update needed: no bug opened, closed, or ruled not-affecting.
no ffmpeg-patches impact: no public C-API or header change.
🤖 Generated with Claude Code